Sanet ST

Advanced Studies in Theoretical and Applied Econometrics 53
Felix Chan
László Mátyás Editors
Econometrics
with
Machine
Learning
Advanced Studies in Theoretical and Applied
Econometrics
Volume 53
Series Editors
Badi Baltagi , Center for Policy Research, Syracuse University, Syracuse, NY,
USA
Yongmiao Hong, Department of Economics, Cornell University, Ithaca, NY, USA
Gary Koop, Department of Economics, University of Strathclyde, Glasgow, UK
Walter Krämer, Business and Social Statistics Department, TU Dortmund
University, Dortmund, Germany
László Mátyás, Department of Economics, Central European University,
Budapest, Hungary and Vienna, Austria
This book series aims at addressing the most important and relevant current issues
in theoretical and applied econometrics. It focuses on how the current data revolution
has affected econometric modeling, analysis and forecasting, and how applied work
has benefitted from this newly emerging data-rich environment. The series deals with
all aspects of macro-, micro-, financial-, and econometric methods and related
disciplines, like, for example, program evaluation or spatial analysis.
The volumes in the series are either monographs or edited volumes, mainly
targeting researchers, policymakers, and graduate students.
This book series is listed in Scopus.
Felix Chan • László Mátyás
Editors
Econometrics with Machine

Learning
Editors
Felix Chan László Mátyás
School of Accounting, Economics Department of Economics
& Finance Central European University
Curtin University Budapest, Hungary and Vienna, Austria
Bentley, Perth, WA, Australia
ISSN 1570-5811 ISSN 2214-7977 (electronic)

Advanced Studies in Theoretical and Applied Econometrics
ISBN 978-3-031-15148-4 ISBN 978-3-031-15149-1 (eBook)
https://doi.org/10.1007/978-3-031-15149-1
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword
Felix Chan and László Mátyás have provided great service to the profession by writing
and editing this volume of contributions to econometrics and machine learning. This
is a remarkably fast moving area of research with long-term consequences for the analysis
of high dimensional, ‘big data’, and its prospects for model and policy evaluation.
The book reflects well the current direction of research focus that is relevant to
professionals who are more concerned with ‘partial effects’ in the statistical analysis of
data, so-called ‘causal analysis’. Partial effects are of central interest to policy and
treatment effect evaluation, as well as optimal decision making in all applied fields,
such as market research, evaluation of treatment outcomes in economics and health,
finance, in counterfactual analysis, and model building in all areas of science. The
holy grail of quantitative economics/econometrics research in the last 100 years has
been the identification and development of ‘causal’ models, with a primary focus on
conditional expectation of one or more target variables/outcomes, conditioned on
several ‘explanatory’ variables, ‘features’ in the ML jargon. This edifice depends
crucially on two decisions: correct functional form and a correct set of both target
explanatory variables and additional control variables, reflecting various degrees of
departure from the gold standard randomized experiment and sampling.
This holy grail has been a troubled path since there is little or no guidance in
substantive sciences on functional forms, and certainly little or no indication on
sampling and experimental failures, such as selection. Most economists would admit,
at least privately, that quantitative models fail to perform successfully in a consistent
manner. This failure is especially prominent in out of sample forecasting and rare event
prediction, that is, in counterfactual analysis, a central activity in policy evaluation
and optimal decision making. The dominant linear, additively separable multiple
regression, the workhorse of empirical research for so many decades, has likely
created a massive reservoir of misleading ‘stylised facts’ which are the artefacts of
linearity, and its most obvious flaw, constant partial effect (coefficients).
Nonlinear, nonparametric, semiparametric and quantile models and methods have
developed at a rapid pace, with related approximate/asymptotic inference theory,
to deal with these shortcomings. This development has been facilitated by rapidly
expanding computing capacity and speed, and greater availability of rich data samples.
v
vi Foreword
These welcome developments and movements are meant to provide more reliable and
‘robust’ empirical findings and inferences. For a long time, however, these techniques
have been limited to a small number of conditioning variables, or a small set of ‘moment
conditions’, and subject to the curse of dimensionality in nonparametric and other robust
methods.
The advent of regularization and penalization methods and algorithms has opened
the door to allow for model searches in which an impressive degree of allowance
may be made for both possible nonlinearity of functional forms (filters, ‘learners’),
and potential explanatory/predictive variables, even possibly larger in number than
the sample size. Excitement about these new ‘machine learning (ML)’ methods is
understandable with generally impressive prediction performance. Stupendously
fast and successful ‘predictive text’ search is a common point of contact with this
experience for the public. Fast and cheap computing is the principal facilitator, some
consider it ‘the reason’ for mass adoption.
It turns out that an exclusive focus on prediction criteria has some deliterious
consequences for the identification and estimation of partial effects and model objects
that are central to economic analysis, and other substantive areas. Highly related
‘causes’ and features are quite easily removed in ‘sparsity’ techniques, such as LASSO,
producing ‘biased’ estimation of partial effects. In my classes, I give an example of
a linear model with some exact multicollinear variables, making the identification of
some partial effects impossible (‘biased’?), without impacting the estimation of the
conditional mean, the ‘goodness of fit’, or the prediction criteria. The conditional
mean is an identified/estimable function irrespective of the degree of multicollinearity!
The general answer to this problem has been to, one way or other, withhold a set of
target features from ‘selection’ and possible elimination by LASSO and other naive
‘model selection’ techniques. Double machine learning (DML), Random Forests, and
subset selection and aggregation/averaging are examples of such methods, some being
pre-selection approaches and others being various types of model averaging. There
are numerous variations to these examples, but I choose the taxonomy of ‘model
selection’ vs ‘model averaging’ approaches. These may be guided by somewhat
traditional econometrics thinking and approaches, or algorithmic, computer science
approaches that are less concerned with rigorous statistical ‘inference’.
Econometricians are concerned with rigorous inference and rigorous analysis of
identification. Given their history of dealing with the examination of poorly collected
‘observational’ data, on the one hand, and the immediacy of costly failures of models
in practical applications and governance, econometricians are well placed to lead in
the development of ‘big data’ techniques that accommodate the dual goals of accurate
prediction and identification (unbiased?) of partial effects.
This volume provides a very timely set of 10 chapters that help readers to appreciate
the nature of the challenges and promise of big data methods (data science!?), with
a timely and welcome emphasis on ‘debiased’ and robust estimation of partial and
treatment effects. The first three chapters are essential reading, helped by a bridge
over the ocean of the confusing new naming of old concepts and objects in statistics (see
the Appendix on the terminology). Subsequent chapters contribute by delving further
Foreword vii
into some of the rapidly expanding techniques, and some key application areas, such
as the health sciences and treatment effects.
Doubly robust ML methods are introduced in several places in this volume, and
will surely be methods of choice for sometime to come. The book lays bare the
challenges of causal model identification and analysis which reflect familiar challenges
of model uncertainty and sampling variability. It makes clear that larger numbers of
variables and moment conditions, as well as greater flexibility in functrional forms,
ironically, produce new challenges (e,g., highly dependent features, reporting and
summary of partial effects which are no longer artificial constant, increased danger
of endogeneity,...). Wise and rigorous model building and expert advice will remain
an art, especially as long as we deal with subfield (economics) models which cannot
take all other causes into account. Including everything and the kitchen sink turns out
to be a hindrance in causal identification and counterfactual analysis, but a boon to
black box predictive algorithms. A profit making financial returns algorithm is ‘king’
until it fails. When it fails, we have tough time finding out why! Corrective action
will be by trial and error. That is a hard pill when economic policy decisions take so
long to take effect, if any.
A cautionary note, however, is that rigorous inference is also challenging, as in the
rather negative, subtle finding of a lack of ‘uniformity’ in limit results for efficiency
bounds. This makes ‘generalisability’ out of any subsample and sample problematic
since the limiting results are pointwise, not uniform.
The authors are to be congratulated for this timely contribution and for the breadth
of issues they have covered. Most readers will find this volume to be of great value,
especially in teaching, even though they all surely wish for coverage of even more
techniques and algorithms.
I enjoyed reading these contributions and have learned a great deal from them.
This volume provides invaluable service to the profession and students in econometrics,
statistics, computer science, data sciences, and more.
Emory University, Atlanta, USA, June 2022 Esfandiar Maasoumi

Preface
In his book The Invention of Morel, the famous Argentinian novelist, Adolfo Bioy
Casares, creates what we would now call a parallel universe, which can hardly be
distinguished from the real one, and in which the main character gets immersed
and eventually becomes part of. Econometricians in the era of Big Data feel a bit
like Bioy Casares’ main character: We have a hard time making up our mind about
what is real and imagined, what is a fact or an artefact, what is evidence or just
perceived, what is a real signal or just noise, whether our data is or represents the
reality we are interested in, or whether they are just some misleading, meaningless
numbers. In this book, we aim to provide some assistance to our fellow economists
and econometricians in this respect with the help of machine learning. What we hope
to add is, as the German poet and scientist, Johann Wolfgang von Goethe said (or
not) in his last words: Mehr Licht (More light).
In the above spirit, the volume aims to bridge the gap between econometrics and
machine learning and promotes the use of machine learning methods in economic
and econometric modelling. Big Data not only provide a plethora of information,
but also often the kind of information that is quite different from what traditional
statistical and econometric methods are grown to rely upon. Methods able to uncover
deep and complex structures in (very) large data sets, let us call them machine
learning, are ripe to be incorporated into the econometric toolbox. However, this is
not painless as machine learning methods are rooted in a different type of enquiry
than econometrics. Unlike econometrics, they are not focused on causality, model
specification, or hypothesis testing and the like, but rather on the underlying properties
of the data. They often rely on algorithms to build models geared towards prediction.
They represent two cultures: one motivated by prediction, the other by explanation.
Mutual understanding is not promoted by their use of different terminology. What in
econometrics is called sample (or just data) used to estimate the unknown parameters
of a model, in machine learning is often referred to as a training sample. The unknown
parameters themselves may be known as weights, which are estimated through a
learning or training process (the ‘machine’ or algorithm itself). Machine learning talks
about supervised learning where both the covariates (explanatory variables, features,
ix
x Preface
or predictors) and the dependent (outcome) variables are observed and unsupervised
learning, where only the covariates are observed. Machine learning’s focus on
prediction is most often structural, does not involve time, while in econometrics this
mainly means forecasting (about some future event).
The purpose of this volume is to show that despite this different perspective,
machine learning methods can be quite useful in econometric analyses. We do not
claim to be comprehensive by any means. We just indicate the ways this can be done,
what we know and what we do not and where research should be focused.
The first three chapters of the volume lay the foundations of common machine
learning techniques relevant to econometric analysis. Chapter 1 on Linear Econo-
metrics Models presents the foundation of shrinkage estimators. This includes ridge,
Least Absolute Shrinkage and Selection Operator (LASSO) and their variants as
well as their applications to linear models, with a special focus on valid statistical
inference for model selection and specification. Chapter 2 extends the discussion to
nonlinear models and also provides a concise introduction to tree based methods
including random forest. Given the importance of policy evaluation in economics,
Chapter 3 presents the most recent advances in estimating treatment effects using
machine learning. In addition to discussing the different machine learning techniques
in estimating average treatment effects, the chapter presents recent advances in
identifying the treatment effect heterogeneity via a conditional average treatment
effect function.
The next parts extend and apply the foundation laid by the first three chapters
to specific problems in applied economics and econometrics. Chapter 4 provides
a comprehensive introduction to Artificial Neural Networks and their applications
to economic forecasting with a specific focus on rigorous evaluation of forecast
performances between different models. Building upon the knowledge presented in
Chapter 3, Chapter 5 presents a comprehensive survey of the applications of causal
treatment effects estimation in Health Economics. Apart from Health Economics,
machine learning also appears in development economics, as discussed in Chapter 9.
Here a comprehensive survey of the applications of machine learning techniques in
development economics is presented with a special focus on data from Geographical
Information Systems (GIS) as well as methods in combining observational and (quasi)
experimental data to gain an insight into issues around poverty and inequality.
In the era of Big Data, applied economists and econometricians are exposed to
a large number of additional data sources, such as data collected by social media
platforms and transaction data captured by financial institutions. Interdependence
between individuals reveals insights and behavioural patterns relevant to policy
makers. However, such analyses require technology and techniques beyond the
traditional econometric and statistical methods. Chapter 6 provides a comprehensive
review of this subject. It introduces the foundation of graphical models to capture
network behaviours and discusses the most recent procedures to utilise the large
volume of data arising from such networks. The aim of such analyses is to reveal
the deep patterns embedded in the network. The discussion of graphical models and
their applications continues in Chapter 8, which discusses how shrinkage estimators
presented in Chapters 1 and 2 can be applied to graphical models and presents its
Preface xi
applications to portfolio selection problems via a state-space framework. Financial

applications of machine learning are not limited to portfolio selection, as shown in
Chapter 10, which provides a comprehensive survey of the contribution of machine
learning techniques in identifying the relevant factors that drive empirical asset
pricing.
Since data only capture historical information, any bias or prejudice induced by
humans in their decision making is also embedded into the data. Predictions from
these data therefore continue with such bias and prejudice. This issue is becoming
increasingly important and Chapter 7 provides a comprehensive survey of recent
techniques to enforce fairness in data-driven decision making through Structural
Econometric Models.
Perth, June 2022 Felix Chan

Budapest and Vienna, June 2022 László Mátyás
Acknowledgements
We address our thanks to all those who have facilitated the birth of this book: the
contributors who produced quality work, despite onerous requests and tight deadlines;
Esfandiar Maasoumi, who supported this endeavour and encouraged the editors from
the very early planning stages; and last but not least, the Central European University
and Curtin University, who financially supported this project. Some chapters have
been polished with the help of Eszter Timár. Her English language editing made them
easier and more enjoyable to read.
The final camera–ready copy of the volume has been prepared with LATEX and
Overleaf by the authors, the editors and some help from Sylvia Soltyk and the LATEX
wizard Oliver Kiss.
xiii
Contents
1 Linear Econometric Models with Machine Learning . . . . . . . . . . . . . . 1

Felix Chan and László Mátyás
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Shrinkage Estimators and Regularizers . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 𝐿 𝛾 norm, Bridge, LASSO and Ridge . . . . . . . . . . . . . . . . . 6
1.2.2 Elastic Net and SCAD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3 Adaptive LASSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.4 Group LASSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.1 Computation and Least Angular Regression . . . . . . . . . . . 13
1.3.2 Cross Validation and Tuning Parameters . . . . . . . . . . . . . . 14
1.4 Asymptotic Properties of Shrinkage Estimators . . . . . . . . . . . . . . . . 15
1.4.1 Oracle Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.2 Asymptotic Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.3 Partially Penalized (Regularized) Estimator . . . . . . . . . . . . 20
1.5 Monte Carlo Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.5.1 Inference on Unpenalized Parameters . . . . . . . . . . . . . . . . . 23
1.5.2 Variable Transformations and Selection Consistency . . . . 25
1.6 Econometrics Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.6.1 Distributed Lag Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.6.2 Panel Data Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.6.3 Structural Breaks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.7 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Proof of Proposition 1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2 Nonlinear Econometric Models with Machine Learning . . . . . . . . . . . 41
Felix Chan, Mark N. Harris, Ranjodh B. Singh and Wei (Ben) Ern Yeo
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.2 Regularization for Nonlinear Econometric Models . . . . . . . . . . . . . . 43
xv
xvi Contents
2.2.1 Regularization with Nonlinear Least Squares . . . . . . . . . . 44

2.2.2 Regularization with Likelihood Function . . . . . . . . . . . . . . 46
Continuous Response Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Discrete Response Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.2.3 Estimation, Tuning Parameter and Asymptotic Properties 50
Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Tuning Parameter and Cross-Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Asymptotic Properties and Statistical Inference . . . . . . . . . . . . . . . . . . . . . . 52
2.2.4 Monte Carlo Experiments – Binary Model with shrinkage 56
2.2.5 Applications to Econometrics . . . . . . . . . . . . . . . . . . . . . . . 61
2.3 Overview of Tree-based Methods - Classification Trees and
Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.3.1 Conceptual Example of a Tree . . . . . . . . . . . . . . . . . . . . . . . 66
2.3.2 Bagging and Random Forests . . . . . . . . . . . . . . . . . . . . . . . 68
2.3.3 Applications and Connections to Econometrics . . . . . . . . . 70
Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3 The Use of Machine Learning in Treatment Effect Estimation . . . . . . 79
Robert P. Lieli, Yu-Chin Hsu and Ágoston Reguly
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.2 The Role of Machine Learning in Treatment Effect Estimation: a
Selection-on-Observables Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.3 Using Machine Learning to Estimate Average Treatment Effects . . 84
3.3.1 Direct versus Double Machine Learning . . . . . . . . . . . . . . 84
3.3.2 Why Does Double Machine Learning Work and Direct
Machine Learning Does Not? . . . . . . . . . . . . . . . . . . . . . . . 87
3.3.3 DML in a Method of Moments Framework . . . . . . . . . . . . 89
3.3.4 Extensions and Recent Developments in DML . . . . . . . . . 90
3.4 Using Machine Learning to Discover Treatment Effect Heterogeneity 92
3.4.1 The Problem of Estimating the CATE Function . . . . . . . . 92
3.4.2 The Causal Tree Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.4.3 Extensions and Technical Variations on the Causal Tree
Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3.4.4 The Dimension Reduction Approach . . . . . . . . . . . . . . . . . 99
3.5 Empirical Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Contents xvii
4 Forecasting with Machine Learning Methods . . . . . . . . . . . . . . . . . . . . 111

Marcelo C. Medeiros
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.2 Modeling Framework and Forecast Construction . . . . . . . . . . . . . . . 113
4.2.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.2.2 Forecasting Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.2.3 Backtesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.2.4 Model Choice and Estimation . . . . . . . . . . . . . . . . . . . . . . . 117
4.3 Forecast Evaluation and Model Comparison . . . . . . . . . . . . . . . . . . . 120
4.3.1 The Diebold-Mariano Test . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.3.2 Li-Liao-Quaedvlieg Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.3.3 Model Confidence Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.4 Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.4.1 Factor Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.4.2 Bridging Sparse and Dense Models . . . . . . . . . . . . . . . . . . 127
4.4.3 Ensemble Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.5 Nonlinear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.5.1 Feedforward Neural Networks . . . . . . . . . . . . . . . . . . . . . . . 131
4.5.2 Long Short Term Memory Networks . . . . . . . . . . . . . . . . . 136
4.5.3 Convolution Neural Networks . . . . . . . . . . . . . . . . . . . . . . . 139
4.5.4 Autoenconders: Nonlinear Factor Regression . . . . . . . . . . 145
4.5.5 Hybrid Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5 Causal Estimation of Treatment Effects From Observational Health
Care Data Using Machine Learning Methods . . . . . . . . . . . . . . . . . . . 151
William Crown
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.2 Naïve Estimation of Causal Effects in Outcomes Models with
Binary Treatment Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.3 Is Machine Learning Compatible with Causal Inference? . . . . . . . . 154
5.4 The Potential Outcomes Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5.5 Modeling the Treatment Exposure Mechanism–Propensity Score
Matching and Inverse Probability Treatment Weights . . . . . . . . . . . 157
5.6 Modeling Outcomes and Exposures: Doubly Robust Methods . . . . 158
5.7 Targeted Maximum Likelihood Estimation (TMLE) for Causal
Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.8 Empirical Applications of TMLE in Health Outcomes Studies . . . . 163
5.8.1 Use of Machine Learning to Estimate TMLE Models . . . . 163
5.9 Extending TMLE to Incorporate Instrumental Variables . . . . . . . . . 164
5.10 Some Practical Considerations on the Use of IVs . . . . . . . . . . . . . . . 165
5.11 Alternative Definitions of Treatment Effects . . . . . . . . . . . . . . . . . . . 166
xviii Contents
5.12 A Final Word on the Importance of Study Design in Mitigating Bias168

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
6 Econometrics of Networks with Machine Learning . . . . . . . . . . . . . . . 177
Oliver Kiss and Gyorgy Ruzicska
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
6.2 Structure, Representation, and Characteristics of Networks . . . . . . . 179
6.3 The Challenges of Working with Network Data . . . . . . . . . . . . . . . . 182
6.4 Graph Dimensionality Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
6.4.1 Types of Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
6.4.2 Algorithmic Foundations of Embeddings . . . . . . . . . . . . . . 187
6.5 Sampling Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
6.5.1 Node Sampling Approaches . . . . . . . . . . . . . . . . . . . . . . . . . 190
6.5.2 Edge Sampling Approaches . . . . . . . . . . . . . . . . . . . . . . . . . 191
6.5.3 Traversal-Based Sampling Approaches . . . . . . . . . . . . . . . . 192
6.6 Applications of Machine Learning in the Econometrics of Networks196
6.6.1 Applications of Machine Learning in Spatial Models . . . . 196
6.6.2 Gravity Models for Flow Prediction . . . . . . . . . . . . . . . . . . 203
6.6.3 The Geographically Weighted Regression Model and ML 205
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
7 Fairness in Machine Learning and Econometrics . . . . . . . . . . . . . . . . . 217
Samuele Centorrino, Jean-Pierre Florens and Jean-Michel Loubes
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.2 Examples in Econometrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
7.2.1 Linear IV Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
7.2.2 A Nonlinear IV Model with Binary Sensitive Attribute . . 223
7.2.3 Fairness and Structural Econometrics . . . . . . . . . . . . . . . . . 223
7.3 Fairness for Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
7.4 Full Fairness IV Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
7.4.1 Projection onto Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
7.4.2 Fair Solution of the Structural IV Equation . . . . . . . . . . . . 230
7.4.3 Approximate Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
7.5 Estimation with an Exogenous Binary Sensitive Attribute . . . . . . . . 240
7.6 An Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
7.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8 Graphical Models and their Interactions with Machine Learning in
the Context of Economics and Finance . . . . . . . . . . . . . . . . . . . . . . . . . 251
Ekaterina Seregina
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
8.1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
8.2 Graphical Models: Methodology and Existing Approaches . . . . . . . 253
8.2.1 Graphical LASSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Contents xix
8.2.2 Nodewise Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258

8.2.3 CLIME . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
8.2.4 Solution Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
8.3 Graphical Models in the Context of Finance . . . . . . . . . . . . . . . . . . . 262
8.3.1 The No-Short-Sale Constraint and Shrinkage . . . . . . . . . . 267
8.3.2 The 𝐴-Norm Constraint and Shrinkage . . . . . . . . . . . . . . . . 270
8.3.3 Classical Graphical Models for Finance . . . . . . . . . . . . . . . 272
8.3.4 Augmented Graphical Models for Finance Applications . 273
8.4 Graphical Models in the Context of Economics . . . . . . . . . . . . . . . . 278
8.4.1 Forecast Combinations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
8.4.2 Vector Autoregressive Models . . . . . . . . . . . . . . . . . . . . . . . 280
8.5 Further Integration of Graphical Models with Machine Learning . . 283
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
9 Poverty, Inequality and Development Studies with Machine Learning 291
Walter Sosa-Escudero, Maria Victoria Anauati and Wendy Brau
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
9.2 Measurement and Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
9.2.1 Combining Sources to Improve Data Availability . . . . . . . 294
9.2.2 More Granular Measurements . . . . . . . . . . . . . . . . . . . . . . . 298
9.2.3 Dimensionality Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . 304
9.2.4 Data Imputation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
9.2.5 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
9.3 Causal Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
9.3.1 Heterogeneous Treatment Effects . . . . . . . . . . . . . . . . . . . . 307
9.3.2 Optimal Treatment Assignment . . . . . . . . . . . . . . . . . . . . . . 312
9.3.3 Handling High-Dimensional Data and Debiased ML . . . . 313
9.3.4 Machine-Building Counterfactuals . . . . . . . . . . . . . . . . . . . 315
9.3.5 New Data Sources for Outcomes and Treatments . . . . . . . 316
9.3.6 Combining Observational and Experimental Data . . . . . . 319
9.4 Computing Power and Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
10 Machine Learning for Asset Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . 337

Jantje Sönksen
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
10.2 How Machine Learning Techniques Can Help Identify Stochastic
Discount Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
10.3 How Machine Learning Techniques Can Test/Evaluate Asset
Pricing Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
10.4 How Machine Learning Techniques Can Estimate Linear Factor
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
10.4.1 Gagliardini, Ossola, and Scaillet’s (2016) Econometric
Two-Pass Approach for Assessing Linear Factor Models . 349
xx Contents
10.4.2 Kelly, Pruitt, and Su’s (2019) Instrumented Principal

Components Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
10.4.3 Gu, Kelly, and Xiu’s (2021) Autoencoder . . . . . . . . . . . . . . 351
10.4.4 Kozak, Nagel, and Santosh’s (2020) Regularized
Bayesian Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
10.4.5 Which Factors to Choose and How to Deal with Weak
Factors? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
10.5 How Machine Learning Can Predict in Empirical Asset Pricing . . . 356
Appendix 1: An Upper Bound for the Sharpe Ratio . . . . . . . . . . . . . . . . . . . 359
Appendix 2: A Comparison of Different PCA Approaches . . . . . . . . . . . . . 360
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
A Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
A.2 Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
List of Contributors
Maria Victoria Anauati

Universidad de San Andres and CEDH-UdeSA, Buenos Aires, Argentina, e-mail:
vanauati@udesa.edu.ar
Wendy Brau
Universidad de San Andres and CEDH-UdeSA, Buenos Aires, Argentina, e-mail:
wbrau@udesa.edu.ar
Samuele Centorrino
Stony Brook University, Stony Brook, New York, USA, e-mail:
samuele.centorrino@stonybrook.edu
Felix Chan
Curtin University, Perth, Australia, e-mail: felix.chan@cbs.curtin.edu.au
William Crown
Brandeis University, Waltham, Massachusetts, USA, e-mail: wcrown@brandeis.edu
Marcelo Cunha Medeiros
Pontifical Catholic University of Rio de Janeiro, Rio de Janeiro, Brasil, e-mail:
mcm@econ.puc-rio.br
Jean-Pierre Florens
Toulouse School of Economics, Toulouse, France, e-mail: jean-pierre.florens@tse-
fr.eu
Mark N, Harris
Curtin University, Perth, Australia, e-mail: mark.harris@curtin.edu.au
Yu-Chin Hsu
Academia Sinica, National Central University and National Chengchi University,
Taiwan, e-mail: ychsu@econ.sinica.edu.tw
Oliver Kiss
Central European University, Budapest, Hungary and Vienna, Austria, e-mail:
xxi
xxii List of Contributors
Kiss_Oliver@phd.ceu.edu
Robert Lieli
LieliR@ceu.edu
Jean-Michel Loubes
Institut de Mathematiques de Toulouse, Toulouse, France, e-mail: loubes@math.univ-
toulouse.fr
László Mátyás
matyas@ceu.edu
Ágoston Reguly
regulyagoston@gmail.com
Gyorgy Ruzicska
Central European University, Budapest, Hungary and Vienna, Austria e-mail:
Ruzicska_Gyorgy@phd.ceu.edu
Ekaterina Seregina
Colby College, Waterville, ME, USA, e-mail: eseregin@colby.edu
Ranjodh Singh
Curtin University, Perth, Australia, e-mail: Ranjodh.Singh@curtin.edu.au
Walter Sosa-Escudero
Universidad de San Andres, CONICET and Centro de Estudios para el Desarrollo
Humano (CEDH-UdeSA), Buenos Aires, Argentina, e-mail: wsosa@udesa.edu.ar
Jantje Sönksen
Eberhard Karls University, Tübingen, Germany, e-mail: jantje.soenksen@uni-
tuebingen.de
Ben Weiern Yeo
Curtin University, Perth, Australia, e-mail: weiern.yeo@student.curtin.edu.au
Chapter 1
Linear Econometric Models with Machine
Learning
Felix Chan and László Mátyás
Abstract This chapter discusses some of the more popular shrinkage estimators in the
machine learning literature with a focus on their potential use in econometric analysis.
Specifically, it examines their applicability in the context of linear regression models.
The asymptotic properties of these estimators are discussed and the implications on
statistical inference are explored. Given the existing knowledge of these estimators, the
chapter advocates the use of partially penalized methods for statistical inference. Monte
Carlo simulations suggest that these methods perform reasonably well. Extensions of
these estimators to a panel data setting are also discussed, especially in relation to
fixed effects models.
1.1 Introduction
This chapter has two main objectives. First, it aims to provide an overview of the most
popular and frequently used shrinkage estimators in the machine learning literature,
including the Least Absolute Shrinkage and Selection Operator (LASSO), Ridge,
Elastic Net, Adaptive LASSO and Smoothly Clipped Absolute Deviation (SCAD).
The chapter covers their definitions and theoretical properties. Then, the usefulness
of these estimators is explored in the context of linear regression models from the
perspective of econometric analysis. While some of these shrinkage estimators, such
as, the Ridge estimator as proposed by Hoerl and Kennard (1970b), have a long
history in Econometrics, the evolution of the shrinkage estimators has become one
of the main focuses in the development of machine learning techniques. This is
partly due to their excellent results in obtaining superior predictive models when
the number of covariates (explanatory variables) is large. They are also particularly
Felix Chan B
László Mátyás
Central European University, Budapest, Hungary and Vienna, Austria, e-mail: matyas@ceu.edu
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1

F. Chan, L. Mátyás (eds.), Econometrics with Machine Learning, Advanced Studies
in Theoretical and Applied Econometrics 53, https://doi.org/10.1007/978-3-031-15149-1_1
2 Chan and Mátyás
useful when traditional estimators, such as the Ordinary Least Squares (OLS), are
no longer feasible e.g., when the number of covariates is larger than the number of
observations. In these cases and, in the absence of any information on the relevance
of each covariate, shrinkage estimators provide a feasible approach to potentially
identify relevant variables from a large pool of covariates. This feature highlights the
fundamental problem in sparse regression, i.e., a linear regression model with a large
parameter vector that potentially contains many zeros. The fundamental assumption
here is that, while the number of covariates is large, perhaps much larger than the
number of observations, the number of associated non-zero coefficients is relatively
small. Thus, the fundamental problem is to identify the non-zero coefficients.
While this seems to be an ideal approach to identify economic relations, it is
important to bear in the mind that the fundamental focus of shrinkage estimators
is to construct the best approximation of the response (dependent) variable. The
interpretation and statistical significance of the coefficients do not necessarily play an
important role from the perspective of machine learning. In practice, a zero coefficient
may manifest itself from two different scenarios in the context of shrinkage estimators:
(i) its true value is 0 in the data generating process (DGP) or (ii) the true value
of the coefficient is close enough to 0 that shrinkage estimators cannot identify its
importance, e.g., because of the noise in the data. The latter is related to the concept of
uniform signal strength, which is discussed in Section 1.4. Compared to conventional
linear regression analysis in econometrics, zero coefficients are often inferred from
statistical inference procedures, such as the 𝐹-t or 𝑡-tests. An important question is
whether shrinkage estimators can provide further information to improve conventional
statistical inference typically used in econometric analysis. Putting this into a more
general framework, machine learning often views data as ‘pure information’, while in
econometrics the signal to noise ratio plays an important role. When using machine
learning techniques in econometrics this ‘gap’ has to be bridged in some way.
In order to address this issue, this chapter explores the possibility of conducting
valid statistical inference for shrinkage estimators. While this question has received
increasing attention in recent times, it has not been the focus of the literature.
This, again, highlights the fundamental difference between machine learning and
econometrics in the context of linear models. Specifically, machine learning tends
to focus on producing the best approximation of the response variable, while
econometrics often focuses on the interpretations and the statistical significance of
the coefficient estimates. This clearly highlights the above gap in making shrinkage
estimators useful in econometric analysis. This chapter explores this gap and provides
an evaluation of the scenarios in which shrinkage estimators can be useful for
econometric analysis. This includes the overview of the most popular shrinkage
estimators and their asymptotic behaviour, which is typically analyzed in the form of
the so-called Oracle Properties. However, the practical usefulness of these properties
seems somewhat limited as argued by Leeb and Pötscher (2005) and Leeb and
Pötscher (2008).
Recent studies show that valid inference may still be possible on the subset
of the parameter vector that is not part of the shrinkage. This means shrinkage
estimators are particular useful in identifying variables that are not relevant from
1 Linear Econometric Models with Machine Learning 3
an economics/econometrics interpretation point of view. Let us call them control

variables, when the number of such potential variables is large, especially when larger
than the number of observations. In this case, one can still conduct valid inference on
the variables of interests by applying shrinkage (regularization) on the list of potential
control variables only. This chapter justifies this approach by obtaining the asymptotic
distribution of this partially ‘shrunk’ estimator under the Bridge regularizer, which
has LASSO and Ridge as special cases. Monte Carlo experiments show that the result
may also be true for other shrinkage estimators such as the adaptive LASSO and
SCAD.
The chapter also discusses the use of shrinkage estimators in the context of fixed
effects panel data models and some recent applications for deriving an ‘optimal’
set of instruments from a large list of potentially weak instrumental variables (see
Belloni, Chen, Chernozhukov & Hansen, 2012). Following similar arguments, this
chapter also proposes a novel procedure to test for structural breaks with unknown
breakpoints.1
The overall assessment is that shrinkage estimators are useful when the number of
potential covariates is large and conventional estimators, such as OLS, are not feasible.
However, because statistical inference based on shrinkage estimators is a delicate and
technically demanding problem that requires careful analysis, users should proceed
with caution.
The chapter is organized as follows. Section 1.2 introduces some of the more
popular regularizers in the machine learning literature and discusses their properties.
Section 1.3 provides a summary of the algorithms used to obtain the shrinkage
estimates and the associated asymptotic properties are discussed in Section 1.4.
Section 1.5 provides some Monte Carlo simulation results examining the finite sample
performance of the partially penalized (regularized) estimators. Section 1.6 discusses
three econometric applications using shrinkage estimators including fixed effects
estimators with shrinkage and testing for structural breaks with unknown breakpoints.
Some concluding remarks are made in Section 1.7.
1.2 Shrinkage Estimators and Regularizers
This section introduces some of the more popular shrinkage estimators in the machine
learning literature. Interestingly, some of them have a long history in econometrics.
The goal here is to provide a general framework for the analysis of these estimators
and to highlight their connections. Note that the discussion focuses solely on linear
models which can be written as
𝑦 𝑖 = x𝑖′ 𝛽 0 + 𝑢 𝑖 , 𝑢 𝑖 ∼ 𝐷 (0, 𝜎𝑢2 ), 𝑖 = 1, . . . , 𝑁, (1.1)
1 The theoretical properties and the finite sample performance of the proposed procedure are left
for further researchers. The main objective here is to provide an example of other potential useful
applications of the shrinkage estimators in econometrics.
4 Chan and Mátyás
′
where x𝑖 = 𝑥 1𝑖 , . . . , 𝑥 𝑝𝑖 is a 𝑝 × 1 vector containing 𝑝 explanatory variables, which
in the machine learning literature, are often referred to as covariates, features or
′
predictors with the parameter vector 𝛽 0 = 𝛽10 , . . . , 𝛽 𝑝0 . The response variable
is denoted by 𝑦 𝑖 , which is often called the endogenous, or dependent, variable in
econometrics and 𝑢 𝑖 denotes the random disturbance term with finite variance, i.e.,
𝜎𝑢2 < ∞. Equation (1.1) can also be expressed in matrix form
𝑦 = X𝛽𝛽 0 + 𝑢, (1.2)

where y = (𝑦 1 , . . . , 𝑦 𝑁 ), X = x1 , . . . , x 𝑝 , and u = (𝑢 1 , . . . , 𝑢 𝑁 ).
An important deviation from the typical econometric textbook setting is that some
of the elements in 𝛽 0 can be 0 with the possibility that 𝑝 ≥ 𝑁. Obviously, the familiar
Ordinary Least Square (OLS) estimator
𝛽ˆ 𝑂𝐿𝑆 = (X′X) −1 X′y (1.3)
cannot be computed when 𝑝 > 𝑁 since the Gram Matrix, X′X, does not have full
rank in this case. Note that in the shrinkage estimator literature, it is often, if not
always, assumed that 𝑝 1 << 𝑁 < 𝑝 or 𝑝 1 << 𝑝 < 𝑁 where 𝑝 1 denotes the number of
non-zero coefficients. In other words, it is assumed that the OLS estimator could be
computed if the 𝑝 1 covariates with non-zero coefficients could be detected a priori.
At this point, it may be useful to define some terminology and notations that
would aid subsequent discussions. The size of a parameter vector 𝛽 is the number of
elements in the vector and the length of 𝛽 is the length of the vector as measured by
an assigned norm. While the 𝐿 𝛾 norm2 is the most popular family of norms, there
are others in the literature, such as the maximum norm and zero norm. The 𝐿 𝛾 norm
′
of a vector 𝛽 = 𝛽1 , . . . , 𝛽 𝑝 denoted by the notation, ||𝛽𝛽 || 𝛾 , is defined as
𝑝
! 1/𝛾
∑︁
𝛾
||𝛽𝛽 || 𝛾 = |𝛽𝑖 | 𝛾 > 0, (1.4)
𝑖=1
where |𝛽| denotes the absolute value of 𝛽. When 𝛾 = 2, the 𝐿 𝛾 norm is known
as the Euclidean Norm, or Euclidean Distance, which is perhaps more familiar to
econometricians. As a√︁simple example, let 𝛽 = (𝛽1 , 𝛽2 ), then Euclidean Norm, of 𝐿 2
norm of 𝛽 is ||𝛽𝛽 || 2 = |𝛽1 | 2 + |𝛽2 | 2 .
The idea of a shrinkage estimator is to impose a restriction on the length of the
estimated parameter vector 𝛽ˆ . Since the length of the vector is fixed, if the objective
is to construct the best approximation to the response variable, then the coefficients
corresponding to the 0 elements in 𝛽 0 should be among the first to reach 0 before the
other coefficient estimates, as these variables would not be useful in predicting 𝑦 𝑖
and setting their coefficients to 0 would allow more ‘freedom’ for the other non-zero
coefficients. In other words, the idea is to ‘shrink’ the parameter vector in order
to identify the 0 elements in 𝛽 0 . This can be framed as the following optimization
2 The literature tends to refer to this as the 𝐿 𝑝 norm, 𝛾 is used instead of 𝑝 in this chapter as 𝑝 is
used to denote the number of covariates, i.e., the size of the vector.
problem
𝛽ˆ = arg min 𝑔 (𝛽𝛽 ; y, X) (1.5)

𝛽
𝛼 ) ≤ 𝑐,
s.t. 𝑝 (𝛽𝛽 ;𝛼 (1.6)
where 𝑔 (𝛽𝛽 ; y, X) denotes the objective (loss) function, 𝑝 (𝛽𝛽 ;𝛼 𝛼 ) denotes a function
that can be used to regulate the total length of the estimated parameter vector and is
often called the regularizer in the machine learning literature. However, it can also be
interpreted as the penalty function, which may be more familiar to econometricians.
As indicated in the constraint, the total length of the estimated parameter vector
is bounded by the constant 𝑐 > 0, which has to be determined (or selected) by the
researchers a priori. Unsurprisingly, this choice is of fundamental importance as it
affects the ability of the shrinkage estimator to correctly identify those coefficients
whose value is indeed 0. If 𝑐 is too small, then it is possible that coefficients with
a sufficiently small magnitude are incorrectly identified as coefficients with zero
value. Such mis-identification is also possible when the associated variables are noisy,
such as those suffering from measurement errors. In addition to setting zero values
incorrectly, small 𝑐 may also induce substantial bias to the estimates of non-zero
coefficients.
In the case of a least squares type estimator, 𝑔 (𝛽𝛽 ; y, X) = (y − X𝛽𝛽 ) ′ (y − X𝛽𝛽 ) but
there can also be other objective functions, such as a log-likelihood function in the case
of non-linear models or a quadratic form typically seen in the Generalized Method
of Moments estimation. Unless otherwise stated, in this chapter the focus is solely
on the least squares loss. Different regularizers, i.e., different definitions of 𝑝(𝛽𝛽 ;𝛼 𝛼 ),
lead to different shrinkage estimators, including the Least Absolute Shrinkage and
Selection Operator (LASSO) i.e., 𝑝(𝛽𝛽 ) = 𝑝𝑗=1 |𝛽 𝑗 |, the Ridge Estimator i.e., 𝑝(𝛽𝛽 ) =
Í
Í𝑝 2 Í 𝑝 2
𝑗=1 |𝛽 𝑗 | , Elastic Net i.e., 𝑝(𝛽 ) = 𝑗=1 𝛼|𝛽 𝑗 | + (1 − 𝛼)|𝛽 𝑗 | , Smoothly Clipped
𝛽
Absolute Deviation (SCAD), and other regularizers. The theoretical properties of
some of these estimators are discussed later in the chapter, including their Oracle
Properties and the connection of these properties to the more familiar concepts such
as consistency and asymptotic distributions.
The optimization as defined in Equations (1.5) – (1.6) can be written in its
Lagrangian form as
𝛽ˆ = arg min 𝑔 (𝛽𝛽 ; y, X) + 𝜆𝑝 (𝛽𝛽 ;𝛼

𝛼) . (1.7)
𝛽
A somewhat subtle difference between the Lagrangian as defined Equation (1.7)

is that the Lagrange multiplier, 𝜆, is fixed by the researcher, rather than being a
choice variable along with 𝛽 . This reflects the fact that 𝑐, the length of the parameter
vector, is fixed by the researcher a priori before the estimation procedure. It can be
shown that there is a one-to-one correspondence between 𝜆 and 𝑐 with 𝜆 being a
decreasing function of 𝑐. This should not be surprising if one interprets 𝜆 as the
penalty induced by the constraints. In the extreme case that 𝜆 → 0 as 𝑐 → ∞, Equation
6 Chan and Mátyás
(1.7) approaches the familiar OLS estimator, under the assumption that 𝑝 < 𝑁.3 Since
𝜆 is pre-determined, it is often called the tuning parameter. In practice, 𝜆 is often
selected by cross validation which is discussed further in Section 1.3.2.
1.2.1 𝑳𝜸 norm, Bridge, LASSO and Ridge
A particularly interesting class of regularizers is called the Bridge estimator as defined

by Frank and Friedman (1993), which proposed the following regularizer in Equation
(1.7)
𝑝
∑︁
𝛾 ∈ R+ .
𝛾
𝑝(𝛽𝛽 ; 𝛾) = ||𝛽𝛽 || 𝛾 = |𝛽 𝑗 | 𝛾 , (1.8)
𝑗=1
The Bridge estimator encompasses at least two popular shrinkage estimators as special
cases. When 𝛾 = 1, the Bridge estimator becomes the Least Absolute Shrinkage and
Selection Operator (LASSO) as proposed by Tibshirani (1996) and when 𝛾 = 2, tbe
Bridge estimator becomes the Ridge estimator as defined by Hoerl and Kennard
(1970b, 1970a). Perhaps more importantly, the asymptotic properties of the Bridge
estimator were examined by Knight and Fu (2000) and subsequently, the results shed
light on the asymptotic properties of both the LASSO and Ridge estimators. This is
discussed in more details in Section 1.4.
As indicated in Equation (1.8), the Bridge uses the 𝐿 𝛾 norm as the regularizer. This
leads to the interpretation that LASSO (𝛾 = 1) regulates the length of the coefficient
vector using the 𝐿 1 (absolute) norm while Ridge (𝛾 = 2) regulates the length of the
coefficient vector using the 𝐿 2 (Euclidean) norm. An advantage of the 𝐿 1 norm i.e.,
LASSO, is that it can produce estimates with exactly zero values, i.e., elements in 𝛽ˆ
can be exactly 0, while the 𝐿 2 norm, i.e., Ridge, does not usually produce estimates
with values that equal exactly 0. Figure 1.1 illustrates the difference between the three
regularizers for 𝑝 = 2. Figure 1.1a gives the plot of LASSO when |𝛽1 | + |𝛽2 | = 1 and
as indicated in the figure, if one of the coefficients is in fact zero, then it is highly
likely that the contour of the least squares will intersect with one of the corners first
and thus identifies the appropriate coefficient as 0.
In contrast, the Ridge contour does not have the ‘sharp’ corner as indicated in
Figure 1.1b and hence the likelihood of reaching exactly 0 in one of the coefficients
is low even if the true value is 0. However, the Ridge does have a computational
advantage over other variations of the Bridge estimator. When 𝛾 = 2, there is a close
form solution, namely
𝛽ˆ 𝑅𝑖𝑑𝑔𝑒 = (X′X + 𝜆I) −1 X′y. (1.9)
3 While the optimization problem for shrinkage estimator approaches the optimization problem for
OLS regardless on the relative size between 𝑝 and 𝑁 , the solution of the latter does not exist when
𝑝 > 𝑁.
(a) LASSO (b) Ridge
(c) Elastic Net
Fig. 1.1: Contour plots of LASSO, Ridge and Elastic Net
When 𝛾 ≠ 2, there is no closed form solution to the associated constrained optimization

problems so it must be solved numerically. The added complexity is that when 𝛾 ≥ 1,
the regularizer is a convex function, which means a whole suite of algorithms is
available for solving the optimization as defined in Equations (1.5) and (1.6), at
least in the least squares case. When 𝛾 < 1, the regularizer is no longer convex
and algorithms for solving this problem are less straightforward. Interestingly, this
also affects the asymptotic properties of the estimators (see Knight & Fu, 2000).
Specifically, the asymptotic distributions are different when 𝛾 < 1 and 𝛾 ≥ 1. This is
discussed briefly in Section 1.4.
1.2.2 Elastic Net and SCAD
The specification of the regularizer can be more general than a norm. One example
is Elastic Net as proposed by Zou and Hastie (2005), which is a linear combination
between 𝐿 1 and 𝐿 2 norms. Specifically,
𝑝(𝛽𝛽 ; 𝛼) = 𝛼1 ||𝛽𝛽 || 1 + 𝛼2 ||𝛽𝛽 || 22 , 𝛼 ∈ [0, 1]. (1.10)
Clearly, Elastic Net has both LASSO and Ridge as special cases. It reduces to the
former when (𝛼1 , 𝛼2 ) = (1, 0) and the latter when (𝛼1 , 𝛼2 ) = (0, 1). The exact value
8 Chan and Mátyás
of (𝛼1 , 𝛼2 ) is to be determined by the researchers, along with 𝜆. Thus, Elastic Net

requires more than one tuning parameter. While these can be selected via cross
validation (see, for example, Zou & Hastie, 2005), a frequent choice is 𝛼2 = 1 − 𝛼1
with 𝛼1 ∈ [0, 1]. In this case, the Elastic Net is an affine combination of the 𝐿 1 and
𝐿 2 regularizers which reduces the number of tuning parameters. The motivation of
Elastic Net is to overcome certain limitations of the LASSO by striking a balance
between LASSO and the Ridge. Figure 1.1c contains the contour of Elastic Net. Note
that the contour is generally smooth but there is a distinct corner in each of the four
cases when one of the coefficients is 0. As such, Elastic Net also has the ability to
identify coefficients with 0 values. However, unlike the LASSO, Elastic Net can select
more than one variable from a group of highly correlated covariates.
An alternative to LASSO is Smoothly Clipped Absolute Deviation regularizer
as proposed by Fan and Li (2001). The main motivation of SCAD is to develop a
regularizer that satisfies the following three conditions:
1. Unbiasedness. The resulting estimates should be unbiased, or at the very least,
nearly unbiased. This is particularly important when the true unknown parameter
is large with a relatively small 𝑐.
2. Sparsity. The resulting estimator should be a thresholding rule. That is, it satisfies
the role of a selector by setting the coefficient estimates of all ‘unnecessary’
variables to 0.
3. Continuity. The estimator is continuous in data.
Condition 1 is to address a well-known property of LASSO, namely that it often
produces biased estimates. Under the assumption that X′X = I, Tibshirani (1996)
showed the following relation between LASSO and OLS
𝛽ˆ 𝐿 𝐴𝑆𝑆𝑂,𝑖 = sgn 𝛽ˆ𝑂𝐿𝑆,𝑖 | 𝛽ˆ𝑂𝐿𝑆,𝑖 | − 𝜆 ,

(1.11)
where sgn 𝑥 denotes the sign of 𝑥. The equation above suggests that the greater 𝜆
is (or the smaller 𝑐 is) the larger is the bias in LASSO, under the assumption that
the OLS is unbiased or consistent. It is worth noting that the distinction between
unbiased and consistent is not always obvious in the machine learning literature.
For the purposes of the discussion in this chapter, Condition 1 above is treated as
consistency as typically defined in the econometric literature. Condition 2 in this
context refers to the ability of a shrinkage estimator to produce estimates that are
exactly 0. While LASSO satisfies this condition, Ridge in generally does not produce
estimates that are exactly 0. Condition 3 is a technical condition that is typically
assumed in econometrics to ensure the continuity of the loss (objective) function and
is often required to prove consistency.
Conditions 1 and 2 are telling. In the language of conventional econometrics and
statistics, if an estimator is consistent, then it should automatically satisfy these two
conditions, at least asymptotically. The introduction of these conditions suggests
that shrinkage estimators are not generally consistent, at least not in the traditional
sense. In fact, while LASSO satisfies sparsity, it is not unbiased. Equation 1.11 shows
that LASSO is a shifted OLS estimator when X′X = I. Thus, if OLS is unbiased
or consistent, then LASSO will be biased (or inconsistent) with the magnitude of
the bias determined by 𝜆, or equivalently, 𝑐. This should not be a surprise, since if

𝑐 < ||𝛽𝛽 || 1 then it is obviously not possible for 𝛽ˆ 𝐿 𝐴𝑆𝑆𝑂 to be unbiased (or consistent)
as the total length of the estimated parameter vector is less than the total length
of the true parameter vector. Even if 𝑐 > ||𝛽𝛽 || 1 , the unbiasedness of LASSO is not
guaranteed as shown by Fan and Li (2001). The formal discussion of these properties
leads to the development of the Oracle Properties, which are discussed in Section
1.4. For now, it is sufficient to point out that a motivation for some of the alternative
shrinkage estimators is to obtain, in a certain sense, unbiased or consistent parameter
estimates, while having the ability to identify the unnecessary explanatory variable
by assigning 0 to their coefficients i.e., sparsity.
The SCAD regularizer can be written as
 |𝛽|

 if |𝛽| ≤ 𝜆,
 2𝑎𝜆|𝛽| − 𝛽2 − 𝜆2


if 𝜆 < |𝛽| ≤ 𝑎𝜆,

𝑝(𝛽 𝑗 ; 𝑎, 𝜆) = − 1)𝜆 (1.12)
 𝜆(𝑎2(𝑎

 + 1)

 if |𝛽| > 𝑎𝜆,
 2
where 𝑎 > 2. The SCAD has two interesting features. First, the regularizer is itself
a function of the tuning parameter 𝜆. Second, SCAD divides the coefficient into
three different regions namely, |𝛽 𝑗 | ≤ 𝜆, 𝜆 < |𝛽 𝑗 | < 𝑎𝜆 and |𝛽 𝑗 | ≥ 𝑎𝜆. When |𝛽 𝑗 |
is less than the tuning parameter, 𝜆, the penalty is equivalent to the LASSO. This
helps to ensure the sparsity feature of LASSO i.e., it can assign zero coefficients.
However, unlike the LASSO, the penalty does not increase when the magnitude of
the coefficient is large. In fact, when |𝛽𝛽 𝑗 | > 𝑎𝜆 for some 𝑎 > 2, the penalty is constant.
This can be better illustrated by examining the derivative of the SCAD regularizer,
(𝑎𝜆 − |𝛽 𝑗 |)+
𝑝 ′ (𝛽 𝑗 ; 𝑎, 𝜆) = 𝐼 (|𝛽 𝑗 | ≤ 𝜆) + 𝐼 (|𝛽 𝑗 | > 𝜆), (1.13)
(𝑎 − 1)𝜆
where 𝐼 ( 𝐴) is an indicator function that equals 1 if 𝐴 is true and 0 otherwise and

(𝑥)+ = 𝑥 if 𝑥 > 0 and 0 otherwise. As shown in the expression above, when |𝛽 𝑗 | ≤ 𝜆,
the rate of change of the penalty is constant, when |𝛽 𝑗 | ∈ (𝜆, 𝑎𝜆], the penalty increases
linearly and becomes 0 when |𝛽 𝑗 | > 𝑎𝜆. Thus, there is no additional penalty when |𝛽 𝑗 |
exceeds a certain magnitude. This helps to ease the problem of biased estimates related
to the standard LASSO. Note that the derivative as shown in Equation (1.13) exists
for all |𝛽 𝑗 | > 0 including the two boundary points. Thus SCAD can be interpreted as
a quadratic spline with knots at 𝜆 and 𝑎𝜆.
Figure 1.2 provides some insight on the regularizers through their plots. As shown
in Figure 1.1a, the penalty increases as the coefficient increases. This means more
penalty is applied to coefficients with a large magnitude. Moreover, the rate of change
of the penalty equals the tuning parameter, 𝜆. Both Ridge and Elastic Net exhibit
similar behaviours for large coefficients as shown in Figures 1.1b and 1.1c but they
behave differently for coefficients close to 0. In case of the Elastic Net, the rate of
change of the penalty for coefficients close to 0 is larger than in the case of the Ridge,
which makes it more likely to push small coefficients to 0. In contrast, the SCAD
10 Chan and Mátyás
is quite different from the other penalty functions. While it behaves exactly like the
LASSO when the coefficients are small, the penalty is a constant for coefficients
with large magnitude. This means once coefficients exceed a certain limit, there is
no additional penalty imposed, regardless how much larger the coefficients are. This
helps to alleviate the bias imposed on large coefficients, as in the case of LASSO.
(a) LASSO Penalty (b) Ridge Penalty
(c) Elastic Net Penalty (d) SCAD Penalty
Fig. 1.2: Penalty Plots for LASSO, Ridge, Elastic Net and SCAD
1.2.3 Adaptive LASSO
While LASSO does not possess the Oracle Properties in general, a minor modification
of it can lead to a shrinkage estimator with Oracle Properties, while also satisfying
sparsity and continuity. The Adaptive LASSO (adaLASSO) as proposed in Zou
(2006) can be defined as
𝑝
∑︁
𝛽ˆ 𝑎𝑑𝑎 = arg min 𝑔(𝛽𝛽 ; X, Y) + 𝜆 𝑤 𝑗 |𝛽 𝑗 |, (1.14)
𝛽 𝑗=1
where 𝑤 𝑗 > 0 for all 𝑗 = 1, . . . , 𝑝 are weights to be pre-determined by the researchers.

As shown by Zou (2006), an appropriate data-driven determination of 𝑤 𝑗 would
lead to adaLASSO with Oracle ′ Properties. The term adaptive reflects the fact the
weight vector w = 𝑤 1 , . . . , 𝑤 𝑝 is based on any consistent estimator of 𝛽 . In other
words, the adaLASSO takes the information provided by the consistent estimator
and allocates significantly more penalty to the coefficients that are close to 0. This
is often achieved by assigning ŵ = 1./| 𝛽ˆ | 𝜂 , where 𝜂 > 0, ./ indicates element-wise
division and |𝛽𝛽 | denotes the element-by-element absolute value operation. 𝛽ˆ can be
chosen based on any consistent estimator of 𝛽 . OLS is a natural choice under standard
assumptions but it is only valid when 𝑝 < 𝑁. Note that using a consistent estimator to
construct the weight is a limitation particularly for the case 𝑝 > 𝑁, where consistent
estimator is not always possible to obtain. In this case, LASSO can actually be used
to construct the weight but suitable adjustments must be made for the case when
𝛽ˆ 𝐿 𝐴𝑆𝑆𝑂, 𝑗 = 0 (see Zou, 2006 for one possible adjustment).
Both LASSO and adaLASSO have been widely used and extended in various
settings in recent times, especially for time series applications in terms of lag order
selection. For examples, see, Wang, Li and Tsai (2007), Hsu, Hung and Chang
(2008) and Huang, Ma and Zhang (2008). Two particularly interesting studies
are by Medeiros and Mendes (2016), and Kock (2016). The former extended the
adaLASSO for time series models with non-Guassian and conditional heteroskedastic
errors, while the latter established the validity of using adaLASSO with non-stationary
time series data.
A particularly convenient feature of the adaLASSO in the linear regression setting
is that under ŵ = 1./|𝛽𝛽 | 𝜂 , the estimation can be transformed into a standard LASSO
problem. Thus, adaLASSO imposes no additional computational cost other than
obtaining initial consistent estimates. To see this, rewrite Equation (1.14) using the
least squares objective
𝛽ˆ 𝑎𝑑𝑎 = arg min (y − X𝛽𝛽 ) ′ (y − X𝛽𝛽 ) + 𝜆w′ |𝛽𝛽 |, (1.15)

𝛽
′
since 𝑤 𝑗 > 0 for all 𝑗 = 1, . . . , 𝑝, w′ |𝛽𝛽 |= |w′ 𝛽 |. Define 𝜃 = 𝜃 1 , . . . , 𝜃 𝑝 with 𝜃 𝑗 = 𝑤 𝑗 𝛽 𝑗
for all 𝑗 and Z = x1 /𝑤 1 , . . . , x 𝑝 /𝑤 𝑝 is a 𝑁 × 𝑝 matrix transforming each column in
X by dividing the appropriate element in w. Thus, adaLASSO can now be written as
𝑝
∑︁
𝜃ˆ 𝑎𝑑𝑎 = arg min (y − Z𝜃𝜃 ) ′ (y − Z𝜃𝜃 ) + 𝜆 |𝜃 𝑗 |, (1.16)
𝜃 𝑗=1
which is a standard LASSO problem with 𝛽ˆ 𝑗 = 𝜃ˆ 𝑗 /𝑤 𝑗 .
1.2.4 Group LASSO
There are frequent situations when the interpretation of the coefficients make sense
only if all of them in a subset of variables are non-zero. For example, if an explanatory
variable is a categorical variable (or factor) with 𝑀 options, then a typical approach
is to create 𝑀 dummy variables,4 each representing a single category. This leads

to 𝑀 columns in X and 𝑀 coefficients in 𝛽 . The interpretation of the coefficients
can be problematic if some of these coefficients are zeros, which often happens in
the case of LASSO, as highlighted by Yuan and Lin (2006). Thus, it would be more
appropriate to ‘group’ these coefficients together to ensure that the sparsity happens
at the categorical variable (or factor) level, rather than at the individual dummy
variables level. One way to capture this is to rewrite Equation (1.2) as
𝐽
∑︁
y= X 𝑗 𝛽 0 𝑗 + u, (1.17)
𝑗=1

where X 𝑗 = x 𝑗1 , . . . , x 𝑗 𝑀 𝑗 and 𝛽 0 𝑗 = 𝛽0 𝑗1 , . . . , 𝛽0 𝑗 𝑀 𝑗 . Let X = (X1 , . . . , X 𝐽 ) and
′
𝛽 = 𝛽 1′ , . . . , 𝛽 ′𝐽 , the group LASSO as proposed by Yuan and Lin (2006) is defined as
𝐽
∑︁
𝛽ˆ 𝑔𝑟 𝑜𝑢 𝑝 = arg min (y − X𝛽𝛽 ) ′ (y − 𝛽 X) + 𝜆 ||𝛽𝛽 𝑗 || 𝐾 𝑗 , (1.18)
𝛽 𝑗=1
where the penalty function is the root of the quadratic form

1/2
||𝛽𝛽 𝑗 || 𝐾 𝑗 = 𝛽 ′𝑗 𝐾 𝑗 𝛽 𝑗 𝑗 = 1, . . . , 𝐽 (1.19)
for some positive semi-definite matrix 𝐾 𝑗 to be chosen by the researchers. Note that
when 𝑀 𝑗 = 1 with 𝐾 𝑗 = I for all 𝑗, then group LASSO is reduced into a standard
LASSO. Clearly, it is possible for 𝑀 𝑗 = 1 and 𝐾 𝑗 = I for some 𝑗 only. Thus, group
LASSO allows the possibility of mixing categorical variables (or factors) with
continuous variables. Intuitively, the construction of the group LASSO imposes a 𝐿 2
norm, like the Ridge regularizer, to the coefficients that are being ‘grouped’ together,
while imposing an 𝐿 1 norm, like the LASSO, to each of the coefficients of the
continuous variables and the collective coefficients of the categorical variables. This
helps to ensure that if all categorical variables are relevant, all associated coefficients
are likely to have non-zero estimates.
While SCAD and adaLASSO may have more appealing theoretical properties,
LASSO, Ridge and Elastic Net remain popular due to their computational convenience.
Moreover, LASSO, Ridge and Elastic Net have been implemented in several software
packages and programming languages, such as R, Python and Julia, which also
explains their popularity. Perhaps more importantly, these routines are typically
capable of identifying appropriate tuning parameters, i.e., 𝜆, which makes them more
appealing to researchers. In general, the choice of regularizers is still an open question
as the finite sample performance of the associated estimator can vary with problems.
For further discussion and comparison see Hastie, Tibshirani and Friedman (2009).
4 Assuming there is no intercept.

1.3 Estimation
This section provides a brief overview of the computation of various shrinkage

estimators discussed earlier, including the determination of the tuning parameters.
1.3.1 Computation and Least Angular Regression
It should be clear that each shrinkage estimator introduced above is a solution to

a specific (constrained) optimization problem. With the exception of the Ridge
estimator, which has a closed form solution, some of these optimization problems
are difficult to solve in practice. The popularity of the LASSO is partly due to its
computation convenience via the Least Angular Regression (LARS, proposed by
Efron, Hastie, Johnstone & Tibshirani, 2004). In fact, LARS turns out to be so flexible
and powerful that the solutions to most of the regularization problems above can be
solved using a variation of LARS. This applies also to regularizers such as SCAD,
where it is a nonlinear function. As shown by Zou and Li (2008), it is possible
to obtain SCAD estimates by using LARS with local linear approximation. The
basic idea is to approximate Equations (1.7) and (1.12) using Taylor approximation,
which gives a LASSO type problem. Then, iteratively solve the associated LASSO
problem until convergence. Interestingly, in the context of linear models, the number
of iterations required until convergence is often a single step! This greatly facilitates
the estimation process using SCAD.
There are some developments of LARS focusing on improving the objective
function by including an additional variable to better the fitted value of y, ŷ. In
other words, the algorithm focuses on constructing the best approximation of the
response variable, ŷ, the accuracy of which is measured by the objective function,
rather than the coefficient estimates, 𝛽ˆ . This, once again, highlights the difference
between machine learning and econometrics. The former focuses on the predictions
of y while the latter also focuses on the coefficient estimate 𝛽ˆ . This can also be seen
via the determination of the tuning parameter 𝜆 which is discussed in Section 1.3.2.
The outline of the LARS algorithm can be found below:5
Step 1. Standardize the predictors to have mean zero and unit norm. Start with the
residual û = y − ȳ, where ȳ denotes the sample mean of y with 𝛽 = 0.
Step 2. Find the predictor 𝑥 𝑗 most correlated with û.
Step 3. Move 𝛽 𝑗 from 0 towards its least-squares coefficient until some other
predictors, 𝑥 𝑘 has as much correlation with the current residuals as does 𝑥 𝑗 .
Step 4. Move 𝛽 𝑗 and 𝛽 𝑘 in the direction defined by their joint least squares coefficient
of the current residual on (𝑥 𝑗 , 𝑥 𝑘 ) until some other predictor, 𝑥𝑙 , has as much
correlation with the current residuals.
Step 5. Continue this process until all 𝑝 predictors have been entered. After min(𝑁 −
1, 𝑝) steps, this arrives at the full least-squares solution.
5 This has been extracted from Hastie et al. (2009)

The computation of LASSO requires a small modification to Step 4 above, namely:

4a. If a non-zero coefficient hits zero, drop its variable from the active set of variables
and recompute the current joint least squares direction.
This step allows variables to join and leave the selection as the algorithm progresses
and, thus, allows a form of ‘learning’. In other words, every step revises the current
variable selection set, and if certain variables are no longer required, the algorithm
removes them from the selection. However, such variables can ‘re-enter’ the selection
set at later iterations.
Step 5 above also implies that LARS will produce at most 𝑁 non-zero coefficients.
This means if the intercept is non-zero, it will identify at most 𝑁 − 1 covariates with
non-zero coefficients. This is particularly important in the case when 𝑝 1 > 𝑁 and
LARS cannot identify more than 𝑁 relevant covariates. The same limitation is likely
to be true for any algorithms but a formal proof of this claim is still lacking and could
be an interesting direction for future research.
As mentioned above, LARS has been implemented in most of the popular open
source languages, such as R, Python and Julia. This implies LASSO and any related
shrinkage estimators that can be computed in the form of a LASSO problem can be
readily calculated in these packages.
LARS is particularly useful when the regularizer is convex. When the regularizer
is non-convex, such as the case of SCAD, it turns out that it is possible to approximate
the regularizer via local linear approximation as shown by Zou and Li (2008). The
idea is to transform the estimation problem into a sequence of LARS, which then can
be conducted iteratively.
1.3.2 Cross Validation and Tuning Parameters
The discussion so far has assumed that the tuning parameter, 𝜆, is given. In practice,
𝜆 is often obtained via K-folds cross validation. This approach yet again highlights
the difference between machine learning and econometrics, where the former focuses
pre-dominantly on the prediction performance of y. The basic idea of cross validation
is to divide the sample randomly into 𝐾 partitions and randomly select 𝐾 − 1 partitions
to estimate the parameters. The estimated parameters can then be used to construct
predictions for the remaining (unused) partition, called the left-out partition, and the
average prediction errors are computed based on a given loss function (prediction
criterion) over the left-out partition. The process is then repeated 𝐾 times each with
a different left-out partition. The tuning parameter, 𝜆, is chosen by minimizing the
average prediction errors over the 𝐾 folds. This can be summarized as follows:
Step 1. Divide the dataset into 𝐾 partitions randomly such that D = ⋓ 𝑘=1
𝐾 D .
𝑘
Step 2. Let ŷ 𝑘 be the prediction of y in D 𝑘 based on the parameter estimates from
the other 𝐾 − 1 partitions.
Step 3. The total prediction error for a given 𝜆 is
∑︁
𝑒 𝑘 (𝜆) = (𝑦 𝑖 − 𝑦ˆ 𝑘𝑖 ) 2 .
𝑖 ∈D 𝑘
Step 4. For a given 𝜆, the average prediction errors over the 𝐾-folds is
𝐾
∑︁
𝐶𝑉 (𝜆) = 𝐾 −1 𝑒 𝑘 (𝜆).
𝑘=1
Step 5. The tuning parameter can then be chosen based on
𝜆ˆ = arg min 𝐶𝑉 (𝜆).

𝜆
The process discussed here is known to be unstable for moderate sample sizes. In
order to ensure robustness, the 𝐾-fold process can be repeated 𝑁 − 1 times and the
tuning parameter, 𝜆, can be obtained as the average of these repeated 𝐾-fold cross
validations.
It is important to note that the discussion in Section 1.2 explicitly assumed 𝑐 is
fixed and by implication, this means 𝜆 is also fixed. In practice, however, 𝜆 is obtained
via statistical procedures such as cross validation introduced above. The implication
is that 𝜆 or 𝑐 should be viewed as a random variable in practice, rather than being
fixed. This impacts on the properties of the shrinkage estimators but to the best of the
authors’ knowledge, this issue has yet to be examined properly in the literature. Thus,
this would be another interesting avenue for future research.
The methodology above explicitly assumes that the data are independently distrib-
uted. While this may be reasonable in a cross section setting, it is not always valid for
time series data, especially in terms of autoregressive models. It may also be problem-
atic in a panel data setting with time effects. In those cases, the determination of 𝜆 is
much more complicated. It often reverts to evaluating some forms of goodness-of-fit
via information criteria for different values of 𝜆. For examples of such approaches,
see Wang et al. (2007), Zou, Hastie and Tibshirani (2007) and Y. Zhang, Li and
Tsai (2010). In general, if prediction is not the main objective of a study, then these
approaches can also be used to determine 𝜆. See also, Hastie et al. (2009) and Fan,
Li, Zhang and Zou (2020) for more comprehensive treatments of cross validation.
1.4 Asymptotic Properties of Shrinkage Estimators
Valid statistical inference often relies on the asymptotic properties of estimators.

This section provides a brief overview of the asymptotic properties of the shrinkage
estimators presented in Section 1.2 and discusses their implications for statistical
inference. The literature in this area is highly technical, but rather than focusing
on these aspects, the focus here is on the extent to which these can facilitate valid
statistical inference typically employed in econometrics, with an emphasis on the
qualitative aspects of the results (see the references for technical details).
The asymptotic properties in the shrinkage estimators literature for linear models
can broadly be classified into three focus areas, namely:
1. Oracle Properties,
2. Asymptotic distribution of shrinkage estimators, and
3. Asymptotic properties of estimators for parameters that are not part of the
shrinkage.
1.4.1 Oracle Properties
The asymptotic properties of the shrinkage estimators presented above are often
investigated through the so-called Oracle Properties. Although the origin of the term
Oracle Properties can be traced back to Donoho and Johnstone (1994), Fan and
Li (2001) are often credited as the first to formalize its definition mathematically.
Subsequent presentation can be found in Zou (2006) and Fan, Xue and Zou (2014),
with the latter having possibly the most concise definition to-date. The term Oracle
is used to highlight the feature that Oracle estimator shares the same properties as
estimators with the correct set of covariates. In other words, the Oracle estimator can
‘foresee’ the correct set of covariates. While presentations can be slightly different,
the fundamental idea is very similar. To aid the presentation, let us rearrange the true
parameter vector, 𝛽 0 , so that all parameters with non-zero values can be grouped into
a sub-vector and all the parameters with zero values can be grouped into another
sub-vector. It is also helpful to create a set that contains the indexes of all non-zero
coefficients as well as another one that contains the indexes of all zero-coefficients.
Formally, let A = { 𝑗 : 𝛽0 𝑗 ≠ 0} and Â = { 𝑗 : 𝛽ˆ 𝑗 ≠ 0}, and without loss of generality,
′ ′ ′ ′
partition 𝛽 0 = 𝛽 ′ , 𝛽 ′ 𝑐 and 𝛽ˆ = 𝛽ˆ A , 𝛽ˆ A 𝑐 , where 𝛽 0A denotes the sub-vector
0A 0A
of 𝛽 0 containing all the non-zero elements of 𝛽 0 i.e., those with indexes that belong to
A, while 𝛽 A 𝑐 is the sub-vector of 𝛽 0 containing all the zero elements i.e., those with
indexes that do not belong to 𝛽 A . Similar definitions apply to 𝛽ˆ A and 𝛽ˆ A 𝑐 . Then the
estimator 𝛽ˆ is said to have the Oracle Properties if it has

1. Selection Consistency: lim Pr Â = A = 1, and
𝑁 →∞
√ 𝑑
2. Asymptotic normality: 𝑁 𝛽ˆ A − 𝛽 A → 𝑁 (0,Σ Σ ),
where Σ is the variance-covariance matrix of the following estimator
𝛽ˆ 𝑜𝑟 𝑎𝑐𝑙𝑒 = arg min 𝑔(𝛽𝛽 ). (1.20)

𝛽 :𝛽
𝛽 A 𝑐 =0
Equation (1.20) is called the Oracle estimator by Fan et al. (2014) because it
shares the same properties as the estimator that contains only the variables with
non-zero coefficients. A shrinkage estimator is said to have the Oracle Properties if
it is selection consistent, i.e., is able to identify a variable with zero and non-zero
coefficients, and has the same asymptotic distribution as a consistent estimator that
contains only the correct set of variables. Note that selection consistency is a weaker
condition than consistency in the traditional sense. The requirement of selection
consistency is to discriminate coefficients with zero and non-zero values but it does
not require the estimates to be consistent if they have non-zero values.
It should be clear from the previous discussions that neither LASSO nor Ridge
have Oracle Properties in general since LASSO is typically inconsistent and Ridge
does not usually have selection consistency. However, adaLASSO, SCAD and group
LASSO have been shown to possess Oracle Properties (for technical details, see Zou,
2006, Fan & Li, 2001 and Yuan & Lin, 2006, respectively). While these shrinkage
estimators possess Oracle Properties, their proofs usually rely on the following three
assumptions:
Assumption 1. u is a vector of independent, identically distributed random variables
with finite variance.
Assumption 2. There exist a matrix C with finite elements such that 𝑁 −1 𝑖=1 x𝑖 x𝑖′ −
Í𝑁
C = 𝑜 𝑝 (1).
Assumption 3. 𝜆 ≡ 𝜆 𝑁 = 𝑂 (𝑁 𝑞 ) for some 𝑞 ∈ (0, 1].
While Assumption 2 is fairly standard in the econometric literature, Assumption 1
appears to be restrictive as it does not include common issues in econometrics such
as serial correlation or heteroskedasticity. However, several recent studies, such as
Wang et al. (2007) and Medeiros and Mendes (2016) have attempted to relax this
assumption to non-Gaussian, serial correlated or conditional heteroskedastic errors.
Assumption 3 is perhaps the most interesting. First, it once again highlights the
importance of the tuning parameter, not just for the performance of the estimators, but
also for their asymptotic properties. Then, the assumption requires that 𝑁 −𝑞 𝜆 𝑁 − 𝜆 0 =
𝑜 𝑝 (1) for some 𝜆0 ≥ 0. This condition trivially holds when 𝜆 𝑁 = 𝜆 stays constant
regardless of the sample size. However, this is unlikely to be the case in practice
when 𝜆 𝑁 is typically chosen based on cross validation as described in Section 1.3.2.
Indeed, when 𝜆 𝑁 remains constant, 𝜆0 = 0, which implies that the constraint imposes
no penalty on the loss asymptotically and in the least squares case, the estimator
collapses to the familiar OLS estimator asymptotically. In the case when 𝜆 𝑁 is a
non-decreasing function of 𝑁, Assumption 3 assets an upper bound of the growth rate.
This should not be surprising since 𝜆 𝑁 is an inverse function of 𝑐. If 𝜆 𝑁 increases, the
total length of 𝛽 decreases and that may increase the amount of bias in the estimator.
Another perspective of this assumption is its relation to the uniform signal strength
condition as discussed by C.-H. Zhang and Zhang (2014). The condition asserts that
all non-zero coefficients must be greater in magnitude than an inflated level of noise
which can be expressed by
√︂
2 log 𝑝
𝐶𝜎 , (1.21)
𝑁
where 𝐶 is the inflation factor (see C.-H. Zhang and Zhang (2014)). Essentially, the
coefficients are required to be ‘large’ enough (relative to the noise) to ensure selection
consistency and this assumption has also been widely used in some of the literature
(for examples, see the references within C.-H. Zhang & Zhang, 2014). While there
seems to be an intuitive link between Assumption 3 and the uniform signal strength
condition, there does not appear to be any formal discussion about their connection
and this could be an interesting area for future research. Perhaps more importantly,
the relation between choosing 𝜆 via cross validation, Assumption 3, and the uniform
signal strength condition is still somewhat unclear. This also means that if the true
value of the coefficient is sufficiently small, there is a good chance that shrinkage
estimators will be unable to identify them and it is unclear how to statistically verify
this situation in finite samples. Theoretically, however, if the shrinkage estimator is
selection consistent, then it should be able to discriminate coefficients with a small
magnitude and coefficients with zero value.
The Oracle Properties are appealing because they seem to suggest that one can
select the right set of covariates via selection consistency, and simultaneously estimate
the non-zero parameters consistently, which are also asymptotically normal. It is
therefore tempting to
1. conduct statistical inference directly on the shrinkage estimates in the usual way
or
2. use a shrinkage estimator as a variable selector, then estimate the coefficients
using OLS on model with only the selected covariates and conduct statistical
inference on the OLS estimates in the usual way.
The last approach is particularly tempting especially for the shrinkage estimators that
satisfy selection consistency but not necessarily asymptotic normality, such as the
original LASSO. This approach is often called Post-Selection OLS or in the case of
LASSO, Post-LASSO OLS. These two approaches turn out to be overly optimistic in
practice as argued by Leeb and Pötscher (2005) and Leeb and Pötscher (2008). The
main issue has to do with the mode of convergence in proving the Oracle Properties.
In most cases, the convergence is pointwise rather than uniform. An implication of the
pointwise convergence is that a different 𝛽 0 may require different sample sizes before
the asymptotic distribution becomes a reasonable approximation. Since the sample
size is typically fixed in practice with 𝛽 0 unknown, blindly applying the asymptotic
result for inference may lead to misleading conclusion, especially if the sample size is
not large enough for the asymptotic to ‘kick in’. Worse still, it is typically not possible
to examine if the sample size is large enough since it depends on the unknown 𝛽 0 .
Thus, the main message is that while Oracle Properties are interesting theoretical
properties and provide an excellent framework to aid the understanding of different
shrinkage estimators, it does not necessary provide the assurance one wishes for in
practice for statistical inference.
1.4.2 Asymptotic Distributions
Next, let us examine the distributional properties of shrinkage estimators directly.

The seminal work of Knight and Fu (2000) derived the asymptotic distribution of the
Bridge estimator as defined in Equations (1.7) and (1.8). Under Assumptions 1 – 3
with 𝑔(𝛽𝛽 ) = (y − X𝛽𝛽 ) ′ (y − X𝛽𝛽 ), 𝑞 = 0.5 and 𝛾 ≥ 1, Knight and Fu (2000) showed that
√ 𝑑
𝑁 𝛽ˆ 𝐵𝑟𝑖𝑑𝑔𝑒 − 𝛽 0 → arg min𝑉 (𝜔 𝜔 ), (1.22)
where
𝑝
∑︁
′ ′
𝑉 (𝜔
𝜔 ) = −2𝜔 𝜔 +𝜔
𝜔 W𝜔 𝜔 + 𝜆0
𝜔 C𝜔 𝑢 𝑗 sgn 𝛽0 𝑗 |𝛽0 𝑗 | 𝛾−1 , (1.23)
𝑗=1
with W having a 𝑁 (0, 𝜎 2 C) distribution. For 𝛾 < 1, Assumption 3 needs to be

adjusted such that 𝑞 = 𝛾/2 with 𝑉 (𝜔
𝜔 ) changes to
𝑝
∑︁
𝑉 (𝜔 𝜔 ′W𝜔
𝜔 ) = −2𝜔 𝜔 ′C𝜔
𝜔 +𝜔 𝜔 + 𝜆0 |𝑢 𝑗 | 𝛾 𝐼 (𝛽0 𝑗 = 0). (1.24)
𝑗=1
Recall that the Bridge regularizer is convex for 𝛾 ≥ 1, but not for 𝛾 < 1. This is
reflected by the difference in the asymptotic distributions implied by Equations (1.23)
and (1.24). Interestingly, but perhaps not surprisingly, the main difference between
the two expressions is the term related to the regularizer. Moreover, the growth rate
of 𝜆 𝑁 is also required to be much slower for the 𝛾 < 1 case than for the 𝛾 ≥ 1 one.
Let 𝜔 ∗ = arg min𝑉 (𝜔𝜔 ), then for 𝛾 ≥ 1 it is straightforward to show that
𝜔 ∗ = C−1 W − 𝜆0 sgn 𝛽 0 |𝛽𝛽 0 | 𝛾−1 ,

(1.25)
where sgn 𝛽 , |𝛽𝛽 | 𝛾−1 and the product of the two terms are understood to be taken
element wise. This means
√ 𝑑
𝑁 𝛽ˆ 𝐵𝑟𝑖𝑑𝑔𝑒 − 𝛽 0 → C−1 W − 𝜆0 sgn 𝛽 0 |𝛽𝛽 0 | 𝛾−1 .

(1.26)
Note that setting 𝛾 = 1 and 𝛾 = 2 yield the asymptotic distribution of LASSO and
Ridge, respectively. However, as indicated in Equation (1.25), the distribution depends
on 𝛽 0 . the true parameter vector. The result has two important implications. First,
Bridge estimator is generally inconsistent and second, the asymptotic distribution
depends on the true parameter vector, which means it is subject to the criticism
of Leeb and Pötscher (2005). The latter means that the sample size required for
the asymptotic distribution to be a ‘reasonable’ approximation to the finite sample
distribution depends on the true parameter vector. Since the true parameter vector is
not typically known, this means the asymptotic distribution may not be particularly
helpful in practice, at least not in the context of hypothesis testing.
While directly statistical inference on the shrinkage estimates appears to be difficult,
there are other studies that take slightly different approaches. The two studies that are
worth mentioning are Lockhart, Taylor, Tibshirani and Tibshirani (2014) and Lee,
Sun, Sun and Taylor (2016). The former developed a test statistics for coefficients
as they enter the model during the LARS process, while the latter derived the
asymptotic distribution of the least squares estimator conditional on model selection
from shrinkage estimator, such as the LASSO.
In addition to the two studies above, it is also worth noting that in some cases, such
as LASSO, where the estimates can be bias-corrected and through such correction, the
asymptotic distribution of the bias-corrected estimator can be derived (for examples,
see C.-H. Zhang & Zhang, 2014 and Fan et al., 2020). Despite the challenges as
shown by Leeb and Pötscher (2005) and Leeb and Pötscher (2008), the asymptotic
properties of the shrinkage estimators, particularly for purposes of valid statistical
inference, remains an active area of research.
Overall, the current knowledge regarding the asymptotic properties of various
shrinkage estimators can be summarized as follows:
1. Some shrinkage estimators have shown to possess Oracle Properties, which
means asymptotically they can select the right covariates i.e., correctly assign 0
to the coefficients with true 0 value. It also means that the estimators have an
asymptotic normal distribution.
2. Despite the Oracle Properties and other asymptotic results, such as those of
Knight and Fu (2000), the practical usefulness of these results are still somewhat
limited. The sample size required for the asymptotic results to ‘kick in’ depends
on the true parameter vector, which is unknown in practice. Thus, one can never
be sure about the validity of using the asymptotic distribution for a given sample
size.
1.4.3 Partially Penalized (Regularized) Estimator
While valid statistical inference on shrinkage estimators appears to be challenging,

there are situations where the parameter of interests may not be part of the shrinkage.
This means the regularizer does not have to be applied to the entire parameter vector.
Specifically, let us rewrite Equation (1.2) as
y = X1 𝛽 1 + X2 𝛽 2 + u, (1.27)
where X1 and X2 are 𝑁 × 𝑝1 and 𝑁 × 𝑝 2 matrices such that 𝑝 = 𝑝 1 + 𝑝 2 with
′
X = [X1 , X2 ], and 𝛽 = 𝛽 1′ , 𝛽 2′ such that 𝛽 1 and 𝛽 2 are 𝑝 1 × 1 and 𝑝 2 × 1 parameter
vectors. Assume that only 𝛽 2 is sparse i.e., contains elements with zero value, and
consider the following shrinkage estimator
′ ′
′
𝛽ˆ 1 , 𝛽ˆ 2 = arg min (y − X1 𝛽 1 − X2 𝛽 2 ) ′ (y − X1 𝛽 1 − X2 𝛽 2 ) + 𝜆𝑝(𝛽𝛽 2 ). (1.28)
𝛽2
𝛽 1 ,𝛽
Note that the penalty function (regularizer) applies only to 𝛽 2 but not 𝛽 1 . A natural
question in this case is whether the asymptotic properties of 𝛽ˆ 1 could facilitate valid
statistical inference in the usual manner?
In the case of the Bridge estimator, it is possible to show that 𝛽ˆ 1 has an asymptotic
normal distribution similar to the OLS estimator. This is formalized in Proposition
1.1.
Proposition 1.1 Consider the linear model as definedÍin Equation (1.27) and the
estimator as defined in Equation (1.28) with 𝑝(𝛽𝛽 2 ) = 𝑝𝑗= 𝑝1 +1 |𝛽 𝑗 | 𝛾 for some 𝛾 > 0.
√ √
Under Assumptions 1 and 2 along with 𝜆 𝑁 / 𝑁 → 𝜆0 ≥ 0 for 𝛾 ≥ 1 and 𝜆 𝑁 / 𝑁 𝛾 →
𝜆0 ≥ 0, then
√ 𝑑
𝑁 𝛽ˆ 1 − 𝛽 01 → 𝜔∗1
where
𝜔∗1 = C−1
1 W1 ,
with W1 ∼ 𝑁 (0, 𝜎𝑢2 I) denoting a 𝑝 1 × 1 random vector and 𝐶1−1 the 𝑝 1 × 𝑝 1 matrix
consisting of the first 𝑝 1 rows and columns of C.
Proof See the Appendix. □

The implication of Proposition 1.1 is that valid inference in the usual manner
should be possible for parameters that are not part of the shrinkage process, at least
for the Bridge estimator. This is verified by some Monte Carlo experiment in Section
1.5.1. This leads to the question on whether this idea can be generalized further. For
example, consider
y = X1 𝛽 1 + u, (1.29)
where X1 is a 𝑁 × 𝑝 1 matrix containing (some) endogenous variables. Now assume
there are 𝑝 2 potential instrumental variables where 𝑝 2 can be very large. In a Two
Stage Least Squares setting, one would typically construct instrumental variables by
first estimating
X1 = X2Π + v
Π . Given the estimation of Π is separate
and set the instrumental variables Z = X2Π̂
from 𝛽 1 and perhaps more importantly, the main target is Z, which in a sense, should
be the best approximation of X1 given X2 . As shown by Belloni et al. (2012), when
𝑝 2 is large, it is possible to
1. leverage shrinkage estimators to produce the best instruments i.e., the best
approximation of X1 given X2 and
2. reduce the number of instruments given the sparse nature of shrinkage estimators
and thus alleviate the issue of well too many instrumental variables.
The main message is that it is possible to obtain a consistent and asymptotically normal
estimator for 𝛽 1 by constructing optimal instruments from shrinkage estimators using
a large number of potential instrumental variables. There are two main ingredients
which make this approach feasible. The first is that it is possible to obtain ‘optimal’
instruments Z based on Post-Selection OLS, that is, OLS after a shrinkage procedure
as shown by Belloni and Chernozhukov (2013). Given Z, Belloni et al. (2012)
show that the usual IV type estimators, such as 𝛽ˆ 𝐼𝑉 = (Z′X1 ) −1 Z′y, follow standard
asymptotic results. This makes intuitive sense, as the main target in this case is
the best approximation of X1 rather than the quality of the estimator for Π . Thus,
in a sense, this approach leverages the intended usage of shrinkage estimators of
producing optimal approximation and use this approximation as instruments to resolve

endogeneity.
This idea can be expanded further to analyze a much wider class of econometric
problems. Chernozhukov, Hansen and Spindler (2015) generalized this approach by
considering the estimation problem where the parameters would satisfy the following
system of equations
𝑀 (𝛽𝛽 1 , 𝛽 2 ) = 0. (1.30)
Note that least squares, maximum likelihood and Generalized Method of Moments
can be captured in this framework. Specifically, the system of equation as denoted
by 𝑀 can be viewed as the first order derivative of the objective function and thus
Equation (1.30) represents the First Order Necessary Condition for a wide class of
M-estimators. Along with some relatively mild assumptions, Chernozhukov et al.
(2015) show that the following:
𝜕
𝑀 (𝛽𝛽 1 , 𝛽 2 ) 𝛽 2 =𝛽ˆ = 0, (1.31)
𝜕𝛽𝛽 2′ 2
where 𝛽ˆ 2 denotes a good quality shrinkage estimator of 𝛽 2 , are sufficient to ensure

valid statistical inference on 𝛽ˆ 1 . Equation (1.31) is often called the immunization
condition. Roughly speaking, it asserts that if the system of equations as defined in
Equation (1.30) is not sensitive, and therefore immune, to small changes of 𝛽 2 for a
given estimator of 𝛽 2 , then statistical inference based on 𝛽ˆ 1 is possible.6
Using the IV example above, Π , the coefficient matrix for constructing optimal
instruments, can be interpreted as 𝛽 2 in Equation (1.30). Equation (1.31) therefore
requires the estimation of 𝛽 1 not to be sensitive to small changes in the shrinkage
estimators used to estimate Π . In other words, the condition requires that a small
Π does not affect the estimation of 𝛽 1 . This is indeed the case as long as
change in Π̂
the small changes in Π̂Π do not affect the estimation of the instruments Z significantly.
1.5 Monte Carlo Experiments
This section provides a brief overview of some of the possible applications of

shrinkage estimators in econometrics. A handful of Monte Carlo Simulation results
are also presented to shed more light on a few issues discussed in the Chapter.
6 Readers are referred to Chernozhukov et al. (2015) and Belloni and Chernozhukov (2013) for
technical details and discussions.
1.5.1 Inference on Unpenalized Parameters
As seen in Section 1.4.3, while valid inference remains challenging for Shrinkage
estimators in general, it is possible to obtain valid inference on parameters that are not
part of the shrinkage. In other words, if the parameters do not appear in the regularizer,
then inference of these parameter estimates in the case of the Bridge estimator is
asymptotically normal, as shown in Proposition 1.1. Here some Monte Carlo evidence
is provided about the finite sample performance of this estimator. In addition to
the Bridge estimator, this section also examines the finite sample properties of the
unpenalized (unregularized) parameter with other regularizers, including Elastic Net,
adaLASSO and SCAD. This may be helpful in identifying if these regularizers share
the same properties as the Bridge for the unpenalized parameters.
Unless otherwise stated, all Monte Carlo simulations in this section involve
simulating six covariates, 𝑥𝑖 , 𝑖 = 1, . . . , 6. The six covariates are generated through a
multivariate lognormal distribution, 𝑀𝑉 𝐿𝑁 (0,Γ Γ ). The semi-positive definite matrix
Γ controls the degree of correlations between the six covariates. In this section,
Γ = {𝜌𝑖 𝑗 } with
 𝜎𝑖


 if 𝑖 = 𝑗,
𝜌𝑖 𝑗 = 𝜌 𝑖+ 𝑗−2

 𝜎𝜎 if 𝑖 ≠ 𝑗,
 𝑖 𝑗
where 𝜎𝑖2 denotes the diagonal of Γ , which is assigned to be {2, 1, 1.5, 1.25, 1.75, 3}.
Three different values of 𝜌 are considered, namely 𝜌 = 0, 0.45, 0.9, which covers
cases of non-correlated, moderately correlated and highly correlated covariates. The
rationale in choosing the multivariate lognormal distribution is to allow further
investigations on the impact of variable transformations, such as the logarithmic trans-
formation frequently used in economic and econometric analyses, on the performance
of shrinkage estimators. This is the focus in Section 1.5.2.
In each case, the Monte Carlo experiment has 5000 replication over five different
sample sizes: 30, 50, 100, 500 and 1000. The collection of estimators considered
includes OLS, LASSO, Ridge, Elastic Net, adaLASSO with 𝜂 = 1 and 𝜂 = 2, and
SCAD. The tuning parameters are selected based on five-fold cross validation as
discussed in Section 1.3.2.
The first set of experiments examines the size and power of a simple 𝑡 test. The
data generating process (DGP) in this case is
𝑦 𝑖 = 𝛽1 log 𝑥1𝑖 + 2 log 𝑥2𝑖 − 2 log 𝑥 3𝑖 + 𝑢 𝑖 , 𝑢 𝑖 ∼ 𝑁 (0, 1).
Four different values of 𝛽1 are considered namely, 0, 0.1, 1 and 2. In each case, the
following model is estimated
𝑦 𝑖 = 𝛽0 + x𝑖′ 𝛽 + 𝑢 𝑖 , (1.32)
where x𝑖 = (log 𝑥 1𝑖 , . . . , log 𝑥6𝑖 ) ′ with 𝛽 = (𝛽1 , . . . , 𝛽6 ) ′. In this experiment, the coef-
ficient of log 𝑥1 , 𝛽1 , is not part of the shrinkage. In other words, the regularizers
are applied only to 𝛽2 , . . . , 𝛽6 . After the estimation, the 𝑡-test statistics for testing
𝐻0 : 𝛽1 = 0 are computed in the usual way. This means for the case 𝛽1 = 0, the
experiment examines the size of the simple 𝑡-test on 𝛽1 . In all other cases, the
experiments examine the finite sample power of the simple 𝑡-test on 𝛽1 with varying
signal-to-noise ratio.
The results of each case are summarized in Tables 1.1 – 1.4. Given that the
significance level of the test is set at 0.05, the 𝑡-test has the expected size as shown
in Table 1.1 for all estimators when the covariates are uncorrelated. However, when
the correlation between the covariates increases, the test size for the Ridge estimator
seems to be higher than expected, albeit not drastically. This is related to Theorem
4 by Knight and Fu (2000) where the authors considered the case when C (the
variance-covariance matrix of the covariates) is nearly singular. The near singularity
of C is reflected through the increasing value of 𝜌 in the experiment. As a result, a
different, more restrictive bound, on 𝜆 𝑛 is required for the asymptotic result to remain
valid in the case when 𝛾 > 1. Interestingly, the test size under all other estimators
remained reasonably close to expectation as 𝑁 increased for the values of 𝜌 considered
in the study.
Table 1.1: Size of a Simple 𝑡-test for Unpenalized Parameter
𝜌 𝑁 OLS LASSO adaLASSO adaLASSO2 Ridge Elastic Net SCAD
30 6.02 5.4 5.44 5.4 5.62 5.52 5.62

50 5.74 5.46 5.54 5.46 5.62 5.6 5.52
0 100 5.14 5.34 5.28 5.26 5.26 5.26 5.16
500 5.42 5.34 5.38 5.34 5.32 5.46 5.24
1000 5.2 5.18 5.14 5.08 5.08 5.16 5.12
30 6.04 6.02 6.12 6.22 5.8 5.94 6.06

50 5.52 5.46 5.42 5.54 5.94 5.62 5.5
0.45 100 5.8 6.16 6.0 6.16 6.84 6.14 5.98
500 5.48 5.96 5.86 5.98 6.44 5.8 7.4
1000 5.06 5.0 4.96 5.04 6.24 4.9 8.98
30 6.44 5.86 5.84 5.72 3.8 5.6 5.92

50 5.78 5.18 5.26 5.16 3.66 4.88 4.94
0.9 100 5.22 5.22 5.26 5.28 4.3 5.26 4.68
500 5.14 4.68 4.84 4.9 6.74 4.92 4.62
1000 4.76 4.26 4.22 4.2 6.08 4.4 4.92
Table 1.2: Power of a Simple 𝑡-test for Unpenalized Parameter: 𝛽1 = 2
30 100.0 100.0 100.0 100.0 100.0 100.0 100.0

50 100.0 100.0 100.0 100.0 100.0 100.0 100.0
0 100 100.0 100.0 100.0 100.0 100.0 100.0 100.0
500 100.0 100.0 100.0 100.0 100.0 100.0 100.0
1000 100.0 100.0 100.0 100.0 100.0 100.0 100.0
30 100.0 100.0 100.0 100.0 100.0 100.0 100.0

50 100.0 100.0 100.0 100.0 100.0 100.0 100.0
0.45 100 100.0 100.0 100.0 100.0 100.0 100.0 100.0
500 100.0 100.0 100.0 100.0 100.0 100.0 100.0
1000 100.0 100.0 100.0 100.0 100.0 100.0 100.0
30 98.38 99.08 98.92 98.94 99.04 98.92 98.78

50 99.98 100.0 100.0 100.0 100.0 100.0 100.0
0.9 100 100.0 100.0 100.0 100.0 100.0 100.0 100.0
500 100.0 100.0 100.0 100.0 100.0 100.0 100.0
1000 100.0 100.0 100.0 100.0 100.0 100.0 100.0
When 𝛽1 = 2, the power of the test is nearly 1 for all cases as shown in Table 1.2. As
the true value decreases, the performance of the 𝑡-test from all shrinkage estimators
remained comparable to OLS as shown in Tables 1.3 - 1.4. This provides some
evidence to support Proposition 1.1 as well as an indication that its theoretical results
may also be applicable to Adaptive LASSO, Elastic Net, and SCAD.
1.5.2 Variable Transformations and Selection Consistency
Another important issue examined in this section is the finite sample performance of
selection consistency. Specifically, this section examines the selection consistency
in the presence of correlated covariates and covariates with a different signal to
noise ratio in the form of different variable transformations as well as different
variances. The variable transformations considered include logarithmic and quadratic
transforms.
The DGP considered is
2 3
𝑦 𝑖 = 2 log 𝑥2𝑖 − 0.1 log 𝑥3𝑖 + 1.2𝑥 2𝑖 − 𝑥3𝑖 + 𝑢𝑖 , 𝑢 𝑖 ∼ 𝑁 (0, 1).
Table 1.3: Power of a Simple 𝑡-test for Unpenalized Parameter: 𝛽1 = 1
30 90.94 90.2 90.24 90.1 88.86 90.2 91.02

50 99.24 99.26 99.3 99.26 98.92 99.24 99.32
0 100 100.0 100.0 100.0 100.0 100.0 100.0 100.0
500 100.0 100.0 100.0 100.0 100.0 100.0 100.0
1000 100.0 100.0 100.0 100.0 100.0 100.0 100.0
30 84.12 85.86 85.94 85.88 86.04 85.98 85.32

50 97.8 98.4 98.5 98.46 98.66 98.34 98.46
0 100 100.0 100.0 100.0 100.0 100.0 100.0 100.0
500 100.0 100.0 100.0 100.0 100.0 100.0 100.0
1000 100.0 100.0 100.0 100.0 100.0 100.0 100.0
30 21.94 25.18 24.9 25.06 23.08 23.84 23.0

50 34.28 40.16 40.0 40.0 40.22 38.08 37.34
0 100 61.3 68.32 68.32 68.06 72.98 65.98 66.48
500 99.88 99.94 99.94 99.94 100.0 99.94 99.98
1000 100.0 100.0 100.0 100.0 100.0 100.0 100.0
The estimated model is similar to Equation (1.32) but the specification of x𝑖 changes
2 , . . . , 𝑥 2 ′ and 𝛽 = (𝛽 , . . . , 𝛽 ) ′ .
to x𝑖 = 𝑥 1𝑖 , . . . , 𝑥6𝑖 , log 𝑥1𝑖 , . . . , log 𝑥6𝑖 , 𝑥 1𝑖 2𝑖 1 18
Tables 1.5 and 1.6 contains the percentage of the replications where the estimators
identified the coefficients of log 𝑥2 and 𝑥22 as zero, respectively. Note that the DGP
suggests that neither coefficients should be zero. Thus, these results highlight the finite
sample performance of the selection consistency. Given the size of each coefficient
and the variance of 𝑥 2 , it is perhaps not surprising that shrinkage estimators generally
assigned non-zero coefficient to 𝑥22 but they tend to penalize the coefficient of log 𝑥 2
to 0 for large portion of the replications. In fact, as sample size increases, the number
of replications where zero coefficients had been assigned to log 𝑥 2 increases, but the
opposite is true for 𝑥22 . This can be explained by two factors. First, log 𝑥2 and 𝑥22 are
highly correlated and it is well known that LASSO type estimators will generally
select only one variable from a set of highly correlated covariates, see for example
Tibshirani (1996). Second, the variance of log 𝑥 2 is smaller than that of 𝑥22 , which
implies a lower signal to noise ratio. This also generally affects the ability of shrinkage
estimators to select such variables, as discussed earlier.
These results seem to suggest that one should not take selection consistency for
granted. While the theoretical results are clearly correct, the assumptions underlying
the result, specifically the assumption on the tuning parameter, 𝜆, and its relation to
Table 1.4: Power of a Simple 𝑡-test for Unpenalized Parameter: 𝛽1 = 0.1
30 12.54 11.6 11.68 11.6 11.56 12.22 12.16

50 16.22 15.56 15.86 15.88 15.44 15.9 16.18
0 100 28.48 27.88 28.1 27.92 27.24 28.06 28.08
500 87.64 87.4 87.4 87.38 86.3 87.42 87.18
1000 99.42 99.44 99.46 99.46 99.16 99.46 99.46
30 11.18 12.54 12.26 12.4 13.54 12.48 12.0

50 14.9 17.58 17.78 17.42 19.62 17.4 16.7
0.45 100 23.9 28.6 28.54 28.42 33.94 27.86 28.58
500 77.3 82.26 81.94 81.96 93.02 81.2 87.6
1000 97.12 98.2 98.14 98.08 99.88 98.0 99.38
30 7.8 7.74 7.94 7.78 5.58 7.32 7.68

50 6.78 7.42 7.44 7.24 6.46 6.96 6.8
0.9 100 7.84 9.1 9.24 9.28 10.14 8.7 8.54
500 17.56 22.62 22.36 22.38 42.82 20.8 25.5
1000 31.12 38.32 38.1 38.12 74.44 35.64 48.08
the signal to noise ratio in the form of uniform signal strength condition clearly plays
an extremely important role in variable selection using shrinkage estimators. Thus,
caution and further research in this area seems warranted.
1.6 Econometrics Applications
Next, some recent examples of econometric application using shrinkage estimators

are provided. The first application examines shrinkage estimators in a distributed lag
model setting. The second example discusses the use of shrinkage estimators in panel
data with fixed effects, and the third one proposes a new test for structural breaks
with unknown break points.
Table 1.5: Percentage of Replications Excluding log 𝑥2
30 0.0 52.24 52.22 52.26 0.0 46.82 79.58

50 0.0 56.44 56.44 56.44 0.0 50.56 86.48
0 100 0.0 64.3 64.3 64.3 0.0 62.98 94.16
500 0.0 88.08 88.08 88.08 0.0 89.16 99.98
1000 0.0 95.02 95.02 95.02 0.0 96.26 100.0
30 0.0 50.08 50.18 50.22 0.0 47.9 79.74

50 0.0 56.0 55.96 56.0 0.0 53.7 85.8
0.45 100 0.0 64.08 64.08 64.08 0.0 63.72 94.34
500 0.0 89.88 89.88 89.88 0.0 91.28 99.96
1000 0.0 96.64 96.64 96.64 0.0 96.7 100.0
30 0.0 47.46 47.68 47.66 0.0 47.06 73.68

50 0.0 53.06 53.34 53.12 0.0 54.16 79.42
0.9 100 0.0 65.44 65.5 65.42 0.0 64.94 89.76
500 0.0 93.98 93.98 93.98 0.0 92.38 99.62
1000 0.0 98.82 98.82 98.82 0.0 97.62 99.94
1.6.1 Distributed Lag Models
The origin of the distributed lag model can be traced back to Tinbergen (1939).
While there have been studies focusing on lag selection in an Autoregressive-Moving
Average (ARMA) setting (for examples, see Wang et al., 2007; Hsu et al., 2008 and
Huang et al., 2008), the application of the partially penalized estimator as discussed
in Section 1.4.3 has not been discussed in this context.
Consider the following DGP
𝐿
∑︁
𝑦 𝑖 = x𝑖′ 𝛽 + ′
x𝑖− 𝑗 𝛼 𝑗 + 𝑢𝑖 , (1.33)
𝑗=1
where 𝐿 < 𝑁. Choosing the appropriate lag order for a particular variable is a
challenging task. Perhaps more importantly, as the number of observations increases,
the number of potential (lag) variables increases and this creates additional difficulties
in identifying a satisfactorily model. If one is only interested in statistical inference
on the estimates of 𝛽 , then the results from Section 1.4.3 may seem to be useful. In
this case, one can apply a Partially Penalized Estimator and obtain the parameter
Table 1.6: Percentage of Replications Excluding 𝑥22
30 0.0 3.32 3.28 3.3 0.0 2.08 7.72

50 0.0 2.08 2.06 2.06 0.0 1.18 5.16
0 100 0.0 1.16 1.16 1.16 0.0 0.68 3.58
500 0.0 0.64 0.62 0.62 0.0 0.18 1.88
1000 0.0 0.5 0.5 0.5 0.0 0.1 1.88
30 0.0 4.22 4.22 4.24 0.0 3.4 7.96

50 0.0 2.4 2.42 2.4 0.0 1.52 5.08
0.45 100 0.0 1.22 1.2 1.2 0.0 0.5 4.02
500 0.0 0.7 0.7 0.7 0.0 0.12 1.94
1000 0.0 0.34 0.34 0.34 0.0 0.1 1.28
30 0.0 14.28 14.3 14.32 0.0 13.6 19.36

50 0.0 9.6 9.52 9.5 0.0 8.46 14.32
0.9 100 0.0 5.62 5.6 5.62 0.0 5.3 9.52
500 0.0 2.54 2.54 2.58 0.0 2.0 5.24
1000 0.0 1.9 1.84 1.88 0.0 1.3 4.24
estimates as follows
2
𝑁
∑︁ ∑︁𝐿
′ ′
𝛽ˆ , 𝛼ˆ = arg min 𝑦 𝑖 − x𝑖 − x𝑖− 𝑗 𝛼 𝑗 ® + 𝜆𝑝 (𝛼
𝛼) , (1.34)
© ª
𝛽
𝛼
𝛽 ,𝛼 𝑖=1 « 𝑗=1 ¬
′
where 𝛼 = 𝛼 1′ , . . . ,𝛼𝛼 ′𝐿 and the 𝑝(𝛼𝛼 ) is a regularizer applied only to 𝛼 . Since 𝛽ˆ is not
part of the shrinkage, under the Bridge regularizer and the assumptions in Proposition
1.1, 𝛽ˆ has an asymptotically normal distribution which facilitates valid inferences on
𝛽 . Obviously, 𝛽 does not have to be the coefficients associated with the covariates
at the same time period as the response variable. The argument above should apply
to any coefficients of interests. The main idea is that if a researcher is interested in
conducting statistical inference on a particular set of coefficients with a potentially
high number of possible control variables, as long as the coefficients of interests are
not part of the shrinkage, valid inference on these coefficients may still be possible.
1.6.2 Panel Data Models
Following the idea above, another potentially useful application is the panel data
model with fixed effect. Consider
𝑦 𝑖𝑡 = x𝑖𝑡′ 𝛽 + 𝛼𝑖 + 𝑢 𝑖𝑡 , 𝑖 = 1, . . . , 𝑁, 𝑡 = 1, . . . ,𝑇 . (1.35)
The parameter vector 𝛽 is typically estimated by the fixed effect estimator

𝑁 ∑︁
∑︁ 𝑇
2
𝛽ˆ 𝐹𝐸 = arg min 𝑦¤ 𝑖𝑡 − x¤ 𝑖𝑡′ 𝛽 , (1.36)
𝛽 𝑖=1 𝑡=1
where 𝑦¤ 𝑖𝑡 = 𝑦 𝑖𝑡 − 𝑦¯ 𝑖 with 𝑦¯ 𝑖 = 𝑇 −1 𝑇𝑡=1 𝑦 𝑖𝑡 and 𝑥¤𝑖𝑡 = x𝑖𝑡 − x̄𝑖 with x̄𝑖 = 𝑇 −1 𝑇𝑡=1 x𝑖𝑡 .
Í Í
In practice, the estimator is typically computed using the dummy variable approach.
However, when 𝑁 is large, the number of 𝛼𝑖 is also large. Since it is common
for 𝑁 >> 𝑇 in panel data, the number of dummy variables required for the fixed
effect estimator can also be unacceptably large. Since the main focus is 𝛽 , under the
assumption that 𝛼𝑖 are constants for 𝑖 = 1, . . . , 𝑁, it seems also possible to apply the
methodology from the previous example and consider the following estimator
∑︁ 𝑇
𝑁 ∑︁
2
𝛽ˆ , 𝛼ˆ = arg min 𝑦 𝑖𝑡 − x𝑖𝑡′ 𝛽 − 𝛼𝑖 + 𝜆𝑝 (𝛼
𝛼) . (1.37)
𝛼
𝛽 ,𝛼 𝑖=1 𝑡=1
where 𝛼 = (𝛼1 , . . . , 𝛼 𝑁 ) ′ and 𝑝(𝛼

𝛼 ) denotes a regularizer applied only to the coefficients
of the fixed effect dummies. Proposition 1.1 should apply in this case without any
modification.
This can be extended to a higher dimensional panel with more than two indexes.
In such cases, the number of dummies required grows exponentially. While it is
possible to obtain a fixed effect estimators in a higher dimensional panel through
various transformations proposed by Balazsi, Mátyás and Wansbeek (2018), these
transformations are not always straightforward to derive and the dummy variable
approach could be more practically convenient. The dummy variable approach,
however, suffers from the curse of dimensionality and the proposed method here
seems to be a feasible way to resolve this issue. Another potential application is to
incorporate interacting fixed effects of the form 𝛼𝑖𝑡 into model (1.35). This is, of
course, not possible in a usual two-dimensional panel data setting, but feasible with
this approach.
Another possible application, which has been proposed in the literature, is to
incorporate a regularizer in Equation (1.36) and therefore define a shrinkage estimator
in a standard fixed effects framework. Specifically, fixed effects with shrinkage can be
defined as
𝑁 ∑︁
∑︁ 𝑇
2
𝛽ˆ 𝐹𝐸 = arg min 𝑦¤ 𝑖𝑡 − x¤ 𝑖𝑡′ 𝛽 + 𝜆𝑝(𝛽𝛽 ;𝛼
𝛼 ), (1.38)
𝛽 𝑖=1 𝑡=1
where 𝑝(𝛽𝛽 ) denotes the regularizer, which in principle, can be any regularizers such
as those introduced in Section 1.2. Given the similarity between Equations (1.38) and
all the other shrinkage estimators considered so far, it seems reasonable to assume
that the results in Knight and Fu (2000) and Proposition 1 would apply possibly
with only some minor modifications. This also means that fixed effect models with a
shrinkage estimator are not immune to the shortfall of shrinkage estimators in general.
The observations and issues highlighted in this chapter would apply equally in this
case.
1.6.3 Structural Breaks
Another econometric example where shrinkage type estimators could be helpful is

the testing for structural breaks with unknown breakpoints. Consider the following
DGP
𝑦 𝑖 = x𝑖′ 𝛽 + x𝑖′𝛿 0 𝐼 (𝑖 > 𝑡 1 ) + 𝑢 𝑖 , 𝑢 𝑖 ∼ 𝐷 (0, 𝜎𝑢2 ), (1.39)

where the break point, 𝑡 1 is unknown. Equation (1.39) implies that the parameter
vector when 𝑖 ≤ 𝑡 1 is 𝛽 0 and when 𝑖 > 𝑡 1 , it is 𝛽 0 + 𝛿 0 . In other words, a structural
break occurs at 𝑖 = 𝑡1 and 𝛿 denotes the shift in the parameter vector before and after
the break point.
Such models have a long history in econometrics, for example, see Andrews (1993)
and Andrews (2003) as well as the references within. However, the existing tests are
bounded by the 𝑝 < 𝑁 restriction. That is, the number of variables must be less than
the number of observations. Given that these tests are mostly residuals based tests,
this means that it is possible to obtain post-shrinkage (or post selection) residuals
and use these residuals in the existing tests. To illustrate the idea, consider the simple
case when 𝑡1 is known. In this case, a typical approach is to consider the following
𝐹-test statistics as proposed by Chow (1960)
𝑅𝑆𝑆 𝑅 − 𝑅𝑆𝑆𝑈𝑅1 − 𝑅𝑆𝑆𝑈𝑅2 𝑁 − 2𝑝
𝐹= , (1.40)
𝑅𝑆𝑆𝑈𝑅1 + 𝑅𝑆𝑆𝑈𝑅2 𝑝
where 𝑅𝑆𝑆 𝑅 denotes the residual sum-of-squares from the restricted model (𝛿𝛿 = 0),
while 𝑅𝑆𝑆𝑈𝑅1 and 𝑅𝑆𝑆𝑈𝑅2 denote the unrestricted sum-of-squares before and after
the break, respectively. Specifically, 𝑅𝑆𝑆𝑈𝑅1 denotes the residual sum-of-squares from
the residuals 𝑢ˆ 𝑡 = 𝑦 𝑡 − x𝑡′ 𝛽ˆ for 𝑡 ≤
𝑡1 and
𝑅𝑆𝑆𝑈𝑅2 denotes the residuals sum-of-squares
′ ˆ ˆ
from the residuals 𝑢ˆ 𝑖 = 𝑦 𝑡 − x𝑡 𝛽 + 𝛿 for 𝑡 > 𝑡 1 .
It is well known that under the null hypothesis 𝐻0 : 𝛿 = 0, the 𝐹-test statistics in
Equation (1.40) follows an 𝐹 distribution under the usual regularity conditions. When
𝑡1 is not known, Andrews (1993) derived the asymptotic distribution for
𝐹 = sup 𝐹 (𝑠), (1.41)

𝑠
where 𝐹 (𝑠) denotes the 𝐹-statistics as defined in Equation (1.40) assuming 𝑠 as the
breakpoint for 𝑠 = 𝑝 + 1, . . . , 𝑁 − 𝑝 − 1. The idea is to select a breakpoint 𝑠, such that
the test has the highest chance to reject the null of 𝐻0 : 𝛿 = 0. The distribution based
on this approach is non-standard as shown by Andrews (1993) and must therefore be
tabulated or simulated.
Note that the statistics in Equation (1.40) is based on the residuals rather than the
individual coefficient estimates, so it is possible to use the arguments by Belloni et al.
(2012) and construct the statistics as follows:
Step 1. Estimate the parameter vector 𝑦 𝑖 = x𝑖′ 𝛽 + 𝑢 𝑖 using a LASSO type estimator,
called it 𝛽ˆ 𝐿 𝐴𝑆𝑆𝑂 .
Step 2. Obtain a Post-Selection OLS. That is, estimate the linear regression model
using OLS with the covariates selected in the previous step.
Step 3. Construct the residuals using the estimates from the previous step 𝑢ˆ 𝑅,𝑖 = 𝑦 𝑖 − 𝑦ˆ 𝑖
where 𝑦ˆ 𝑖 = x𝑖′ 𝛽ˆ 𝑂𝐿𝑆 .
Í𝑁 2
Step 4. Compute 𝑅𝑆𝑆𝑈𝑅 = 𝑖=1 𝑢ˆ 𝑅,𝑖 .
Step 5. Estimate the following model using a LASSO-type estimator
𝑁
∑︁−1
𝑦 𝑖 = x𝑖′𝑖𝛽𝛽 + 𝛿 𝑗 x𝑖 𝛽 𝐼 (𝑖 ≤ 𝑗) + 𝑢 𝑖 (1.42)
𝑗=2
and denotes the estimates for 𝛽 , 𝛿 and 𝑗 as 𝛽ˆ 𝐿 𝐴𝑆𝑆𝑂−𝑈𝑅 , 𝛿ˆ𝐿 𝐴𝑆𝑆𝑂 and 𝑗ˆ,
respectively. Under the assumption there is only one break 𝛿 𝑗 = 0 for all 𝑗
except when 𝑗 = 𝑡 1 .
Step 6. Obtain the Post-Selection OLS for the pre-break unrestricted model,
𝛽ˆ 𝑈𝑅1−𝑂𝐿𝑆 . That is, estimate the linear regression model using OLS with the
covariates selected in Step 5 for 𝑖 ≤ 𝑗ˆ.
Step 7. Obtain the Post-Selection OLS for the post-break unrestricted model,
𝛽ˆ 𝑈𝑅2−𝑂𝐿𝑆 . That is, estimate the linear regression model using OLS with the
covariates selected in Step 5 for 𝑖 > 𝑗ˆ.
Step 8. Construct the pre-break residuals using 𝛽ˆ 𝑈𝑅1−𝑂𝐿𝑆 . That is, 𝑢ˆ𝑈𝑅1,𝑖 = 𝑦 𝑖 − 𝑦ˆ 𝑖 ,
where 𝑦ˆ 𝑖 = x𝑖′ 𝛽ˆ 𝑈𝑅1−𝑂𝐿𝑆 for 𝑖 ≤ 𝑗ˆ.
Step 9. Construct the post-break residuals using 𝛽ˆ 𝑈𝑅2−𝑂𝐿𝑆 . That is, 𝑢ˆ𝑈𝑅2,𝑖 = 𝑦 𝑖 − 𝑦ˆ 𝑖 ,
where 𝑦ˆ 𝑖 = x𝑖′ 𝛽ˆ 𝑈𝑅2−𝑂𝐿𝑆 for 𝑖 > 𝑗ˆ.
Í 𝑗ˆ 2
𝑢ˆ 2
Í𝑁
Step 10. Compute 𝑅𝑆𝑆𝑈𝑅1 = 𝑖=1 𝑢ˆ𝑈𝑅1,𝑖 and 𝑅𝑆𝑆𝑈𝑅2 = 𝑖= 𝑗ˆ+1 𝑈𝑅2,𝑖
.
Step 11. Compute the test statistics as defined in Equation (1.40).
Essentially, the proposal above uses LASSO as a variable selector as well as a
break point identifier. It then generates the residuals sum-of-squares using OLS based
on the selection given by LASSO. This approach can potentially be justified by the
results by Belloni et al. (2012) and Belloni and Chernozhukov (2013). Unlike the
conventional approach when the breakpoint is unknown, such as those studied by
Andrews (1993) whose test statistics have non-standard distributions, the test statistics
proposed here is likely to follow the 𝐹 distribution similar to the original test statistics
as proposed by Chow (1960) and can accommodate the case when 𝑝 > 𝑁. To the best
of the authors’ knowledge, this approach is novel with both the theoretical properties
and finite sample performance to be further evaluated. However, given the results of
Belloni et al. (2012), this seems like a plausible approach to tackle of the problem of
detecting structural breaks with unknown breakpoints.
1.7 Concluding Remarks
This chapter has provided a brief overview of the most popular shrinkage estimators
in the machine learning literature and discussed their potential applications in
econometrics. While valid statistical inference may be challenging to obtain directly
for shrinkage estimators, it seems possible, at least in the case of the Bridge estimator,
to conduct valid inference on the statistical significance of a subset of the parameter
vector, for the elements which are not part of the regularization. In the case of the
Bridge estimator, this chapter has provided such a result by modifying the arguments
of Knight and Fu (2000). Monte Carlo evidence suggested that similar results may
also be applicable to other shrinkage estimators, such as the adaptive LASSO and
SCAD. However, the results also highlighted that the finite sample performance of a
simple 𝑡-test for an unregularized parameter is no better than those obtained directly
from OLS. Thus, if it is possible to obtain OLS estimates, shrinkage estimators do
not seem to add value for inference purposes. However, when OLS is not possible,
as in the case when 𝑝 > 𝑁, shrinkage estimators provide a feasible way to conduct
statistical inference on the coefficients of interests, as long as they are not part of the
regularization.
Another interesting and useful result from the literature is that while the theoretical
properties of shrinkage estimators may not be useful in practice, shrinkage estimators
do lead to superior fitted values, especially in the case of post-shrinkage OLS. That
is, fitted values obtained by using OLS with the covariates selected by a shrinkage
estimator. The literature relies on this result to obtain optimal instruments in the
presence of many (possibly weak) instrumental variables. Using a similar idea, this
chapter has also proposed a new approach to test for a structural break when the break
point is unknown.7 Finally, the chapter has also highlighted the usefulness of these
methods in a panel data framework.
Table 1.7 contains a brief summary of the various shrinkage estimators introduced
in this chapter. Overall, machine learning methods in the framework of shrinkage
estimators seem to be quite useful in several cases when dealing with linear eco-
nometric models. However, users have to be careful, mostly with issues related to
estimation and variable selection consistency.
7 The theoretical properties and the finite sample performance of this test may be an interesting area
for future research.
Appendix
Proof of Proposition 1.1
For 𝛾 ≥ 1, using the same argument as Theorem 2 in Knight and Fu (2000), it is

straightforward to show that
√ 𝑑
𝑁 𝛽ˆ − 𝛽 0 → arg min𝑉 (𝜔
𝜔 ),
where
𝑝
∑︁
𝑉 (𝜔 𝜔 ′W +𝜔
𝜔 ) = −2𝜔 𝜔 ′C𝜔
𝜔 + 𝜆0 𝜔 𝑗 sgn 𝛽0 𝑗 |𝛽0 𝑗 | 𝛾−1 .
𝑗= 𝑝1 +1
Let 𝜔 ∗= arg min𝑉 (𝜔𝜔 ), which can be obtained by solving the First Order Necessary
Condition and this gives

∗ 𝜆0 𝛾−1
𝜔 𝑗 = c 𝑗 𝑤 𝑗 − sgn 𝛽0 𝑗 |𝛽0 𝑗 | 𝐼 ( 𝑗 > 𝑝 2 ) , 𝑗 = 1, . . . , 𝑝,
2
where c 𝑗 denotes the 𝑗 𝑡 ℎ row of C−1 and 𝑤 𝑗 denotes the 𝑗 𝑡 ℎ element in W. Note
that the last term in the expression above is 0 for 𝑗 ≤ 𝑝 1 . Thus, collect the first 𝑝 1
elements in 𝜔 ∗ gives the result. The argument for 𝛾 < 1 is essentially the same with
the definition of 𝑉 (𝜔
𝜔 ) being replaced by Theorem 3 in Knight and Fu (2000). This
completes the proof.
Table 1.7: Summary of Shrinkage Estimators
Estimator Advantages Disadvantages Software
LASSO
• Can assign 0 to coef- • Lack Oracle Properties in • Available in R, Python and
ficients general. Julia
• Computationally con- • Asymptotic distribution is
venient not practically useful
• Sensitive to the choice of
tuning parameter
• Estimator has no closed
form solution and must
rely on numerical meth-
ods.
Ridge
• Closed form solution • Lack Oracle Properties. • Available in almost all soft-
exists • Cannnot assign 0 to coeffi- ware packages.
• Lots of theoretical cients, • Also easy to implement
results • Sensitive to the choice of given the closed form solu-
tuning parameter tion
Elastic
Net • Aims to strike a • Lack Oracle properties. • Available in R, Python and
balance between • Need to choose two Julia.
LASSO and Ridge tuning parameters. The • Need to adjust the LARS
• Can assign 0 to coef- weight between LASSO algorithms
ficients and Ridge and the penalty
• Researchers can factor
adjust the balance • No closed form solution
between LASSO and and must rely on numerical
Ridge methods
Adaptive
LASSO • Possesses Oracle • Require initial estimates. • Not widely available in
Properties • Practical usefulness of Or- standard software pack-
• Can assign 0 to coef- acle Properties is limited ages
ficients • For linear model, it is
• Less biased than straightforward to com-
LASSO pute based on variable
• It is convenient to transformations
compute for linear
model
Table 1.7 Cont.:Summary of Shrinkage Estimators
Estimator Advantages Disadvantages Software
SCAD
• Possess Oracle prop- • Practical usefulness of Or- • Not widely available in
erties acle properties is limited standard software pack-
• Can assign 0 to coef- • No closed form solution ages.
ficients and must rely on numerical
• Estimates are gener- methods
ally less biased than • Two tuning parameters and
LASSO unclear how to determine
their values in practice
• Regularizer is complicated
and a function of the tun-
ing parameter
Group
LASSO • Useful when covari- • Practical usefulness of Or- • Not widely available in
ates involve categor- acle properties is limited standard software pack-
ical variables • No closed form solution ages
• Allow coefficients of and must rely on numerical • Computationally intensive
a group of variables methods
to be either all zeros • An additional kernel mat-
or all non-zeros rix is required for each
• Possess Oracle prop- group
erties
References 37
References
Andrews, D. W. K. (1993). Tests for Parameter Instability and Structural Change with
Unknown Change Point. Econometrica: Journal of the Econometric Society,
821–856.
Andrews, D. W. K. (2003). End-of-Sample Instability Tests. Econometrica, 71(6),
1661–1694.
Balazsi, L., Mátyás, L. & Wansbeek, T. (2018). The Estimation of Multidimensional
Fixed Effects Panel Data Models. Econometric Reviews, 37, 212-227.
Belloni, A., Chen, D., Chernozhukov, V. & Hansen. (2012). Sparse Models and
Methods for Optimal Instruments With an Application to Eminent Domain.
Econometrica, 80(6), 2369–2429. doi: 10.3982/ECTA9626
Belloni, A. & Chernozhukov, V. (2013, May). Least Squares after Model Selection
in High-dimensional Sparse Models. Bernoulli, 19(2), 521–547. doi: 10.3150/
11-BEJ410
Chernozhukov, V., Hansen, C. & Spindler, M. (2015). Valid Post-Selection and
Post-Regularization Inference: An Elementary, General Approach. Annual
Review of Economics, 7, 649–688.
Chow, G. C. (1960). Tests of Equality between Sets of Coefficients in Two Linear
Regressions. Econometrica: Journal of the Econometric Society, 28(3), 591–
605.
Donoho, D. L. & Johnstone, I. M. (1994). Ideal Spatial Adaptation by Wavelet
Shrinkage. Biometrika, 81, 425-455.
Efron, B., Hastie, T., Johnstone, I. & Tibshirani, R. (2004). Least Angle Regression.
Annals of Statistics, 32(2), 407–451.
Fan, J. & Li, R. (2001). Variable Selection via Nonconcave Penalized Likelihood and
its Oracle Properties. Journal of the American Statistical Association, 96(456),
1348–1360. doi: 10.1198/016214501753382273
Fan, J., Li, R., Zhang, C.-H. & Zou, H. (2020). Statistical Foundations of Data
Science. CRC Press, Chapman and Hall.
Fan, J., Xue, L. & Zou, H. (2014). Strong Oracle Optimality of Folded Concave
Penalized Estimation. Annals of Statistics, 42(3), 819–849.
Frank, I. & Friedman, J. (1993). A Statistical View of Some Chemometrics Regression
Tools. Technometrics, 35, 109-148.
Hastie, T., Tibshirani, R. & Friedman, J. (2009). The elements of statistical learning:
Data mining, inference and prediction. Springer.
Hoerl, A. & Kennard, R. (1970a). Ridge Regression: Applications to Nonorthogonal
Problems. Technometrics, 12, 69-82.
Hoerl, A. & Kennard, R. (1970b). Ridge Regression: Biased Estimation for
Nonorthogonal Problems. Technometrics, 12, 55-67.
Hsu, N.-J., Hung, H.-L. & Chang, Y.-M. (2008, March). Subset Selection for Vector
Autoregressive Processes using LASSO. Computational Statistics & Data
Analysis, 52(7), 3645–3657. doi: 10.1016/j.csda.2007.12.004
Huang, J., Ma, S. & Zhang, C.-H. (2008). Adaptive LASSO for Sparse High-
dimensional Regression Models. Statistica Sinica, 18, 1603-1618.
Knight, K. & Fu, W. (2000). Asymptotics for LASSO-Type Estimators. The Annals
of Statistics, 28(5), 1356–1378.
Kock, A. B. (2016, February). Consistent and Conservative Model Selection
with the Adaptive LASSO in Stationary and Nonstationary Autoregressions.
Econometric Theory, 32(1), 243–259. doi: 10.1017/S0266466615000304
Lee, J. D., Sun, D. L., Sun, Y. & Taylor, J. E. (2016, June). Exact Post-selection
Inference, with Application to the LASSO. The Annals of Statistics, 44(3). doi:
10.1214/15-AOS1371
Leeb, H. & Pötscher, B. M. (2005). Model Selection and Inference: Facts and Fiction.
Econometric Theory, 21(1), 21–59. doi: 10.1017/S0266466605050036
Leeb, H. & Pötscher, B. M. (2008). Sparse Estimators and the Oracle Property, or
the Return of Hodges’ Estimator. Journal of Econometrics, 142(1), 201–211.
doi: 10.1016/j.jeconom.2007.05.017
Lockhart, R., Taylor, J., Tibshirani, R. J. & Tibshirani, R. (2014). A Significance
Test for the LASSO. The Annals of Statistics, 42(2), 413–468.
Medeiros, M. C. & Mendes, E. F. (2016). 𝐿 1 -Regularization of High-dimensional
Time-Series Models with non-Gaussian and Heteroskedastic Errors. Journal
of Econometrics, 191(1), 255–271. doi: 10.1016/j.jeconom.2015.10.011
Tibshirani, R. (1996, January). Regression Shrinkage and Selection Via the Lasso.
Journal of the Royal Statistical Society: Series B (Methodological), 58(1),
267–288. doi: 10.1111/j.2517-6161.1996.tb02080.x
Tinbergen, J. (1939). Statistical Testing of Business-Cycle Theories. League of
Nations, Economic Intelligence Service.
Wang, H., Li, G. & Tsai, C.-L. (2007, February). Regression Coefficient and
Autoregressive Order Shrinkage and Selection via the LASSO. Journal of
the Royal Statistical Society: Series B (Statistical Methodology), 69(1). doi:
10.1111/j.1467-9868.2007.00577.x
Yuan, M. & Lin, Y. (2006, February). Model Selection and Estimation in Regression
with Grouped Variables. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 68(1), 49–67. doi: 10.1111/j.1467-9868.2005.00532
.x
Zhang, C.-H. & Zhang, S. S. (2014, January). Confidence Intervals for Low
Dimensional Parameters in High Dimensional Linear Models. Journal of the
Royal Statistical Society: Series B (Statistical Methodology), 76(1), 217–242.
doi: 10.1111/rssb.12026
Zhang, Y., Li, R. & Tsai, C. (2010). Regularization Parameter Selections via Gener-
alized Information Criterion. Journal of the American Statistical Association,
105, 312-323.
Zou, H. (2006, December). The Adaptive Lasso and Its Oracle Properties. Journal
of the American Statistical Association, 101(476), 1418–1429. doi: 10.1198/
016214506000000735
Zou, H. & Hastie, T. (2005, April). Regularization and Variable Selection via
the Elastic Net. Journal of the Royal Statistical Society: Series B (Statistical
Methodology), 67(2), 301–320. doi: 10.1111/j.1467-9868.2005.00503.x
References 39
Zou, H., Hastie, T. & Tibshirani, R. (2007). On the Degrees of Freedom of the
LASSO. Annals of Statistics, 35, 2173-2192.
Zou, H. & Li, R. (2008, August). One-step Sparse Estimates in Nonconcave
Penalized Likelihood Models. The Annals of Statistics, 36(4). doi: 10.1214/
009053607000000802
Chapter 2
Nonlinear Econometric Models with Machine
Learning
Felix Chan, Mark N. Harris, Ranjodh B. Singh and Wei (Ben) Ern Yeo
Abstract This chapter introduces machine learning (ML) approaches to estimate

nonlinear econometric models, such as discrete choice models, typically estimated by
maximum likelihood techniques. Two families of ML methods are considered in this
chapter. The first, shrinkage estimators and related derivatives, such as the Partially
Penalised Estimator, introduced in Chapter 1. A formal framework of these concepts
is presented as well as a brief literature review. Additionally, some Monte Carlo
results are provided to examine the finite sample properties of selected shrinkage
estimators for nonlinear models. While shrinkage estimators are typically associated
with parametric models, tree based methods can be viewed as their non-parametric
counterparts. Thus, the second ML approach considered here is the application of tree-
based methods in model estimation with a focus on solving classification, or discrete
outcome, problems. Overall, the chapter attempts to identify the nexus between these
ML methods and conventional techniques ubiquitously used in applied econometrics.
This includes a discussion of the advantages and disadvantages of each approach.
Several benefits, as well as strong connections to mainstream econometric methods
are uncovered, which may help in the adoption of ML techniques by mainstream
econometrics in the discrete and limited dependent variable spheres.
Felix Chan B
Mark N. Harris
Curtin University, Perth, Australia, e-mail: mark.harris@curtin.edu.au
Ranjodh B. Singh
Curtin University, Perth, Australia, e-mail: ranjodh.singh@curtin.edu.au
Wei (Ben) Ern Yeo
Curtin University, Perth, Australia, e-mail: weiern.yeo@postgrad.curtin.edu.au

42 Chan at al.
2.1 Introduction
This chapter aims to examine the potential applications of machine learning techniques
in specifying and estimating nonlinear econometric models. The two main objectives
of the chapter are to:
1. provide an overview of a suite of machine learning (ML) techniques that are
relevant to the specification, estimation and testing of nonlinear econometric
models; and
2. identify existing gaps that prevent these techniques from being readily applicable
to econometric analyses.
Here we take ‘nonlinear econometric models’ to mean any of a wide range
of different models used in various areas of applied economics and econometric
analyses. They can generally be classified into the following categories based on the
characteristics of the dependent (response) variable.
1. The dependent (response) variable is discrete.
2. The dependent variable is partly continuous, but is limited in some way; such as
it can only be positive.
3. The response variable is continuous but has a nonlinear relationship, such as
piece-wise step function, with one or more observed covariates.
The first case can be further divided into three different subcases namely, cardinal,
nominal and ordinal responses. Cardinal responses typically manifest as Count Data
in econometrics. In this case, the set of all possible outcomes is often infinite, with
the Poisson model being the fundamental building block of the analysis.
Nominal and ordinal responses, however, often appear in different context in
econometrics to cardinal responses, as they are typically a set of finite discrete
choices, which are usually modelled by a suite of well-known discrete choice models,
such as Logit, Probit and Multinomial models, to name just a few. In the machine
learning (ML) literature, the modelling of nominal and ordinal responses is often
called a classification problem. The aim is to model a discrete choice outcome (among
a set of alternatives) of an individual/entity. Popular examples of discrete choice
analyses include modelling labour force participation (Yes/No), Political beliefs
(Strongly Agree, Agree, Neutral, Disagree and Strongly Disagree) or the choice
of among different products or occupations. The objective of these analyses is to
understand the relation between the covariates (also known as predictors, features
and confounding variables) and the stated/revealed choice.
The second type of nonlinear model typically appears where a continuous variable
has been truncated, or censored, in some manner. Examples abound in modelling
labour supply, charitable donations and financial ratios, among others, where the
response variable is always ≥ 0. This chapter is less focused on this particular type of
model, as it is less common in the ML literature.
The third type of nonlinear models is more popular in the time series econometrics
literature, examples include threshold regression and the threshold autoregressive
(TAR) model; see Tong (2003) for a comprehensive review of threshold models in
2 Nonlinear Econometric Models with Machine Learning 43
time series analysis. These models have been used extensively in modelling regime
switching behaviours. Econometric applications of these models include, for example,
threshold effects between inflation and growth during different phases of a business
cycle (see, e.g., Teräsvirta & Anderson, 1992 and Jansen & Oh, 1999).
Given the extensive nature of nonlinear models, it is beyond the scope of this
chapter to cover the details of ML techniques applicable to all such models. Instead,
the chapter focuses on two ML methods which are applicable to a selected set of
models popular among applied economists and econometricians. These are shrinkage
estimators and tree based techniques in the form of Classification and Regression
Tree (CART). The intention is to provide a flavour on the potential linkage between
ML procedures and econometric models.
The choice of the two techniques is not arbitrary. Shrinkage estimators can be seen
as an extension to the estimators used for parametric models, whereas CART can be
viewed as the ML analogue to nonparametric estimators often used in econometrics.
The connection between CART and nonparametric estimation is also discussed in
this chapter.
The chapter is organised as follows. In Section 2.2 shrinkage estimators and
Partially Penalised Estimators for nonlinear models are introduced. The primary
focus is on variable selection and/or model specification. A brief literature review is
presented in order to outline the applications of this method to a variety of disciplines.
This includes recent developments on valid statistical inferences with shrinkage and
Partially Penalised Estimators for nonlinear models. Asymptotic distributions of the
Partial Penalised Estimators for Logit, Probit and Poisson models are derived. The
results should facilitate valid inferences. Some Monte Carlo results are also provided
to assess the finite sample performance of shrinkage estimators when applied to a
simple nonlinear model.
Section 2.3 provides an overview of ML tree-based methods, including an outline
of how trees are constructed as well as additional methods to improve their predictive
performance. This section also includes an example for demonstration purposes.
Section 2.3.3 highlights the potential connections between tree-based and mainstream
econometric methods. This may help an applied econometrician to make an informed
decision about which method/s are more appropriate for any given specific problem.
Finally, Section 2.4 summarises the main lessons of the chapter.
2.2 Regularization for Nonlinear Econometric Models
This section explores shrinkage estimators for nonlinear econometric models, includ-
ing the case when the response variable is discrete. In the model building phase, one
challenge is the selection of explanatory variables. Especially in the era of big data,
potentially the number of variables can exceed the number of observations. Thus, a
natural question is how can one systematically select relevant variables? As already
seen in Chapter 1, obtaining model estimates using traditional methods is not feasible
when the number of covariates exceeds the number of observations.
44 Chan at al.
As in the case of linear models, one possible solution is to apply shrinkage

estimators with regularization. This method attempts to select relevant variables from
the entire set of variables, where the number of such variables may be greater than
the number of observations. In addition to variable selection, regularization may also
be used to address model specification. Specifically, here we are referring to different
transformations of variables entering the model: squaring and/or taking logarithms,
for example.
While Chapter 1 explores shrinkage estimators for linear models, their application
to nonlinear ones requires suitable modifications on the objective function. Recall, a
shrinkage estimator can be expressed as solution to an optimisation as presented in
Equations (1.5) and (1.6). namely,
𝛽ˆ = arg min 𝑔 (𝛽𝛽 ; y, X)

𝛽
𝛼 ) ≤ 𝑐,
s.t. 𝑝 (𝛽𝛽 ;𝛼
where 𝛽 denotes the parameter vector, y and X are the vectors containing the
observations for the response variable and the covariates, respectively, 𝑝(𝛽𝛽 ;𝛼 𝛼 ) is
the regularizer where 𝛼 denotes additional tuning parameters, and 𝑐 is a positive
constant.
In the case of linear models, the least squares is often used as the objective function,
that is, 𝑔(𝛽𝛽 ; y, X) = (y − X𝛽𝛽 ) ′ (y − X𝛽𝛽 ) coupled with different regularizers, 𝑝(𝛽𝛽 ;𝛼
𝛼 ).
Popular choices include LASSO, Ridge, Adaptive LASSO and SCAD. See Chapter 1
for more information on the different regularizers under the least square objective. In
the econometric literature, nonlinear least squares and maximum likelihood are the
two most popular estimators for nonlinear models. A natural extension of shrinkage
estimators then is to define the objective function based on nonlinear least squares or
the likelihood function.
2.2.1 Regularization with Nonlinear Least Squares
Consider the following model
𝑦 𝑖 = ℎ(x𝑖 ; 𝛽 0 ) + 𝑢 𝑖 , 𝑢 𝑖 ∼ 𝐷 (0, 𝜎𝑢2 ), (2.1)

where 𝑢 𝑖 is a continuous random variable, which implies that the response variable,
𝑦 𝑖 , is also a continuous random variable. The expectation of 𝑦 𝑖 conditional on the
covariates, x𝑖 , is a twice differentiable function ℎ : R 𝑝 → R, which depends on the
covariates x𝑖 as well as the parameter vector 𝛽 0 . Similarly to the linear case in Chapter
1, while the true parameter vector 𝛽 0 is a 𝑝 × 1 vector, many of the elements in 𝛽
can be zeros. Since 𝛽 0 is not known in practice, the objective is to estimate 𝛽 0 using
shrinkage estimators, or at the very least, using shrinkage estimator to identify which
elements in 𝛽 0 are zeros.
Under the assumption that the functional form of ℎ is known, shrinkage estimators
in this case can be defined as the solution to the following optimisation problem
𝑛
∑︁
𝛽ˆ = arg min [𝑦 𝑖 − ℎ (x𝑖 ; 𝛽 )] 2
𝛽 𝑖=1
𝛼 ) ≤ 𝑐,
and similarly to the linear case, the above can be expressed in its Lagrangian form
𝑛
∑︁
𝛽ˆ = arg min [𝑦 𝑖 − ℎ (x𝑖 ; 𝛽 )] 2 + 𝜆𝑝 (𝛽𝛽 ;𝛼
𝛼) , (2.2)
𝛽 𝑖=1
where 𝜆 denotes the tuning parameter. As mentioned in Chapter 1, there are many
choices for regularizers, 𝑝. Popular choices include the Bridge, which has Ridge and
LASSO as special cases, Elastic Net, Adaptive LASSO and SCAD. Each of these
choices applies a different penalty to the elements of 𝛽 (please see Chapter 1 for
further discussions on the properties of these regularizers).
While shrinkage estimators, such as LASSO, are often used as a variable selector
for linear models by identifying which elements in 𝛽 are zeros, such interpretation
does not carry through to nonlinear models in general. This is because the specification
of the conditional mean, ℎ(x𝑖 ; 𝛽 ), is too general to ensure that each element in 𝛽 is
associated with a particular covariate. That is, 𝛽 𝑗 = 0, does not necessary mean that a
covariate 𝑥 𝑗 is to be excluded. Therefore, shrinkage estimators in this case are not
necessarily a selector. At the very least, not a variable selector. One example is
𝑦 𝑖 = 𝛽1 cos(𝛽2 𝑥1𝑖 + 𝛽3 𝑥2𝑖 ) + 𝛽4 𝑥 2𝑖 + 𝑢 𝑖
and clearly 𝛽3 = 0 does not mean 𝑥2𝑖 has not been ‘selected’. One exception is when the
conditional mean can be expressed as a single index function, that is, ℎ(x𝑖 ; 𝛽 ) = ℎ(x𝑖′ 𝛽 ).
In this case, the conditional mean is a function of the linear combination of the
covariates. For this reason, the interpretation of selection consistency and Oracle
properties as introduced in Chapter 1 require extra care. For a nonlinear model,
selection consistency should be interpreted as the ability of a shrinkage estimator to
identify elements in 𝛽 with 0 value. The ability of a shrinkage estimator as a variable
selector is only applicable when the conditional mean is a single index function.
When the functional form of ℎ is unknown, shrinkage estimators are no longer
feasible. In econometrics, nonparametric estimators are typically used in this case. In
the Machine Learning literature, this problem is often solved by techniques such as
CART, to be discussed in Section 2.3.
46 Chan at al.
2.2.2 Regularization with Likelihood Function
The maximum likelihood estimator is arguably one of the most popular estimators
in econometrics, especially for nonlinear models. Regardless of the nature of the
response variables, continuous or discrete, as long as the distribution of the response
variable is known, or can be approximated reasonably well, then it is possible to
define the corresponding shrinkage estimators by using the log-likelihood as the
objective function.
However, there are two additional complications that require some minor ad-
justments. First, the nature of the optimisation is now a maximisation problem i.e.,
maximising the (log-) likelihood, rather than a minimisation problem, i.e., the least
squares. Second, the distribution of the response variable often involves additional
parameters. For example, under the assumption of normality, the least squares estim-
ator generally does not estimate the variance of the conditional distribution jointly
with the parameter vector. In the maximum likelihood case, the variance parameter is
jointly estimated with the parameter vector. This needs to be incorporated into the
formulation of shrinkage estimators. Therefore, a shrinkage estimator can be defined
as the solution to the following optimisation problem

𝛽ˆ , 𝛾ˆ =arg max log 𝐿(𝛽𝛽 ,𝛾𝛾 ; y, X) (2.3)
𝛾
𝛽 ,𝛾
𝛼 ) ≤ 𝑐, (2.4)
where 𝛾 denotes additional parameter vector required for the distribution, 𝐿(𝛽𝛽 ,𝛾𝛾 ; y, x)
denotes the likelihood function and, as usual, 𝑝(𝛽𝛽 ;𝛼𝛼 ) denotes the regularizer (penalty)
function. The optimisation problem above is often presented in its Lagrangian form
for a given 𝑐. That is,

ˆ 𝛾ˆ = arg max [log 𝐿 (𝛽𝛽 ,𝛾𝛾 ; y, X) − 𝜆𝑝( 𝜷; 𝜶)] ,
𝜷, (2.5)
𝜷,𝛾
𝛾
where 𝜆 ≥ 0. Note that the optimisation problem above is a maximisation problem

rather than a minimisation problem, as presented in Equations (1.5) and (1.6).
However, since maximising a function 𝑓 (𝑥) is the same as minimising 𝑔(𝑥) = − 𝑓 (𝑥),
the formulation above can be adjusted, so that it is consistent with Equations (1.5)
and (1.6). namely,

𝛽ˆ , 𝛾ˆ =arg min − log 𝐿 (𝛽𝛽 ,𝛾𝛾 ; y, X) (2.6)
𝛾
𝛽 ,𝛾
𝛼 ) ≤ 𝑐, (2.7)
with the corresponding Lagrangian being

ˆ 𝛾 = arg min [− log 𝐿 (𝛽𝛽 ,𝛾𝛾 ; y, X) + 𝜆𝑝( 𝜷; 𝜶)] .
𝜷,𝛾 (2.8)
𝜷,𝛾
𝛾
Continuous Response Variable
Consider the model as defined in Equation (2.1), under the assumption that 𝑢 𝑖 ∼
𝑁 𝐼 𝐷 (0, 𝜎𝑢2 ), then a shrinkage estimator under the likelihood objective is defined as
[𝑦 𝑖 − ℎ(x𝑖 ; 𝛽 )] 2
𝑛
𝑛 ∑︁
𝛽ˆ , 𝜎
ˆ 𝑢2 = arg max − log 𝜎𝑢2 − 𝛼 ).
− 𝜆𝑝(𝛽𝛽 ;𝛼 (2.9)
𝛽 , 𝜎𝑢2 2 𝑖=1
2𝜎𝑢2
It is well known that the least squares estimator is algebraically equivalent to the
maximum likelihood estimator for 𝛽 under normality. This relation is slightly more
complicated in the context of shrinkage estimators, and it concerns mainly with the
tuning parameter 𝜆. To see this, differentiate Equation (2.2) to obtain the first order
condition for the nonlinear least squares shrinkage, this gives
𝑛
∑︁ 𝜕ℎ 𝜆 𝐿𝑆 𝜕 𝑝
ˆ

𝑦 𝑖 − ℎ(x𝑖 ; 𝛽 ) = (2.10)
𝑖=1
𝜕𝛽𝛽 𝛽 =𝛽ˆ 2 𝜕𝛽𝛽 𝛽 =𝛽ˆ
where 𝜆 𝐿𝑆 denotes the tuning parameter 𝜆 associated with the least squares objective.
Repeat the process above for Equation (2.9) to obtain the first order condition for the
shrinkage estimators with likelihood objective, and this gives
𝑦 𝑖 − ℎ(x𝑖 ; 𝛽ˆ ) 𝜕ℎ
𝑛

∑︁ 𝜕 𝑝
=𝜆 𝑀 𝐿 (2.11)
𝑖=1
ˆ 𝑢2
𝜎 𝜕𝛽𝛽 𝛽 =𝛽ˆ 𝜕𝛽𝛽 𝛽 =𝛽ˆ
2
𝑦 𝑖 − ℎ(x𝑖 ; 𝛽ˆ )
𝑛

∑︁
2
ˆ𝑢 =
𝜎 , (2.12)
𝑖=1
𝑛
where 𝜆 𝑀 𝐿 denotes the tuning parameter 𝜆 for the shrinkage estimator with likelihood
objective. Compare Equation (2.10) and Equation (2.12), it is straightforward to see
that the two estimators will only be algebraically equivalent if their tuning parameters
satisfy
ˆ 𝑢2 𝜆 𝑀 𝐿 .
𝜆 𝐿𝑆 = 2𝜎 (2.13)
This relation provides a link between shrinkage estimators with the least squares
and likelihood objective under normality. Note that this relation holds for both linear
and nonlinear least squares, since ℎ(𝜉𝜉 ; 𝛽 ) can theoretically be a linear function in x𝑖 .
Given the two estimators are algebraically equivalent under the appropriate choice
of tuning parameters, they are likely to share the same properties conditional on the
validity of Equation (2.13). As such, this chapter focuses on shrinkage estimators
with a likelihood objective.
While theoretically they may share the same properties under Equation (2.13), the
choice of the tuning parameter is often based on data-driven techniques in practice,
such as cross validation, see 2.2.3 below. Therefore, it is unclear if Equation (2.13)
necessarily holds in practice. More importantly, since ℎ(x𝑖 ; 𝛽 ) can be a linear function,
this means the shrinkage estimators based on least squares may very well be different
48 Chan at al.
from shrinkage estimators based on maximum likelihood in practice, even under the
same regularizers.
Discrete Response Variables
Another application of shrinkage estimators with a likelihood objective is the

modelling of discrete random variables. There are two common cases in econometrics.
The first concerns with discrete random variables with finite number of outcomes.
More specifically, consider a random variable, 𝑦 𝑖 , that takes on values from a finite,
countable set D = {𝑑1 , . . . , 𝑑 𝑘 } and the probability of 𝑦 𝑖 = 𝑑 𝑗 conditional on a set of
covariates, x𝑖 , can be written as
Pr 𝑦 𝑖 = 𝑑 𝑗 = 𝑔 𝑗 x𝑖′ 𝛽 0 ,

𝑗 = 1, . . . , 𝑘, (2.14)
′ ′
where x𝑖 = 𝑥1𝑖 , . . . , 𝑥 𝑝𝑖 is a 𝑝 × 1 vector of covariates, 𝛽 0 = 𝛽01 , . . . , 𝛽0 𝑝 is the
𝑝 × 1 parameter vector and 𝑔 𝑗 (x) denotes a twice differentiable function in x.
Two simple examples of the above are the binary choice Logit and Probit models
where D = {0, 1}. In the case of Logit model,
exp(x𝑖′ 𝛽 )
Pr(𝑦 𝑖 = 1|x𝑖 ) =
1 + exp(x𝑖′ 𝛽 )
1
Pr(𝑦 𝑖 = 0|x𝑖 ) = ,
1 + exp(x𝑖′ 𝛽 )
and in the case of Probit model,
Pr(𝑦 𝑖 = 1|x𝑖 ) =Φ(x𝑖′ 𝛽 )

Pr(𝑦 𝑖 = 0|x𝑖 ) =1 − Φ x𝑖′ 𝛽 ,

where Φ(𝑥) denotes the standard normal cumulative distribution function (CDF).
Like in the case of linear models as discussed in Chapter 1, the coefficient vector
𝛽 can contain many zeros. In this case, the associated covariates do not affect the
probability of 𝑦 𝑖 . An objective of a shrinkage estimator is to identify which elements
in 𝛽 are zeros.
While there exists variable selection procedures, such as the algorithm proposed
in Benjamini and Hochberg (1995), the resulting forward and backward stepwise
procedures are known to have several drawbacks. In the case of forward stepwise,
it is sensitive to the initial choice of variable, while backward stepwise is often
not possible if the number of variable is greater than the number of observations.
Shrinkage estimators can alleviate some of these drawbacks. Readers are referred
to Hastie, Tibshirani and Friedman (2009) for more detailed discussion on stepwise
procedures.
Similarly to the nonlinear least squares case, shrinkage estimator is a variable
selector only when the log-likelihood function can be expressed as a function of a
linear combination of the covariates. That is 𝐿(𝛽𝛽 ,𝛾𝛾 ; y, X) = 𝐿 (X′ 𝛽 ,𝛾𝛾 ; y). Popular
econometric models that fall within this category include the Logit and Probit models.
For Logit model, or logistic regression, let

𝜋(x)
𝑓 (x; 𝜷) = log = x′ 𝛽 , (2.15)
1 − 𝜋(x)
where 𝜋(x) = Pr(𝑦 = 1|x) denotes the probability of a binary outcome conditional on
x. The log-likelihood for the binary model is well-known and given this, the shrinkage
estimator for the Logit model can be written as
" 𝑛
#
∑︁
𝜷ˆ = arg max {𝑦 𝑖 𝑓 (x𝑖 ; 𝜷) − log(1 + exp 𝑓 (x𝑖 ; 𝜷))} − 𝜆𝑝(𝛽𝛽 ;𝛼
𝛼) . (2.16)
𝜷 𝑖=1
The shrinkage estimator for the Probit model follows a similar formulation. The
log-likelihood function for binary Probit model is
𝑛
∑︁
𝑦 𝑖 log Φ x𝑖′ 𝛽 + (1 − 𝑦 𝑖 ) log 1 − Φ(x𝑖′ 𝛽 )

𝑖=1
and therefore shrinkage estimator for the binary Probit model can be written as
𝑛
∑︁
𝛽ˆ = arg max 𝑦 𝑖 log Φ x𝑖′ 𝛽 + (1 − 𝑦 𝑖 ) log 1 − Φ(x𝑖′ 𝛽 ) − 𝜆𝑝(𝛽𝛽 ;𝛼

𝛼 ). (2.17)
𝛽 𝑖=1
Again, the type of shrinkage estimator depends on the regularizer. In principle, all
regularizers introduced in Chapter 1 can also be used in this case, but their theoretical
properties and finite sample performance are often unknown. This is discussed further
in Section 2.2.3.
The binary choice framework can be extended to incorporate multiple choices.
Based on the above definitions, once the log-likelihood function for a model is known,
it can be substituted into Equation (2.5) with a specific regularizer to obtain the
desired shrinkage estimator.
The second case for a discrete random response variable is when D is an infinite
and countable set, e.g., D = Z+ = {0, 1, 2, . . .}. This type of response variables are
often appeared as count data in econometrics. One popular model for count data is
the Poisson model. This model can be written as
𝑦
exp(−𝜇𝑖 )𝜇𝑖
Pr(𝑦|x𝑖 ) =
𝑦!
where 𝑦! denotes the factorial of 𝑦 and 𝜇𝑖 = exp x𝑖′ 𝛽 . The log-likelihood function

now is
𝑛
∑︁
− exp(x𝑖′ 𝛽 ) + 𝑦 𝑖 (x𝑖′ 𝛽 ) + 𝑦 𝑖 !,
𝑖=1
50 Chan at al.
where the last term 𝑦 𝑖 ! is a constant and is often omitted for purposes of estimation
as it does not affect the computation of the estimator. The shrinkage estimator with a
particular regularizer for Poisson model is defined as
𝑛
∑︁
𝛽ˆ = arg max − exp(x𝑖′ 𝛽 ) + 𝑦 𝑖 (x𝑖′ 𝛽 ) − 𝜆𝑝(𝛽𝛽 ;𝛼
𝛼 ). (2.18)
𝛽 𝑖=1
2.2.3 Estimation, Tuning Parameter and Asymptotic Properties
This section discusses the estimation, determination of the tuning parameter, 𝜆

and the asymptotic properties of shrinkage estimators with the least squares and
maximum likelihood objectives. The overall observation is that theoretical properties
of shrinkage estimators with the least squares and likelihood objectives are sparse. In
fact, the computation of the shrinkage estimators can itself be a challenging problem
given the present knowledge in solving constrained optimisation. Asymptotic results
are also rare and concepts, such as selection consistency, often require additional care
in their interpretations for nonlinear models. This section covers some of these issues
and identify the existing gaps for future research.
Estimation
To the best of our knowledge, the computation of shrinkage estimators with the least
squares or likelihood objectives is still a challenging problem in general. This is due
to the fact that constrained optimisation for nonlinear function is a difficult numerical
problem.
There are specific cases, such as when the log-likelihood function is concave or
when the nonlinear least squares function is convex, efficient algorithms do exist to
solve the associated optimisation problems for certain regularizers, see for example
Kwangmoo, Seung-Jean and Boyd (2007). Another exception is the family of models
that falls under the Generalised Linear Model (GLM) framework. This includes Logit,
Probit and Poisson models as special cases. For these models, efficient algorithms do
𝛾
exist for regularizers that are convex, such as the Bridge regularizers 𝑝(𝛽𝛽 ) = ||𝛽𝛽 || 𝛾
for 𝛾 ≥ 1 as well as Elastic Net and SCAD. These estimators are readily available in
open source languages such as R1, Python2 and Julia3.
1 https://cran.r-project.org/web/packages/glmnet/index.html
2 https://scikit-learn.org/stable/index.html
3 https://www.juliapackages.com/p/glmnet
Tuning Parameter and Cross-Validation
Like in the linear case, the tuning parameter, 𝜆, plays an important role in terms of
the performance of the shrinkage estimators for nonlinear model in practice. In the
case when the response variable is discrete, the cross validation as introduced in
Chapter 1 needs to be modified for shrinkage estimator with likelihood objective. The
main reason for this modification is that the calculation of ‘residuals’, which the least
squares seek to minimise, is not obvious in the context of nonlinear models. This is
particularly true when the response variable is discrete. A different measurement of
‘errors’ in the form of deviance is required for purposes of identifying the optimal
tuning parameter. One such modification can be found below:
1. For each 𝜆 value in a sequence of values (𝜆1 > 𝜆 2 > · · · > 𝜆𝑇 ), estimate the model
for each fold, leaving one fold out at a time. This produces a vector of estimates
for each 𝜆 value: 𝜷ˆ1 . . . 𝜷ˆ𝑻 .
2. For each set of estimates, calculate the deviance based on the left out fold/testing
dataset. The deviance is defined as
2 ∑︁ 𝑘
𝑒 𝑡𝑘 = − log𝑝(𝑦 𝑖 |𝒙 𝒊′ 𝜷ˆ𝑇 ), (2.19)
𝑛𝑘 𝑖
where 𝑛 𝑘 is the number of observations in the fold 𝑘. The quantity computed for
each fold and across all 𝜆 values.
3. Compute the average and standard deviation of the error/deviance resulting from
𝐾 folds.
1 ∑︁ 𝑘
𝑒¯𝑡 = 𝑒 . (2.20)
𝐾 𝑘 𝑡
The above average represents the average error/deviance for each 𝜆.
√︄
1 ∑︁ 𝑘
𝑠𝑑 ( 𝑒¯𝑡 ) = (𝑒 − 𝑒¯𝑡 ) 2 . (2.21)
𝐾 −1 𝑘 𝑡
This is the standard deviation of average error associated with each 𝜆.

4. Choose the best 𝜆 𝑡 value based on the measures provided.
Given the above procedure, it is quite often useful to plot the average error rates
for each 𝜆 value. Figure 2.1 plots the average misclassification error corresponding
to a sequence of log 𝜆 𝑡 values.4 Based on this, we can identify the value of log 𝜆
that minimises the average error.5 In order to be conservative (apply a slightly higher
penalty), a log 𝜆 value of one standard deviation higher from the minimum value can
also be selected. The vertical dotted lines in Figure 2.1 show both these values.
4 This figure is reproduced from Friedman, Hastie and Tibshirani (2010).

5 The process discussed here is known to be unstable for moderate sample sizes. In order to ensure
robustness, the 𝐾-fold process can be repeated 𝑛 − 1 times and the 𝜆 can be obtained as the average
of these repeated 𝑘-fold cross validations.
52 Chan at al.
Fig. 2.1: LASSO CV Plot
Asymptotic Properties and Statistical Inference
The development of asymptotic properties, such as the Oracle properties as defined

in Chapter 1, for shrinkage estimators with nonlinear least squares or likelihood
objectives is still in its infancy. To the best of the authors’ knowledge, the most general
result to-date is provided in, Fan, Xue and Zou (2014) where they obtained Oracle
properties for shrinkage estimators with convex objectives and concave regularizers.
This covers the case of LASSO, Ridge, SCAD and Adaptive LASSO with likelihood
objectives for the Logit, Probit and Poisson models.
However, as discussed in Chapter 1, the practical usefulness of the Oracle properties
has been scrutinised by Leeb and Pötscher (2005) and Leeb and Pötscher (2008).
The use of point-wise convergence in developing these results means the number of
observations required for the Oracle properties to be relevant depends on the true
parameters, which are unknown in practice. As such, it is unclear if valid inference
can be obtained purely based on the Oracle properties.
For linear model, Chapter 1 shows that it is possible to obtain valid inference for
Partially Penalised Estimator, at least for the subset of the parameter vector that is not
subject to regularization. Shi, Song, Chen and Li (2019) show that the same idea can
also apply to shrinkage estimators with likelihood objective for Generalised Linear
Model with canonical link. Specifically, their results apply to response variables with
probability distribution function of the form
𝑦 𝑖 x𝑖′ 𝛽 − 𝜓(x𝑖′ 𝛽 )

𝑓 (𝑦 𝑖 ; x𝑖 ) = exp 𝛿(𝑦 𝑖 ) (2.22)
𝜙0
for some smooth functions 𝜓 and 𝛿. This specification includes the Logit and Poisson
models as special cases, but not the Probit model.
Given Equation (2.22), the corresponding
′ log-likelihood function can be derived in
the usual manner. Let 𝛽 = 𝛽 1′ , 𝛽 2′ where 𝛽 1 and 𝛽 2 are 𝑝 1 × 1 and 𝑝 2 × 1 sub-vectors
with 𝑝 = 𝑝 1 + 𝑝 2 . Assume one wishes to test the hypothesis that B𝛽𝛽 1 = C, where B
and C are a 𝑟 × 𝑝 1 matrix and a 𝑝 1 × 1 vector, respectively. Consider the following
Partially Penalised Estimator

𝛽ˆ 𝑃𝑅 , 𝛾ˆ 𝑃𝑅 = arg max log 𝐿(X𝛽𝛽 ,𝛾𝛾 ; y) − 𝜆𝑝(𝛽𝛽 2 ;𝛼
𝛼) (2.23)
𝛾
𝛽 ,𝛾
and the restricted Partially Penalised Estimator, where 𝛽 1 is assumed to satisfy the
restriction B𝛽𝛽 1 = C,

𝛽ˆ 𝑅𝑃𝑅 , 𝛾ˆ 𝑅𝑃𝑅 =arg max log 𝐿 (X𝛽𝛽 ,𝛾𝛾 ; y) − 𝜆𝑝(𝛽𝛽 2 ;𝛼
𝛼) (2.24)
𝛾
𝛽 ,𝛾
s.t. B𝛽𝛽 1 = C, (2.25)
𝛼 ) is a folded concave regularizer, which includes Bridge with 𝛾 > 1 and

where 𝑝(𝛽𝛽 ;𝛼
SCAD as special cases, then the likelihood ratio test statistics
2 log 𝐿( 𝛽ˆ 𝑃𝑅 , 𝛾ˆ 𝑃𝑅 ) − log 𝐿( 𝛽ˆ 𝑅𝑃𝑅 , 𝛾ˆ 𝑅𝑃𝑅 ) ∼ 𝜒2 (𝑟).

𝑑
(2.26)
The importance of the formulation above is that 𝛽 1 , the subset of parameters that
is subject to the restriction/hypothesis, B𝛽𝛽 1 = C, is not part of the regularization. This
is similar to the linear case, where the subset of the parameters that are to be tested
are not part of the regularization. In this case, Shi et al. (2019) show that log-ratio
test as defined in Equation (2.26) has an asymptotic 𝜒2 distribution.
The result as stated in Equation (2.26), however, only applies when the response
variable has the distribution function in the form of Equation (2.22) with a regularizer
that belongs to the folded concave family. While SCAD belongs to this family, other
popular regularizers, such as LASSO, adaptive LASSO and Bridge with 𝛾 ≤ 1,
are not part of this family. Thus, from the perspective of econometrics, the result
above is quite limited as it is relevant only to Logit and Poisson models with a
SCAD regularizer. It does not cover the Probit model or other popular models and
regularizers that are relevant in econometric analysis. Thus, hypothesis testing using
shrinkage estimators for the cases that are relevant in econometrics is still an open
problem.
However, combining the approach as proposed in Chernozhukov, Hansen and
Spindler (2015) as discussed in Chapter 1 with those considered in Shi et al. (2019)
may prove to be useful in progressing this line of research. For example, it is possible
to derive the asymptotic distribution for the Partially Penalised Estimator for the
models considered in Shi et al. (2019) using the Immunization Condition approach
in, Chernozhukov et al. (2015) as shown in Propositions 2.1 and 2.2.
54 Chan at al.
Proposition 2.1 Let 𝑦 𝑖 be a random variable with the conditional distribution as

defined in Equation (2.22) and consider the following Partially Penalised Estimator
with likelihood objective,
𝛽ˆ = arg max 𝑆(𝛽𝛽 ) =𝑙 (𝛽𝛽 ) − 𝜆𝑝(𝛽𝛽 2 ) (2.27)

𝛽
𝑛
∑︁
𝑙 (𝛽𝛽 ) = 𝑦 𝑖 x𝑖′ 𝛽 − 𝜓(x𝑖′ 𝛽 ), (2.28)
𝑖=1
′ 𝛽 + x ′ 𝛽 , and 𝑝(𝛽
where 𝛽 = 𝛽 1′ 𝛽 2′ and x𝑖 = (x1𝑖 , x2𝑖 ), such that x𝑖′ 𝛽 = x1𝑖

1 2𝑖 2
𝛽 ) denotes
2
𝜕 𝑝
the regularizer, such that exists.
𝜕𝛽𝛽 2 , 𝜕𝛽𝛽 2′
Under the assumptions that x𝑖 is stationary with finite mean and variance and
there exists a well-defined 𝜇 for all 𝛽 , such that
−1
𝜕2 𝑝

𝜕𝜓 ′ 𝜕𝜓 ′
𝜇 =− x1𝑖 x2𝑖 x2𝑖 x2𝑖 +𝜆 , (2.29)
𝜕𝑧 𝑖 𝜕𝑧𝑖 𝜕𝛽𝛽 2 𝜕𝛽𝛽 2′
where 𝑧 𝑖 = x𝑖′ 𝛽 , then

√ 𝑑
Γ −1 ΩΓ
𝑛 𝛽ˆ1 − 𝛽 0 → 𝑁 (0,Γ Γ −1 ), (2.30)
where
!2  ! −1
𝜕2𝜓 𝜕2𝜓 2𝜓 2𝑝

′
 ′ 𝜕 ′ 𝜕 ′ 

Γ = − 2 x1𝑖 x1𝑖 + x1𝑖 x
2𝑖 x x
2𝑖 2𝑖 − 𝜆 ′ x x
2𝑖 1𝑖 
𝜕𝑧𝑖 𝜕𝑧2𝑖 

2
𝜕𝑧 𝑖 𝜕𝛽𝛽 2 𝜕𝛽𝛽 2 
 
Ω =𝑉 𝑎𝑟 𝑛𝑆( 𝛽ˆ ) .
√
Proof See Appendix. □

The
result′ above is quite general and covers Logit model when 𝜓(x𝑖′ 𝛽 ) =
log 1 + exp(x𝑖 𝛽 ) and Poisson model when 𝜓(x𝑖′ 𝛽 ) = exp(x𝑖′ 𝛽 ) with various regular-

izers including Bridge and SCAD. But Proposition 2.1 does not cover Probit model.
The same approach can be used to derive the asymptotic distribution of its shrinkage
estimators with likelihood objective as shown in Proposition 2.2
Proposition 2.2 Consider the following Partially Penalised Estimators for a Probit
model
𝛽ˆ =arg max 𝑆(𝛽𝛽 ) = 𝑙 (𝛽𝛽 ) − 𝜆𝑝(𝛽𝛽 2 ) (2.31)

𝛽
𝑛
∑︁
𝑦 𝑖 log Φ(x𝑖′ 𝛽 ) − (1 − 𝑦 𝑖 ) log 1 − Φ(x𝑖′ 𝛽 ) ,

𝑙 (𝛽𝛽 ) = (2.32)
𝑖=1
where 𝛽 = 𝛽 1′ 𝛽 2′ and x𝑖 = (x1𝑖 , x2𝑖 ), such that x𝑖′ 𝛽 = x1𝑖

′ 𝛽 + x ′ 𝛽 and 𝑝(𝛽

1 2𝑖 2
𝛽 ) denotes
2
𝜕 𝑝
the regularizer, such that exists.
𝜕𝛽𝛽 2 , 𝜕𝛽𝛽 2′
Under the assumptions that x𝑖 is stationary with finite mean and variance and
there exists a well-defined 𝜇 for all 𝛽 , such that
Λ (𝛽𝛽 )Θ
𝜇 = −Λ Θ (𝛽𝛽 ), (2.33)
where
𝑛
∑︁
′
Λ (𝛽𝛽 ) = {−𝑦 𝑖 [𝑧 𝑖 + 𝜙(𝑧𝑖 )] 𝜂(𝑧𝑖 )+[ (1 − 𝑦 𝑖 ) [𝜙(𝑧𝑖 ) − 𝑧𝑖 ] 𝜉 (𝑧 𝑖 )}x1𝑖 x2𝑖 ,
𝑖=1
𝑛
∑︁
′ 𝜕2 𝑝
Θ (𝛽𝛽 ) = {−𝑦 𝑖 [𝑧 𝑖 + 𝜙(𝑧𝑖 )] 𝜂(𝑧𝑖 )+[ (1 − 𝑦 𝑖 ) [𝜙(𝑧𝑖 ) − 𝑧𝑖 ] 𝜉 (𝑧 𝑖 )}x𝑖 x2𝑖 −𝜆 ,
𝑖=1
𝜕𝛽𝛽 2 𝜕𝛽𝛽 2′
𝜙(𝑧𝑖 )
𝜂(𝑧𝑖 ) = ,
Φ(𝑧𝑖 )
𝜙(𝑧 𝑖 )
𝜉 (𝑧𝑖 ) = ,
1 − Φ(𝑧𝑖 )
where 𝑧 𝑖 = x𝑖′ 𝛽 , 𝜙(𝑥) and Φ(𝑥) denote the probability density and cumulative
distribution functions for a standard normal distribution respectively. Then
√ 𝑑
Γ −1 ΩΓ
𝑛 𝛽ˆ1 − 𝛽 0 → 𝑁 (0,Γ Γ −1 ), (2.34)
where
Γ =A1 + A2 B−1 A2′ ,

𝑛
∑︁
′ ′
A1 = −𝑦 𝑖 [𝑧𝑖 + 𝜙(𝑧 𝑖 )] 𝜂(𝑧 𝑖 )x1𝑖 x1𝑖 + (1 − 𝑦 𝑖 ) [𝜙(𝑧𝑖 ) − 𝑧𝑖 ] 𝜉 (𝑧 𝑖 )x1𝑖 x1𝑖 ,
𝑖=1
𝑛
∑︁
′ ′
A2 = −𝑦 𝑖 [𝑧𝑖 + 𝜙(𝑧 𝑖 )] 𝜂(𝑧 𝑖 )x1𝑖 x1𝑖 + (1 − 𝑦 𝑖 ) [𝜙(𝑧𝑖 ) − 𝑧 𝑖 ] 𝜉 (𝑧𝑖 )x1𝑖 x2𝑖 ,
𝑖=1
𝑛
∑︁
′ ′ 𝜕2 𝑝
B= −𝑦 𝑖 [𝑧𝑖 + 𝜙(𝑧 𝑖 )] 𝜂(𝑧𝑖 )x1𝑖 x2𝑖 + (1 − 𝑦 𝑖 ) [𝜙(𝑧𝑖 ) − 𝑧 𝑖 ] 𝜉 (𝑧𝑖 )x1𝑖 x2𝑖 −𝜆 ,
𝑖=1
𝜕𝛽𝛽 2 𝜕𝛽𝛽 2′
𝑛𝑆( 𝛽ˆ ) .
√
Ω =𝑉 𝑎𝑟
Proof See Appendix. □

Given the results in Propositions 2.1 and 2.2, it is now possible to obtain the asymptotic
distribution for the Partially Penalised Estimators for the Logit, Probit and Poisson
models under Bridge and SCAD. This should facilitate valid inferences for the subset
of parameters that are not part of the regularizations for these models. These results
56 Chan at al.
should also lead to similar results as those derived in Shi et al. (2019), which would
further facilitate inferences on parameter restrictions using conventional techniques
such as the log-ratio tests. The formal proof of these is left for further research.
2.2.4 Monte Carlo Experiments – Binary Model with shrinkage
This section contains a small set of Monte Carlo experiments examining the finite
sample performance of shrinkage estimators for a binary model6. The aim of this
experiment is to assess both selection and estimation consistency. The experiments
are carried out under two scenarios; with and without the presence of correlation
between the covariates. The shrinkage estimators chosen for this exercise are the
LASSO and Elastic net, and their performance is compared to the standard maximum
likelihood estimator. The details of the data generation process are outlined below.
1. Generate an 𝑛 x 10 matrix of normally distributed covariates with means and
standard deviations, provided in Table 2.1.
2. In the first scenario, there is no correlation between the covariates. This scenario
serves as benchmark for the second scenario where correlation is present in the
covariates. The correlation coefficient is calculated as
𝜌(𝑋𝑖 , 𝑋 𝑗 ) = (0.5) |𝑖− 𝑗 | , (2.35)
where 𝑋𝑖 is the 𝑖th covariate.

3. The true model parameters 𝛽 0 are provided in Table 2.2. A range of values are
selected for 𝛽 0 and given the relative magnitudes of 𝛽 0 , variables X01 and X02
can be considered strongly relevant variables; X03 a weakly irrelevant variable
and remaining variables are considered to be irrelevant.
4. The response variable is generated as follows:
Step 1. Generate 𝜋 (as per a Logit model specification) as:
exp(x𝑖′ 𝛽 0 )
𝜋𝑖 = .
1 + exp(x𝑖′ 𝛽 0 )
Step 2. Draw 𝑢 𝑖 ∼ 𝑈 [0, 1], where 𝑈 [0, 1] denotes a uniform random distribution
with a 0 to 1 domain.
Step 3. If 𝑢 𝑖 < 𝜋𝑖 set, 𝑦 𝑖 = 1 else set 𝑦 𝑖 = 0.
The results from 4000 repetitions on the selection consistency for the three
estimators; LASSO, Elastic net and the standard maximum likelihood estimator of
a Logit model are provided. The analysis is carried out over five sample sizes of
100, 250, 500, 1000 and 2000. For the LASSO and Elastic net, selection consistency
is defined having a nonzero value for a coefficient. For the maximum likelihood
6 The code used to carry out this Monte Carlo experiment is available in the electronic supplementary
materials.
Table 2.1: Means and standard deviations of covariates
Variables X01 X02 X03 X04 X05 X06 X07 X08 X09 X10
Means 0 0 0 0 0 0 0 0 0 0
Std Dev 2 2.5 1.5 2 1.25 3 5 2.75 4 1
Table 2.2: True values of the parameters
𝜷1 𝜷2 𝜷3 𝜷4 𝜷5 𝜷6 𝜷7 𝜷8 𝜷9 𝜷1 0
Coefficient Values 2 -2 0.05 0 0 0 0 0 0 0
case, selection is defined by a 𝑡-value with an absolute value more than 1.96.7 The
selection consistency results across both scenarios (with and without correlation)
are presented as graphs for variables X01, X03 and X05 as they represent a strongly
relevant, weakly relevant and irrelevant variable respectively.
Selection Consistency of X01

1.00
0.75
Selection Consistency
Estimators
Elastic Net
0.50
LASSO
Max Lik
0.25
0.00
100 250 500 1000 2000

Sample Size
Fig. 2.2: X01 Selection Consistency
7 Results where convergence was not achieved were excluded.

58 Chan at al.
When no correlation is present, the strongly relevant variable X01 is selected

consistently by all estimators for sample sizes equal to and above 250 (Figure 2.2).
Both Elastic net and LASSO selected X01 across all sample sizes. However, for a
sample size of 100, the maximum likelihood estimator had a 75% chance of selecting
X01. Given a strongly relevant variable, this demonstrates that shrinkage estimators
outperform the maximum likelihood estimator at smaller sample sizes. This result
remains relatively unchanged when correlation is introduced (Figure 2.3). This
implies that for a strongly relevant variable, the correlation does not affect selection
consistency. For situations where a strongly relevant variable exits and is correlated
with other variables, shrinkage estimators can be used with confidence.
Selection Consistency of X01 with Correlation

1.00
0.75
Estimators
Elastic Net
0.50
LASSO
Max Lik
0.25
0.00
100 250 500 1000 2000

Sample Size
Fig. 2.3: X01 Selection Consistency with Correlation
When detecting a weakly relevant variable, X03, with no correlations present

(Figure 2.4), both shrinkage estimators outperform the maximum likelihood estimator
in all sample sizes. The Elastic net estimator is the best performer by a large margin.
In fact, its selection consistency improves significantly as the sample size increases,
moving from 50% to 75%. Both the LASSO and the maximum likelihood estimator
have a selection consistency of less than 25% across all sample sizes. The LASSO
estimator’s selection consistency remains relatively constant across all sample sizes,
whereas the maximum likelihood estimator does demonstrate a slight improvement
in selection consistency as the sample size increases.

1.00
0.75
Estimators
Elastic Net
0.50
LASSO
Max Lik
0.25
0.00
100 250 500 1000 2000

Sample Size
The results for selection consistency change significantly once correlation is

introduced (Figure 2.5). The overall selection consistency decreases for both the
shrinkage estimators and the maximum likelihood estimator. Unlike the first scenario,
the Elastic net’s selection consistency remains relatively unchanged across all sample
sizes (approximately 50%). However, it still remains the best performer with regard
to selection consistency. Interestingly, the LASSO’s selection consistency decreases
marginally as the sample size increases. Hence, for large sample sizes, the selection
consistency of the maximum likelihood estimator is similar to or slightly better than
LASSO. These results indicate that for weakly relevant variables, the presence of
correlation does affect the selection consistency of shrinkage estimators. In such
circumstances, using Elastic net may be a good choice.
When detecting the irrelevant variable, X05, with no correlations present (Figure
2.6), the maximum likelihood estimator outperforms the two shrinkage estimators.
However, as the sample size increases, LASSO’s selection consistency almost
matches the maximum likelihood estimator. Elastic net performs the worst in this
case, selecting the completely irrelevant variable more than 50% of the time for small
sample sizes and almost 70% of the time for large sample sizes. The results remain
largely unchanged when correlation is introduced (Figure 2.7). This implies that
correlation does not affect the selection consistency of irrelevant variables. In such
60 Chan at al.

1.00
0.75
Estimators
Elastic Net
0.50
LASSO
Max Lik
0.25
0.00
100 250 500 1000 2000

Sample Size
situations, it may be best not to use shrinkage estimators. Although for large samples,
the difference between the maximum likelihood estimator and LASSO is small.
The overall results on selection consistency indicate that correlation has an
impact on weakly relevant variables. The selection consistency results for strongly
relevant and irrelevant variables are not affected by the presence of correlation. In
addition, shrinkage estimators have a higher selection consistency for both strongly
relevant and weakly relevant variables. But, perform poorly for irrelevant variables.
Given these preliminary results, it may be best to be careful and to carry out
further experiments such as, for example, examining selection consistency for Partial
Penalised Estimations.
With regard to estimation, when no correlation is present, the maximum likelihood
estimator for 𝛽1 corresponding to the strongly relevant variable X01, has a mean
approximately equal to the true value (2). However, both the shrinkage estimators
produce estimates with a significant negative bias (Figure 2.8). This is to be expected
given the role of the penalty functions. In practise, shrinkage estimators are used
for selection purposes and subsequently the selected variables are used in the final
model i.e., post model estimation. The results for the correlated case do not differ
significantly.
When no correlation is present, the maximum likelihood estimator for, 𝛽5 corres-
ponding to irrelevant variable X05, has a mean approximately equal to the true value

1.00
0.75
Estimators
Elastic Net
0.50
LASSO
Max Lik
0.25
0.00
100 250 500 1000 2000

Sample Size
(0). The means of both shrinkage estimators are close to their true values. However,
the main difference is in the variance between the maximum likelihood estimator and
the shrinkage estimators. Figure 2.9 shows that the shrinkage estimators have much
lower variance compared to the maximum likelihood estimator. As in the previous
case, the results for the correlated case are identical.
The overall results for estimation imply that shrinkage estimators do have negative
bias and as such should only be used for selection purposes. In addition, the maximum
likelihood estimator has a lower variance compared to shrinkage estimators.
2.2.5 Applications to Econometrics
A general conclusion from the discussion thus far is that while applications of
shrinkage estimators in nonlinear econometric models is promising, some important
ingredients are still missing in order for these estimators to be widely used. The
answers to the questions below may help to facilitate the use of shrinkage estimators
for nonlinear models and thus, could be the focus of future research.
1. Asymptotic and finite sample properties, such as asymptotic distributions and
Oracle Properties, of shrinkage estimators for nonlinear models with likelihood
62 Chan at al.

1.00
0.75
Estimators
Elastic Net
0.50
LASSO
Max Lik
0.25
0.00
100 250 500 1000 2000

Sample Size
objectives. The contributions thus far do not cover a sufficient set of models used
in econometric analysis.
2. Are the asymptotic properties subject to the criticism of Leeb and Pötscher (2005)
and Leeb and Pötscher (2008) and thus render Oracle properties irrelevant from
the practical viewpoint?
3. What are the properties of Partially Penalised Estimators as introduced in Chapter
1 in the case of nonlinear models in general? In the event that the criticism of Leeb
and Pötscher (2005) and Leeb and Pötscher (2008) hold for nonlinear models,
can Partially Penalised Estimators help to facilitate valid inference? While the
results presented in this chapter look promising, more research in this direction
is required.
4. In the case when 𝑝 > 𝑁, shrinkage estimators have good performance in terms of
selection consistency in the linear case, but their finite sample performance for
nonlinear models is not entirely clear. This is crucial since this is one area where
shrinkage estimators offer a feasible solution over conventional estimators.
Estimator Consistency of X01
800
600
Elastic
Count
400
LASSO
Max Lik
200
1.5 2.0 2.5

Coefficient Estimate
Fig. 2.8: Estimation of strongly relevant variable, X01
2.3 Overview of Tree-based Methods - Classification Trees and

Random Forest
In this section, tree-based methods for classification and regression are covered.
In contrast to the previous section, which covered parametric methods, tree based
methods are nonparametric. As such, they assume very little about the data generating
process, which has led to their use in a variety of different applications.
Tree based methods are similar to decision trees, where the hierarchical structure
allows users to make a series of sequential decisions to arrive at a conclusion. Given
this structure, the decisions can be thought of as inputs 𝑥, and the conclusion as
output, 𝑦. Interestingly, one of the earliest papers on the automatic construction of
decision trees – Morgan and Sonquist (1963) – was coauthored by an economist.
However, it was only since the publication of Breiman, Friedman, Olshen and Stone
(1984) that tree based methods gained prominence. Based on this seminal work, two
types of trees – Classification and Regression Trees (CART) – have become popular.
Both types of trees have a similar framework/structure, and the main difference is the
nature of the output variable, 𝑦. In the case of classification trees, 𝑦 is categorical
or nominal, whereas for regression trees, 𝑦 is numerical. In the ML literature, this
distinction is referred to as qualitative versus quantitative, see James, Witten, Hastie
and Tibshirani (2013) for further discussion.
64 Chan at al.
Estimator Consistency of X05
3000
2000 Elastic
Count
LASSO
Max Lik
1000
−0.2 −0.1 0.0 0.1 0.2

Coefficient Estimate
Fig. 2.9: Estimation of relevant variable, XO5
The popularity of trees is primarily due to their decision-centric framework, i.e.,

by generating a set of rules based inputs 𝑥, trees are able to predict an outcome 𝑦. The
resulting structure of a tree is easy to understand, interpret and visualise. The most
significant inputs 𝑥 are readily identifiable from a tree. This is useful with regard
to data exploration, and requires no or limited statistical knowledge to interpret the
results. In most cases, the data requires relatively less preprocessing to generate a tree
compared to other regression and classification methods. Given these advantages,
trees are used to model numerical and categorical data. Next, an outline of the tree
building process is provided.
The process of building a tree involves recursively partitioning the regressor space
(𝑋𝑖 ) to predict 𝑦 𝑖 . The partitioning yields 𝐽 distinct and non-overlapping regions
𝑅1 , 𝑅2 , . . . 𝑅 𝐽 . For all observations that fall into a region 𝑅 𝑗 , the same prediction
applies to all of them. In the case of regression trees (numerical 𝑦), this is simply the
mean of 𝑦 for that region. For classification trees, the prediction is the most occurring
case/category in that region. A natural question is ‘how is the partitioning/splitting of
the variable carried out?’. The approach used is known as recursive binary splitting.
The tree begins by selecting a variable 𝑋 𝑗 and a corresponding threshold value 𝑠 such
that splitting the regressor space into regions {𝑋 |𝑋 𝑗 < 𝑠} and {𝑋 |𝑋 𝑗 ≥ 𝑠} leads to
the largest possible reduction in error, where ‘error’ is a measure of how well the
predictions match actual values. This process of splitting into regions by considering
all possible thresholds across all variables which minimises the error continues until
a stopping criterion is reached; for example, a minimum number of observations in a
particular region. In the case of regression trees (numerical 𝑦), the residual sum of
squares (RSS) is used, defined as
𝐽 ∑︁
∑︁
RSS = (𝑦 𝑖 − 𝑦ˆ 𝑅 𝑗 ) 2 , (2.36)
𝑗=1 𝑖 ∈𝑅 𝑗
where 𝑦ˆ 𝑅 𝑗 is the mean of 𝑦 for region 𝑅 𝑗 . Minimising RSS implies that all squared
deviations for a given region are minimised across all regions. In a classification
setting, RSS cannot be used as a criterion for splitting variables, since we cannot use
the mean values of 𝑦, as it is categorical. There are however a number of alternative
measures, each with its own strengths and limitations. The classification error rate
(CER) is defined as the fraction of observations that do not belong to the most
common class/category,
𝐶𝐸 𝑅 = 1 − max ( 𝑝ˆ 𝑚𝑘 ), (2.37)
𝑘
where 𝑝ˆ 𝑚𝑘 denotes the proportion of observations in the 𝑚 𝑡 ℎ region that are from
the 𝑘 𝑡 ℎ class. In addition to CER, there are two other measures: the Gini index (also
known as Gini impurity) and Cross Entropy. The Gini index is defined as
𝐾
∑︁
𝐺= 𝑝ˆ 𝑚𝑘 (1 − 𝑝ˆ 𝑚𝑘 ). (2.38)
𝑘=1
Small values of the Gini index indicate that the region predominantly contains
observations from a single category.8 The cross-entropy measure is defined as
𝐾
∑︁
𝐶𝐸𝑛𝑡 = − 𝑝ˆ 𝑚𝑘 log 𝑝ˆ 𝑚𝑘 . (2.39)
𝑘=1
This is based on Shannon’s Entropy (Shannon (1948))) which forms the basis of
Information theory and is closely related to the Gini Index.
Based on the above, the tree building process is as follows:
1. Start with the complete data, consider splitting variables 𝑋 𝑗 at threshold point 𝑠
in order to obtain the first two regions:
𝑅1 ( 𝑗, 𝑠) = {𝑋 |𝑋 𝑗 ≤ 𝑠} and 𝑅2 ( 𝑗, 𝑠) = {𝑋 |𝑋 𝑗 > 𝑠}. (2.40)
2. Evaluate the following in order to determine the splitting variable 𝑋 𝑗 and threshold
point 𝑠
8 Even though both the Gini index in classification and that used in inequality analyses, are measures
of variation, they are not equivalent.
66 Chan at al.
 ∑︁ ∑︁ 
(𝑦 𝑖 − 𝑦¯ 𝑅1 ) 2 + (𝑦 𝑖 − 𝑦¯ 𝑅2 ) 2  ,
 
min  (2.41)
𝑗,𝑠 
𝑥 ∈𝑅
 𝑖 1 ( 𝑗,𝑠) 𝑥𝑖 ∈𝑅2 ( 𝑗,𝑠) 

where 𝑦¯ 𝑅𝑚 is the average of 𝑦 in region 𝑚. Note that this formulation relates to a
regression tree as it is based on RSS; for a classification tree, RSS is replaced
with either the Gini index or Cross-Entropy measure.
3. Having found the first split, repeat this process on the two resulting region 𝑠 (𝑅1
and 𝑅2 ). Continue to apply the above minimisation step on all resulting regions.
4. Stop if there is a low number of observations in a region (say 5 or less).
For a large set of variables, the process described above most likely yields an
overfitted model. As such, it is important to take into consideration the size of a
tree |𝑇 |. This is defined as the number of sub-samples/regions that are not split any
further, or the number of terminal nodes (leaves) of a tree (see section below). A
large tree may overfit the training data, and a small tree may not be able to provide an
accurate fit. Given this trade-off, the act of reducing tree size is done in order to avoid
overfitting. For a given tree, a measure that captures both accuracy and tree size is
defined as
RSS + 𝛼 |𝑇 |, (2.42)
where 𝛼 is the tuning parameter/penalty. The tuning parameter is selected using cross
validation. Note, for classification trees, we substitute the Gini Index with the RSS.
Unlike, the case for shrinkage estimators, this tuning parameter does not have any
theoretical foundations.
2.3.1 Conceptual Example of a Tree
This section introduces some specific CART terms and describes how a small tree
might be constructed. It also provides a brief overview of an evaluation measure for
classification problems. The very top of a tree where the first split takes place is
known as the root node. A branch is a region resulting from a partition of a variable.
If a node does not have branches, it is labelled as a terminal node or leaf. The process
of building or growing a tree consists of increasing the number of branches. The
process of decreasing/cutting the number of branches (Equation 2.42) is known as
pruning.
For ease of presentation, define ∧ and ∨ as the ‘and’ and ‘or’ operator, respectively,
and let the usual indicator function 𝐼 (.) = 1 if its argument is true, and zero otherwise.
Consider the following model where the response variable depends on two
covariates, 𝑥1 and 𝑥2
𝑦 𝑖 =𝛽1 𝐼 (𝑥 1𝑖 < 𝑐 1 ∧ 𝑥 2𝑖 < 𝑐 2 ) + 𝛽2 𝐼 (𝑥1𝑖 < 𝑐 1 ∧ 𝑥2𝑖 ≥ 𝑐 2 )

(2.43)
+ 𝛽3 𝐼 (𝑥 1𝑖 ≥ 𝑐 1 ∧ 𝑥2𝑖 < 𝑐 2 ) + 𝛽4 𝐼 (𝑥1𝑖 ≥ 𝑐 1 ∧ 𝑥 2𝑖 ≥ 𝑐 2 ) + 𝑢 𝑖 .
As shown in Figure 2.10, the tree regression divides the (𝑥1 , 𝑥 2 ) space into four
regions (branches). Within each region, the prediction of the response variable is the
same for all values of 𝑥1 and 𝑥2 belonging to that region. In this particular example,
if 𝑐 1 = 2 and 𝑐 2 = 3 the four regions/branches are:
{(𝑥1 , 𝑥 2 ) : 𝑥 1 < 3 ∧ 𝑥2 < 2}

{(𝑥1 , 𝑥 2 ) : 𝑥 1 < 3 ∧ 𝑥2 ≥ 2}
{(𝑥1 , 𝑥 2 ) : 𝑥 1 ≥ 3 ∧ 𝑥2 < 2}
{(𝑥1 , 𝑥 2 ) : 𝑥 1 ≥ 3 ∧ 𝑥2 ≥ 2}.
Based on this setup, the predicted 𝑦 values in regions is given the corresponding 𝛽.
Figure 2.11 displays the same splits using a hierarchical tree, which is how most of
the tree based models are displayed.
Fig. 2.10: Tree Based Regression
One popular way to evaluate the performance of a binary classification tree is by

using a confusion matrix (Figure 2.12). This matrix directly compares each of the
predicted values to their corresponding actual values. As such, this matrix classifies
all observations into one of four categories. The first two being True Positives (TP)
and True Negatives (TN). These categories count all the observations where the
predictions are consistent with actual/observed value. The remaining two categories
False Positives (FP) and False Negative (FN) count all the observations where the
predictions are not consistent with the actual/observed values. Based on the counts
across the four categories, a variety of metrics can be calculated to measure different
facets of the model, such as overall accuracy. Note that the confusion matrix can be
68 Chan at al.
𝑥2 < 2 𝑥2 ≥ 2
𝑥1 < 3 𝑥1 ≥ 3 𝑥1 < 3 𝑥1 ≥ 3
𝑦ˆ = 𝛽1 𝑦ˆ = 𝛽3 𝑦ˆ = 𝛽2 𝑦ˆ = 𝛽4
Fig. 2.11: Tree view of the Regression Tree
calculated in any classification model. This includes the Logit or Probit model used
in econometrics. This allows the user to not only compare different versions of the
same model, but also compare performance across different types of models. For
example, a confusion matrix based on a Logit model can be compared to a confusion
matrix resulting from a classification tree.
Fig. 2.12: Confusion Matrix
2.3.2 Bagging and Random Forests
As noted by Breiman (1996b), trees are well-known to be unstable as a small change

in the training sample can cause the resulting tree structure to change drastically.
This change implies that the predictions will also change, leading to high variance in
predictions. Given this, it is not surprising that trees are generally out-performed by
other classification and regression methods with regard to predictive accuracy. This is
due to the fact that single trees are prone to overfitting, and pruning does not guarantee
stability. Two other popular methods that can be applied to trees to improve their
predictive performance are Bagging and Random Forests. Both methods are based
on aggregation, whereby the overall prediction is obtained by combining predictions

from many trees. Brief details of both these approaches are discussed below.
Bagging (Bootstrap Aggregation) consists of repeatedly taking samples and
constructing trees from each sample, and subsequently combining the predictions
from each tree in order to obtain the final prediction. The following steps outline this
process:
1. Generate 𝐵 different bootstrapped samples of size 𝑛 (with replacement).
2. Build a tree on each sample and obtain a prediction for a given 𝒙, 𝑓ˆ𝑏 (𝒙).
3. Compute the average of all 𝐵 predictions to get the final bagging prediction,
𝐵
1 ∑︁ ˆ
𝑓ˆbagging (𝒙) = 𝑓𝑏 (𝒙). (2.44)
𝐵 𝑏=1
4. The above steps work fine for regression trees. With regard to classification trees,
the final prediction consists of selecting the most occurring category/class among
the 𝐵 predictions.
As argued in Breiman (1996a), the bagging procedure described above improves
predictions by reducing their variance. The random forests concept extends the
bagging approach by constraining the number of covariates in each bootstrapped
sample. In other words, only a subset of all covariates are used to build each tree.
√
As a rule of thumb, if there are a total of 𝑝 covariates, then only 𝑝 covariates are
selected for each bootstrapped sample. The following steps outline the random forests
approach:
1. Draw a random sample (with replacement) of size 𝑛.
√
2. Randomly select 𝑚 covariates from the full set of 𝑝 covariates (where 𝑚 ≈ 𝑝).
3. Build a tree using the selected 𝑚 covariates and obtain a prediction for a given 𝒙,
𝑓ˆ𝑏,𝑚 (𝒙). This is the prediction of the 𝑏th tree based on 𝑚 selected covariates.
4. Repeat steps 1 to 3 for all 𝐵 bootstrapped samples.
5. Compute the average of individual tree predictions in order to obtain the random
forest prediction, 𝑓ˆ𝑟 𝑓
𝐵
1 ∑︁ ˆ
𝑓ˆ𝑟 𝑓 (𝒙) = 𝑓𝑏,𝑚 (𝒙). (2.45)
𝐵 𝑏=1
6. In the classification setting, each bootstrapped tree predicts the most commonly oc-
curring category/class and the random forest predictor selects the most commonly
occurring class across all 𝐵 trees.
Note that bagging is a special case of random forest, when 𝑚 = 𝑝 (step 2). Similar
to the Bagging approach, the random forests approach reduces the variance, but
unlike bagging, the random forest approach also reduces the correlation across the
bootstrapped trees (due to the fact that not all trees have the same covariates). By
allowing a random selection of covariates, the splitting process is not the same for
each tree, and this seems to improve the overall prediction. Several numerical studies
show that random forests perform well across many applications. These aggregation
70 Chan at al.
methods are related to the concept of forecast combinations, which is popular in

mainstream econometrics. The next section contains further information on the
applications and connection of trees to econometrics.
2.3.3 Applications and Connections to Econometrics
For econometricians, both regression and classification trees potentially offer some
advantages compared to traditional regression-based methods. For a start, they are
displayed graphically. This illustrates the decision-making process, and as such
the results are relatively easy to interpret. The tree based approach allows the user
to explore the variables in order to gauge their partitioning ability, i.e., how well
a given variable is able to classify observations correctly using covariates. This
relates to variable importance, which is discussed below. Trees can handle qualitative
variables directly without the need to create dummy variables, and are relatively
robust to outliers. Given enough data, a tree can estimate nonlinear means and
interaction effects without the econometrician having to specify these in advance.
Furthermore, non-constant variance (heteroskedasticity) is also accommodated by
both classification and regression trees. Many easily accessible and tested packages
can be used to build trees. Examples of such packages in R include rpart (Therneau,
Atkinson and Ripley (2015)), caret (Kuhn (2008)) and tree (Ripley (2021)). For
Python, refer to the scikit-learn library (Pedregosa et al. (2011)).
It is also important to consider some limitations of tree based methods. Trees do
not produce any regression coefficients. Hence, it would be challenging to quantify
the relation between the response variable and the covariates. Another consideration
is computational time of trees, especially as the number of covariates increases. In
addition, repeated sampling and tree building (bagging and random forests) are also
computationally demanding. However, in such instances, parallel processing can help
lower computational time. As mentioned earlier, single trees also have a tendency to
overfit and as such offer lower predictive accuracy relative to other methods. But, this
too can be improved by the use of bagging or random forests.
There are two connections that can be made between trees and topics in economet-
rics: nonparametric regression and threshold regression. With regard to the first, trees
can be seen as a form nonparametric regression. Consider a simple nonparametric
model:
𝑦 𝑖 = 𝑚(𝑥𝑖 ) + 𝜀𝑖 , (2.46)
where 𝑚(𝑥𝑖 ) = 𝐸 [𝑦 𝑖 |𝑥 𝑖 ] is the conditional mean and 𝜀𝑖 are error terms. Here, 𝑚 does
not have a parametric form and its estimation occurs at particular values of 𝑥 (local
estimation). The most popular method of nonparametric regression is to use a local
average of 𝑦. Compute the average of all 𝑦 values with some window of 𝑥. By sliding
this window along the domain, an estimate of the entire regression function can be
obtained. In order to improve the estimation, a weighted local average is used, where
the weights are applied to the 𝑦 values. This is achieved using a kernel function. The
size of the estimation window also known as bandwidth is part of the kernel function
itself. Given this, a nonparametric regression estimator which uses local weighted
averaging can be defined as:
𝑛
∑︁
𝑚ˆ 𝑁 𝑊 (𝑥) = 𝑤𝑖 𝑦𝑖 (2.47)
𝑖=1
Í𝑛
𝐾 ℎ (𝑥 − 𝑥 𝑖 )𝑦 𝑖
= Í𝑖=1
𝑛 , (2.48)
𝑗=1 𝐾 ℎ (𝑥 − 𝑥 𝑗 )
where 𝐾 ℎ (𝑥) denotes the scaled kernel function. Note that the sum of the weights is
equal to 1 (the denominator is the normalising constant). The subscript NW credits
the developers of this estimator: Nadaraya (1964) and Watson (1964).
With an appropriate choice of a kernel function and some modifications on the
bandwidth selection, this nonparametric framework can replicate CART. Begin with
a fixed bandwidth and create a series of non-overlapping 𝑥 intervals. For each one of
these intervals, choose a kernel function that computes the sample mean of 𝑦 values
that lie in that interval. This leads to the estimated regression function being a step
function with regular intervals. The final step is to combine the adjacent intervals
where the difference in the sample means (𝑦) is small. The result is a step function
with varying 𝑥 intervals. Each of the end-points of these intervals represents a variable,
splits between the difference in the mean value of 𝑦 when it is significant. This is
similar to the variable splits in the CART.
Replacing kernels with splines also leads to the CART framework. In equation
(2.46) let 𝑚(𝑥𝑖 ) = 𝑠(𝑥𝑖 ), then
𝑦 𝑖 = 𝑠(𝑥𝑖 ) + 𝜀𝑖 , (2.49)
where 𝑠(𝑥) is a 𝑝th order spline and 𝜀𝑖 is the error term, which is assumed to have
zero mean. The 𝑝th order spline can be written as
𝐽
∑︁
𝑠(𝑥) = 𝛽0 + 𝛽1 𝑥 + · · · + 𝛽 𝑝 𝑥 𝑝 + 𝑏 𝑗 (𝑥 − 𝜅 𝑗 )+𝑝 , (2.50)
𝑗=1
where 𝜅 𝑗 denotes the 𝑗th threshold value and (𝛼)+ = max(0, 𝛼).
Based on the above, trees can be thought of as zero order (𝑝 = 0) splines. These
are essentially piece-wise constant functions (step functions). Using a sufficiently
large number of split points, a step function can approximate most functions. Similar
to trees, splines are prone to overfitting, which can result in a jagged and noisy
fit. However, unlike trees, the covariates need to be processed prior to entering the
nonparametric model. For example, categorical variables need to be transformed to a
set of dummy variables.
It is worth noting that the nonparametric regression can also be extended to
accommodate 𝑦 as a categorical or count variable. This is achieved by applying
a suitable link function to the left-hand side of Equation (2.49). Although, the
zero-spline is flexible, it is still a discrete step function, that is, an approximation of
a nonlinear continuous function. In order to obtain smoother fit, shrinkage can be
applied to reduce the magnitude of 𝑏 𝑗 .
72 Chan at al.
∑︁ 𝐽
∑︁
ˆ = min
ˆ 𝑏)
( 𝛽, {𝑦 𝑖 − 𝑠(𝑥𝑖 )}2 + 𝛼 𝑏 2𝑗 . (2.51)
𝛽,𝑏
𝑖 𝑗=1
Another option would be to use the LASSO penalty, which would allow certain
𝑏 𝑗 to be shrunk to zero. This analogous to pruning a tree.
It is important to note that splines of order greater than zero can model non-
constant relationships between threshold points. As such, this provides greater
flexibility compared to trees. However, unlike trees, the splines require pre-specified
thresholds.9 As such, this does not accurately reflect the variable splitting in CART.
In order to accommodate this shortcoming, the Multivariate Adaptive Regression
Splines (MARS) was introduced. This method is based on linear splines and its
adaptive ability consists of selecting thresholds in order to optimise fit to the data,
see Hazelton (2015) for further explanation. Given this, MARS can be regarded a
generalisation of trees. MARS can also handle continuous numerical variables better
than trees. In the CART framework, midpoints of numerical variable were assigned
as potential thresholds, which is limiting.
Recall that trees are zero-order splines. This is similar to a simplified version
of threshold regression, where the only intercepts (between thresholds) are used to
model the relationship between the response and the covariates. As an example, a
single threshold model is written as
𝑌 = 𝛽01 𝐼 (𝑋 ≤ 𝜅) + 𝛽02 𝐼 (𝑋 > 𝜅) + 𝜀, (2.52)

where 𝜅 is the threshold value, and 𝜀 has a mean of zero. Based on this simple model,
if 𝑋 is less than or equal to the threshold value, the estimated value of 𝑌 is equal to the
constant 𝛽01 . This is similar to a split in a regression tree, where the predicted value
is equal to the average value of 𝑌 in that region. This is also evident in the conceptual
example provided earlier, see Equation (2.43). As in the case for splines, threshold
regression can easily accommodate non-constant relationships, between thresholds.
Similarly to the MARS method, this also represents a generalisation of trees.
Given the two connections, econometricians who have used to nonparametric
and/or threshold regressions are able to relate to the proposed CART methods. The
estimation methods for both nonparametric and threshold regression are well-known
in the econometrics literature. In addition, econometricians can also take advantage
of the rich theoretical developments in both of these areas. This includes concepts
such as convergence rates of estimators and their asymptotic properties. These
developments represent an advantage over the CART methods, which to-date have
no or little theoretical developments (see below). This is especially important to
econometric work which covers inference and/or causality. However, if the primary
purpose is illustrating the decision-making process or for exploratory analysis, then
implementing CART may be advantageous. There are tradeoffs to consider when
selecting a method of analysis. For more details on nonparametric and threshold
regression, see Li and Racine (2006) and Hansen (2000) and references within. Based
9 These are also called knots.

on this discussion, the section below covers the limited (and recent) developments for
inference in trees.
Inference
For a single small tree, it may be possible to visually see the role that each variable
plays in producing partitions. In a loose sense, a variable’s partitioning ability
corresponds to its importance/significance. However, as the number of variables
increases, the visualisation aspect may not be so appealing. Furthermore, when the
bagging and/or random forests approach is implemented, it is not possible to represent
all the results into a single tree. This makes interpretability even more challenging,
despite the improvement in predictions. Nevertheless, the importance/significance of
a variable can still be measured by quantifying the increase in RSS (or Gini index)
if the variable is excluded from the tree. Repeating this for 𝐵 trees, the variable
importance score is then computed as the average increase in RSS (or Gini Index).
This process is repeated for all variables and the variable importance score is plotted.
A large score indicates that a removing this variable leads to large increases in RSS
(Gini index) on average, and hence the variable is considered ‘important’.
Figure 2.13 shows a sample variable importance plot reproduced using Kuhn
(2008). Based on the importance score, variable V11 is considered the most important
variable: removing this variable from the tree/s leads to the greatest increase in
the Gini index, on average. The second most important variable is V12 and so on.
Practitioners often use the variable importance to select variables: keeping the ones
with the largest scores and discarding those with lower ones. However, there seems no
agreed cut-off score for this purpose. In-fact, there are no known theoretical properties
of variable importance. As such, variable importance scores are to be considered with
caution. Based on this, variable importance is of limited use with regard to inference.
More recently, trees have found applications in causal inference settings. This
would be of interest to applied econometricians who are interested in modelling
heterogeneous treatment effects. Athey and Imbens (2016) introduces methods for
constructing trees in order to study causal effects, also providing valid inferences for
such. For further details, see Chapter 3.
Given a single binary treatment, the conditional average treatment effect 𝑦(𝒙) is
given by
𝑦(𝒙) = E[𝑦|𝑑 = 1, 𝒙] − E[𝑦|𝑑 = 0, 𝒙], (2.53)
which is the difference between the conditional expected response for the treated
group (𝑑 = 1) and the control group (𝑑 = 0), given a set of covariates (𝒙). Athey and
Imbens (2016) propose a causal tree framework to estimate 𝑦(𝒙). It is an extension of
the classification and regression tree methods described above. Unlike, trees where
the variable split is chosen based on minimising the RSS or Gini Index, causal trees
choose a variable split (left and right) that maximises the squared difference between
the estimated treatment effects; that is, maximise
74 Chan at al.
Fig. 2.13: Variable Importance Chart
∑︁ ∑︁
( 𝑦¯ 1 − 𝑦¯ 0 ) 2 + ( 𝑦¯ 1 − 𝑦¯ 0 ) 2 , (2.54)
left right
where 𝑦¯ 𝑑 is the sample mean of observations with treatment 𝑑. Furthermore, Athey and
Imbens (2016) uses two samples to build a tree. The first sample determines the variable
splits, and the second one is used to re-estimate the treatment effects conditional
on the splits. Using this setup, Athey and Imbens (2016) derive the approximately
Normal sampling distributions for 𝑦(𝒙). An R package called causalTree is available
for implementing Causal Trees.
The first section of this chapter provided a brief overview on regularization for
nonlinear econometric models. This included regularization with both nonlinear least
squares and the likelihood function. Furthermore, the estimation, tuning parameters
and asymptotic properties were discussed in some detail. One of the important
takeaways is that for shrinkage estimators of nonlinear models, the selection con-
sistency and oracle properties do not simply carry over from shrinkage estimators
for linear models. For example, it is not always the case that shrinkage estimators
of nonlinear models can be used for variable selection. This is due to the functional
form of nonlinear models. The exception is when nonlinear model have a single
index form. Examples of this include both the Logit and Probit model. Another
point of difference compared to the linear models’ case is when using the maximum
likelihood function as an objective function for nonlinear models, there are additional
parameters, such as variance, that are part of the shrinkage estimators. If a given
model can be estimated using both least squares or maximum likelihood, it is possible
to compare the resulting tuning parameters across both shrinkage estimators. Based
on this comparison, it is evident that the tuning parameters are not always equal to
each other. Although, commonly used for empirical work, the theoretical properties
of shrinkage estimators for nonlinear models are often unknown. In addition to this,
the computational aspects of these estimators which involve constrained optimisation
are challenging in general. With regard to asymptotic results, there has been some
progress in laying the theoretical foundations. However, given the wide scope of
nonlinear models, much work remains to be done.
This chapter also provided a brief overview of the tree based methods in the machine
learning literature. This consisted of introducing both regression and classification
trees, as well as methods of building these trees. An advantage of trees is that they are
easy to understand, interpret and visualise. However, trees are prone to overfitting,
which makes them less attractive compared to other regression and classification
methods. Additionally, trees tend to unstable, i.e., small changes in the training
data lead to drastic changes in predictions. To overcome this limitation, bagging
and/or random forests approach can be used. There appears to be no or limited
development with regard to inference for tree based techniques. A pseudo inference
measure, variable importance does not have any theoretical basis and as such cannot
be used with confidence. Causal trees, a recent development, offers some inference
capabilities. However, the asymptotic results are provided for the response variable,
i.e., not directly related to the estimation aspect for trees. Given the links between
trees and nonparametric regression, inference procedures from the nonparametric
framework may extend to tree based techniques. Perhaps, these links could aid in the
development of inference for trees. This is a potential area for further research.
76 Chan at al.
Appendix
The Partially Penalised Estimator satisfies the following First Order Necessary
Conditions
𝜕𝑆
=0
𝜕𝛽𝛽 1
𝜕𝑆
=0.
𝜕𝛽𝛽 2
Given 𝜇 as defined in Equation (2.29), define
𝜕𝑆 𝜕𝑆
𝑀 (𝛽𝛽 1 , 𝛽 2 ) = +𝜇 .
𝜕𝛽𝛽 1 𝜕𝛽𝛽 2
Given the definition of 𝜇, it is straightforward to show that 𝑀 ( 𝛽ˆ 1 , 𝛽ˆ 2 ) = 0 if and only

if the First Order Conditions are satisfied. Moreover, it is also straightforward to show
that
𝜕𝑀
=0
𝜕𝛽𝛽 2
after some tedious algebra. Given the immunization condition, the result follows
directly from Proposition 4 in Chernozhukov et al. (2015). This completes the proof.
The proof of Proposition 2.2 follows the same argument as Proposition 2.1.
References
Athey, S. & Imbens, G. (2016). Recursive Partitioning for Heterogeneous Causal

effects. Proceedings of the National Academy of Sciences, 113(27), 7353–7360.
doi: 10.1073/pnas.1510489113
Benjamini, Y. & Hochberg, Y. (1995). Controlling the False Discovery Rate: A
Practical and Powerful Approach to Multiple Testing. Journal of the Royal
Statistical Society. Series B (Methodological), 57(1), 289–300.
Breiman, L. (1996a). Bagging Predictors. Machine Learning, 24(2), 123-140.
Breiman, L. (1996b). Bias, Variance, and Arcing classifiers (Tech. Rep. No. 460).
Berkeley, CA: Statistics Department, University of California at Berkeley.
References 77
Breiman, L., Friedman, J., Olshen, R. & Stone, C. (1984). Classification and
Regression Trees. Wadsworth and Brooks/Cole.
Chernozhukov, V., Hansen, C. & Spindler, M. (2015). Valid Post-Selection and
Post-Regularization Inference: An Elementary, General Approach. Annual
Review of Economics, 7, 649–688.
Fan, J., Xue, L. & Zou, H. (2014). Strong Oracle Optimality of Folded Concave
Penalized Estimation. Annals of Statistics, 42(3), 819–849.
Friedman, J., Hastie, T. & Tibshirani, R. (2010). Regularization Paths for Generalized
Linear Models via Coordinate Descent. Journal of Statistical Software, 33(1),
1–22. Retrieved from https://www.jstatsoft.org/v33/i01/
Hansen, B. E. (2000). Sample Splitting and Threshold Estimation. Econometrica,
68(3), 575–603.
Hastie, T., Tibshirani, R. & Friedman, J. (2009). The Elements of Statistical Learning:
Data Mining, Inference and Prediction. Springer.
Hazelton, M. L. (2015). Nonparametric regression. In J. D. Wright (Ed.), International
Encyclopedia of the Social & Behavioral Sciences (Second Edition) (Second
Edition ed., p. 867-877). Oxford: Elsevier. doi: https://doi.org/10.1016/
B978-0-08-097086-8.42124-0
James, G., Witten, D., Hastie, T. & Tibshirani, R. (2013). An Introduction to
Statistical Learning: with Applications in R. Springer. Retrieved from
https://faculty.marshall.usc.edu/gareth-james/ISL/
Jansen, D. & Oh, W. (1999). Modeling Nonlinearity of Business Cycles: Choosing
between the CDR and Star Models. The Review of Economics and Statistics,
81, 344-349.
Kuhn, M. (2008). Building Predictive Models in R Using the caret Package.
Journal of Statistical Software, Articles, 28(5), 1–26. Retrieved from https://
www.jstatsoft.org/v028/i05 doi: 10.18637/jss.v028.i05
Kwangmoo, K., Seung-Jean, K. & Boyd, S. (2007). An Interior-Point Method for
Large-Scale Logistic Regression. Journal of Machine Learning Research, 8,
1519–1555.
Leeb, H. & Pötscher, B. M. (2005). Model Selection and Inference: Facts and Fiction.
Econometric Theory, 21(1), 21–59.
Leeb, H. & Pötscher, B. M. (2008). Sparse Estimators and the Oracle Property, or
the Return of Hodges’ Estimator. Journal of Econometrics, 142(1), 201–211.
doi: 10.1016/j.jeconom.2007.05.017
Li, Q. & Racine, J. S. (2006). Nonparametric Econometrics: Theory and Practice
(No. 8355). Princeton University Press.
Morgan, J. N. & Sonquist, J. A. (1963). Problems in the Analysis of Survey Data, and
a Proposal. Journal of the American Statistical Association, 58(302), 415–434.
Retrieved from http://www.jstor.org/stable/2283276
Nadaraya, E. A. (1964). On estimating Regression. Theory of Probability and its
Applications, 9, 141–142.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., . . .
Duchesnay, E. (2011). Scikit-learn: Machine Learning in Python. Journal of
Machine Learning Research, 12, 2825–2830.
78 Chan at al.
Ripley, B. (2021). Tree: Classification and Regression Trees [Computer software

manual]. Retrieved from https://cran.r-project.org/web/packages/tree/index
.html (R pacakge version 1.0-41)
Shannon, C. (1948). The mathematical theory of communication. Bell Systems
Technical Journal, 27, 349–423.
Shi, C., Song, R., Chen, Z. & Li, R. (2019). Linear Hypothesis Testing for High
Dimensional Generalized Linear Models. The Annals of Statistics, 47(5). doi:
10.1214/18-AOS1761
Teräsvirta, T. & Anderson, H. (1992). Characterizing Nonlinearities in Business
Cycles using Smooth Transition Autoregressive Models. Journal of Applied
Econometrics, 7, S119-S136.
Therneau, T., Atkinson, B. & Ripley, B. (2015). rpart: Recursive Partitioning
and Regression Trees [Computer software manual]. Retrieved from http://
CRAN.R-project.org/package=rpart (R package version 4.1-9)
Tong, H. (2003). Non-linear Time Series: A Dynamical System Approach. Oxford
University Press.
Watson, G. S. (1964). Smooth Regression Analysis. Sankhyā Ser., 26, 359–372.
Chapter 3
The Use of Machine Learning in Treatment Effect
Estimation
Robert P. Lieli, Yu-Chin Hsu and Ágoston Reguly
Abstract Treatment effect estimation from observational data relies on auxiliary

prediction exercises. This chapter presents recent developments in the econometrics
literature showing that machine learning methods can be fruitfully applied for this
purpose. The double machine learning (DML) approach is concerned primarily
with selecting the relevant control variables and functional forms necessary for the
consistent estimation of an average treatment effect. We explain why the use of
orthogonal moment conditions is crucial in this setting. Another, somewhat distinct,
strand of the literature focuses on treatment effect heterogeneity through the discovery
of the conditional average treatment effect (CATE) function. Here we distinguish
between methods aimed at estimating the entire function and those that project it on
a pre-specified coordinate. We also present an empirical application that illustrates
some of the methods.
3.1 Introduction
It is widely understood in the econometrics community that machine learning (ML)

methods are geared toward solving prediction tasks (see Mullainathan & Spiess,
2017). Nevertheless, most applied work in economics goes beyond prediction and is
often concerned with estimating the average effect of some policy or treatment on
an outcome of interest in a given population. A question of first order importance
Robert P. Lieli B
Central European University, Budapest, Hungary and Vienna, Austria. e-mail: lielir@ceu.edu
Yu-Chin Hsu
Academia Sinica, Taipei, Taiwan; National Central University and National Chengchi University,
Taipei, Taiwan. e-mail: ychsu@econ.sinica.edu.tw
Ágoston Reguly
Central European University, Budapest, Hungary and Vienna, Austria. e-mail: reguly_agoston@phd
.ceu.edu

80 Lieli at al.
is therefore identification — comparing the average outcome among participants

and non-participants, can the difference be attributed to the treatment? If treatment
status is randomly assigned across population units, then the answer is basically yes.
However, if the researcher works with observational data, meaning that one simply
observes what population units chose to do, then participants may be systematically
different from non-participants and the difference in average outcomes is generally
tainted by selection bias.
There are standard frameworks in econometrics to address the identification of
treatment effects. In the present chapter we adopt one of the most traditional and
fundamental settings in which the researcher has at their disposal a cross-sectional
data set that includes a large set of pre-treatment measurements (covariates) assumed
to be sufficient for adjusting for selection bias. The formalization of this idea is the
widely used unconfoundedness assumption, also known as selection-on-observables,
ignorability, conditional independence, etc. This assumption forms the basis of many
standard methods used for average treatment effect estimation such as matching,
regression, inverse propensity weighting, and their combinations (see e.g., Imbens &
Wooldridge, 2009).
Assessing whether identification conditions hold in a certain setting is necessary
for estimating treatment effects even if ML methods are to be employed. For example,
in case of the unconfoundedness assumption, one has to rely on subject matter theory
to argue that there are no unobserved confounders (omitted variables) affecting
the outcome and the treatment status at the same time.1 If the identification of the
treatment effect is problematic in this sense, the use of machine learning will not fix
it.
Nevertheless, economic theory has its own limitations. It may well suggest missing
control variables, but given a large pool of available controls, it is usually not specific
enough to select the most relevant ones or help decide the functional form with which
they should enter a regression model. This is where machine learning methods can
and do help. More generally, ML is useful in treatment effect estimation because the
problem involves implicit or explicit predictive tasks. For example, the first stage of
the two-stage least squares estimator is a predictive relationship, and the quality of
the prediction affects the precision of the estimator. It is perhaps less obvious that
including control variables in a regression is also a prediction exercise, both of the
treatment status and the outcome. So is the specification and the estimation of the
propensity score function, which is used by many methods to adjust for selection bias.
Another (related) reason for the proliferation of ML methods in econometrics is
their ability to handle high dimensional problems in which the number of variables is
comparable to or even exceeds the sample size. Traditional nonparametric estimators
are well known to break down in such settings (‘the curse of dimensionaliy’) but
even parametric estimators such as ordinary least squares (OLS) cannot handle a high
variable to sample size ratio.
In this chapter we present two subsets of the literature on machine learning aided
causal inference. The first came to be known as double or debiased machine learning
1 There are rare circumstances in which the assumption is testable by statistical methods (see Donald,
Hsu & Lieli, 2014), but in general it requires a substantive theory-based argument.
3 The Use of Machine Learning in Treatment Effect Estimation 81
(DML) and has its roots in an extensive statistical literature on semiparametric

estimation with orthogonal moment conditions. In the typical DML setup the model
(e.g., a regression equation) contains a parameter that captures the treatment effect
under the maintained identifying assumptions as well as nuisance functions (e.g.,
a control function) that facilitate consistent estimation of the parameter of interest.
The nuisance functions are typically unknown conditional expectations; these are
the objects that can be targeted by flexible ML methods. The applicability of ML
turns out to depend on whether the nuisance functions enter the model in a way that
satisfies an orthogonality condition. It is often possible to transform a model to satisfy
this condition, usually at the cost of introducing an additional nuisance function (or
functions) to be estimated.
The second literature that we engage with in this chapter is aimed at estimating
heterogeneous treatment effects using ML methods. More specifically, the object of
interest is the conditional average treatment effect (CATE) function that describes
how the average treatment effect changes for various values of the covariates. This
is equivalent to asking if the average treatment effect differs across subgroups such
as males vs. females or as a function of age, income, etc. The availability of such
information makes it possible to design better targeted policies and concentrate
resources to where the expected effect is higher. One strand of this literature uses
‘causal trees’ to discover the CATE function in as much detail as possible without prior
information, i.e., it tries to identify which variables govern heterogeneity and how
they interact. Another strand relies on dimension reduction to estimate heterogeneous
effects as a function of a given variable as flexibly as possible while averaging out the
rest of the variables.
With machine learning methods becoming ubiquitous in treatment effect estimation,
the number of papers that summarize, interpret and empirically illustrate the research
frontier to broader audiences is also steadily increasing (see e.g., Athey & Imbens,
2019, Kreif & DiazOrdaz, 2019, Knaus, 2021, Knaus, Lechner & Strittmatter, 2021,
Huber, 2021). So, what is the value added of this chapter? We do not claim to provide
a comprehensive review; ML aided causal inference is a rapidly growing field and it
is scarcely possible to cover all recent developments in this space. Rather, we choose
to include fewer papers and focus on conveying the fundamental ideas that underlie
the literature introduced above in an intuitive way. Our hope is that someone who
reads this chapter will be able to approach the vast technical literature with a good
general understanding and find their way around it much easier.
While the chapter also foregoes the discussion of most implementation issues, we
provide an empirical illustration concerned with estimating the effect of a mother’s
smoking during pregnancy on the baby’s birthweight. We use DML to provide an
estimate of the average effect (a well-studied problem) and construct a causal tree to
discover treatment effect heterogeneity in a data-driven way. The latter exercise, to
our knowledge, has not been undertaken in this setting. The DML estimates of the
average ‘smoking effect’ we obtain are in line with previous results, and are virtually
identical to the OLS benchmark and naive applications of the Lasso estimator. Part of
the reason for this is that the number of observations is large relative to the number
of controls, even though the set of controls is also large. The heterogeneity analysis
82 Lieli at al.
confirms the important role of mother’s age already documented in the literature but
also points to other variables that may be predictive of the magnitude of the treatment
effect.
The rest of the chapter is organized as follows. In Section 3.2 we outline a standard
treatment effect estimation framework under unconfoundedness and show more
formally where the DML and the CATE literature fits in, and what the most important
papers are. Section 3.3 is then devoted to the ideas and procedures that define the DML
method with particular attention to why direct ML is not suitable for treatment effect
estimation. Section 3.4 presents ML methods for estimating heterogeneous effects. We
distinguish between the causal tree approach (aimed at discovering the entire CATE
function) and dimension reduction approaches (aimed at discovering heterogeneity
along a given coordinate). Section 3.5 presents the empirical applications and Section
3.6 concludes.
3.2 The Role of Machine Learning in Treatment Effect

Estimation: a Selection-on-Observables Setup
Let 𝑌 (1) and 𝑌 (0) be the potential outcomes associated with a binary treatment
𝐷 ∈ {0, 1} and let 𝑋 stand for a vector of predetermined covariates.2 The (hypothetical)
treatment effect for a population unit is given by 𝑌 (1) −𝑌 (0). Modern econometric
analysis typically allows for unrestricted individual treatment effect heterogeneity and
focuses on identifying the average treatment effect (ATE) or the conditional average
treatment effect (CATE) given the possible values of the full covariate vector 𝑋. These
parameters are formally defined as 𝜏 = 𝐸 [𝑌 (1) −𝑌 (0)] and 𝜏(𝑋) = 𝐸 [𝑌 (1) −𝑌 (0)| 𝑋],
respectively.
The fundamental identification problem is that for any individual unit only the
potential outcome corresponding to their actual treatment status is observed—the
counterfactual outcome is unknown. More formally, the available data consists of
𝑛 , where 𝑌 = 𝑌 (0) + 𝐷 [𝑌 (1) − 𝑌 (0)]. In order to
a random sample {(𝑌𝑖 , 𝐷 𝑖 , 𝑋𝑖 )}𝑖=1
identify ATE and CATE from the joint distribution of the observed variables, we
make the selection-on-observables (unconfoundedness) assumption, which states that
the potential outcomes are independent of the treatment status 𝐷 conditional on 𝑋:

(𝑌 (1),𝑌 (0)) ⊥ 𝐷 𝑋. (3.1)
Condition (3.1) can be used to derive various identification results and corres-
ponding estimation strategies for 𝜏 and 𝜏(𝑋).3 Here we use a regression framework
as our starting point. We can decompose the conditional mean of 𝑌 ( 𝑗) given 𝑋 as
2 Chapter 5 of this volume provides a more detailed account of the potential outcome framework.
3 For these strategies (such as regression, matching, inverse probability weighting) to work in
practice, one also needs the overlap assumption to hold. This ensures that in large samples there is a
sufficient number of treated and untreated observations in the neighborhood of any point 𝑥 in the
support of 𝑋(see Imbens & Wooldridge, 2009).
𝐸 [𝑌 ( 𝑗)|𝑋] = 𝜇 𝑗 + 𝑔 𝑗 (𝑋), 𝑗 = 0, 1,
where 𝜇 𝑗 = 𝐸 [𝑌 ( 𝑗)] and 𝑔 𝑗 (𝑋) is a real-valued function of 𝑋 with zero mean. Then
we can write 𝜏 = 𝜇1 − 𝜇0 for the average treatment effect and 𝜏(𝑋) = 𝜏 + 𝑔1 (𝑋) − 𝑔0 (𝑋)
for the conditional average treatment effect function. Given assumption (3.1), one
can express the outcome 𝑌 as the partially linear regression model
𝑌 = 𝜇0 + 𝑔0 (𝑋) + 𝜏𝐷 + [𝑔1 (𝑋) − 𝑔0 (𝑋)]𝐷 + 𝑈, (3.2)
where 𝐸 [𝑈|𝐷, 𝑋] = 0. One can obtain a textbook linear regression model from
(3.2) by further assuming that 𝑔0 (𝑋) = 𝑔1 (𝑋) = [𝑋 − 𝐸 (𝑋)] ′ 𝛽 or, more generally,
𝑔0 (𝑋) = [𝑋 − 𝐸 (𝑋)] ′ 𝛽0 and 𝑔1 (𝑋) = [𝑋 − 𝐸 (𝑋)] ′ 𝛽1 .
The scope for employing machine learning methods in estimating model (3.2) arises
at least in two different ways. First, the credibility of the unconfoundedness assumption
(3.1) hinges on the researcher’s ability to collect a rich set of observed covariates that
are predictive of both the treatment status 𝐷 and the potential outcomes (𝑌 (1),𝑌 (0)).
Therefore, the vector 𝑋 of potential controls may already be high-dimensional in that
the number of variables is comparable to the sample size. In this case theory-based
variable selection or ad-hoc comparisons across candidate models become inevitable
even if the researcher only wants to estimate simple linear versions of (3.2) by OLS.
Machine learning methods such as the Lasso or 𝐿 2 -boosting (as proposed by Kueck,
Luo, Spindler & Wang, 2022) offer a more principled and data-driven way to conduct
variable selection. Nevertheless, as we will shortly see, how exactly the chosen ML
estimator is used for this purpose matters a great deal.
Second, the precise form of the control functions 𝑔0 (𝑋) and 𝑔1 (𝑋) is unknown.
Linearity is a convenient assumption, but it is just that — an assumption. Misspecifying
the control function(s) can cause severe bias in the estimated value of the (conditional)
average treatment effect (see e.g., Imbens & Wooldridge, 2009). The Lasso handles
the discovery of the relevant functional form by reducing it to a variable selection
exercise from an extended covariate pool or ‘dictionary’ 𝑏(𝑋), which contains a set
of basis functions constructed from the raw covariates 𝑋. Thus, it can simultaneously
address the problem of finding the relevant components of 𝑋 and determining the
functional form with which they should enter the model (3.2).
The previous two points are the primary motivations for what came to be known
as the ‘double’ or ‘debiased’ machine learning (DML) literature in econometrics.
Early works include Belloni, Chen, Chernozhukov and Hansen (2012), Belloni,
Chernozhukov and Hansen (2013) and Belloni, Chernozhukov and Hansen (2014b).
Belloni, Chernozhukov and Hansen (2014a) provides a very accessible and intuitive
synopsis of these papers. The research program was developed further by Belloni,
Chernozhukov, Fernández-Val and Hansen (2017), Chernozhukov et al. (2017) and
Chernozhukov et al. (2018). The last paper implements the double machine learning
method in a general moment-based estimation framework.
A third, somewhat distinct, task where machine learning methods facilitate causal
inference is the discovery of treatment effect heterogeneity, i.e., the conditional
average treatment effect function 𝜏(𝑋). An influential early paper in this area is by
84 Lieli at al.
Athey and Imbens (2016). While maintaining the unconfoundedness assumption, they
move away from the model-based regression framework represented by equation (3.2),
and employ a regression tree algorithm to provide a step-function approximation to
𝜏(𝑋). Several improvements and generalizations are now available, e.g., Wager and
Athey (2018), Athey, Tibshirani and Wager (2019). Other approaches to estimating
treatment effect heterogeneity involve reducing the dimension of 𝜏(𝑋) such as in
Semenova and Chernozhukov (2020), Fan, Hsu, Lieli and Zhang (2020), and Zimmert
and Lechner (2019).
We now present a more detailed review of these two strands of the causal machine
learning literature — double machine learning and methods aimed at discovering
treatment effect heterogeneity. These are very rapidly growing fields, so our goal is
not to present every available paper but rather the central ideas.
3.3 Using Machine Learning to Estimate Average Treatment

Effects
3.3.1 Direct versus Double Machine Learning
As discussed in Section 3.2, DML is primarily concerned with variable selection

and the proper specification of the control function rather than the discovery of
treatment effect heterogeneity. In order to focus on the first two tasks, we actually
restrict treatment effect heterogeneity in model (3.2) by assuming 𝑔0 (𝑋) = 𝑔1 (𝑋).4
This yields:
𝑌 = 𝜇0 + 𝑔0 (𝑋) + 𝜏𝐷 + 𝑈. (3.3)
The need for DML is best understood if we contrast it with the natural (though
ultimately flawed) idea of using, say, the Lasso to estimate model (3.3) directly. In
order to employ the Lasso, one must first set up a dictionary 𝑏(𝑋) consisting of
suitable transformations of the components of 𝑋. In most applications this means
constructing powers and interactions between the raw variables up to a certain
order.5 We let 𝑝 denote the dimension of 𝑏(𝑋), which can be comparable to or even
larger than the sample size 𝑛. The Lasso then uses a linear combination 𝑏(𝑋) ′ 𝛽 to
approximate 𝜇0 + 𝑔0 (𝑋) and the model (3.3) is estimated by solving
𝑛
∑︁ 𝑝
∑︁
min [𝑌𝑖 − 𝜏𝐷 𝑖 − 𝑏(𝑋𝑖 ) ′ 𝛽] 2 + 𝜆 |𝛽 𝑘 |
𝜏,𝛽
𝑖=1 𝑘=1
4 This condition does not mean that 𝑌 (1) −𝑌 (0) is the same for all units; however, it does mean that
the distribution of 𝑌 (1) −𝑌 (0) is mean-independent of 𝑋, i.e., that the CATE function is constant.
5 For example, if 𝑋 = (𝑋1 , 𝑋2 , 𝑋3 ), then the dictionary that contains up to second order polynomial
terms is given by 𝑏 (𝑋) = (1, 𝑋1 , 𝑋2 , 𝑋3 , 𝑋12 , 𝑋22 , 𝑋32 , 𝑋1 𝑋2 , 𝑋1 𝑋3 , 𝑋2 𝑋3 ) ′ .
for some 𝜆 > 0. Let 𝜏ˆ 𝑑𝑖𝑟 and 𝛽ˆ 𝑑𝑖𝑟 denote the solution, where the superscript 𝑑𝑖𝑟
stands for ‘direct’. For sufficiently large values of the penalty 𝜆, many components
of 𝛽ˆ 𝑑𝑖𝑟 are exact zeros, which is the reason why Lasso acts as selection operator.
The coefficient on 𝐷 is left out of the penalty term to ensure that one obtains a
non-trivial treatment effect estimate for any value of 𝜆. In practice, 𝜆 may be chosen
by cross-validation (Chapter 1 of this volume provides a detailed discussion of the
Lasso estimator).
While this direct procedure for estimating 𝜏 may seem reasonable at first glance,
𝜏ˆ 𝑑𝑖𝑟 has poor statistical properties. As demonstrated by Belloni et al. (2014b), 𝜏ˆ 𝑑𝑖𝑟
can be severely biased, and its asymptotic distribution is non-Gaussian in general (it
has a thick right tail with an extra mode in their Figure 1). Thus, inference about 𝜏
based on 𝜏ˆ 𝑑𝑖𝑟 is very problematic. Using some other regularized estimator or selection
method instead of Lasso would run into similar problems (see again Chapter 1 for
some alternatives).
The double (or debiased) machine learning procedure approaches the problem of
estimating (3.3) in multiple steps, mimicking the classic econometric literature on
the semiparametric estimation of a partially linear regression model (e.g., Pagan &
Ullah, 1999, Ch. 5). To motivate this approach, we take the conditional expectation
of (3.3) with respect to 𝑋 to obtain
𝐸 (𝑌 |𝑋) = 𝜇0 + 𝜏𝐸 (𝐷|𝑋) + 𝑔0 (𝑋), (3.4)
given that 𝐸 (𝑈|𝑋) = 0. Subtracting equation (3.4) from (3.3) yields an estimating
equation for 𝜏 that is free from the control function 𝑔0 (𝑋) but involves two other
unknown conditional expectations 𝜉0 (𝑋) = 𝐸 (𝑌 |𝑋) and the propensity score 𝑚 0 (𝑋) =
𝐸 (𝐷 |𝑋) = 𝑃(𝐷 = 1|𝑋):
𝑌 − 𝜉0 (𝑋) = 𝜏(𝐷 − 𝑚 0 (𝑋)) + 𝑈. (3.5)
The treatment effect parameter 𝜏 is then estimated in two stages:

(i) One uses a machine learning method—for example, the Lasso—to estimate the
two ‘nuisance functions’ 𝜉0 (𝑋) and 𝑚 0 (𝑋) in a flexible way. It is this twofold
application of machine learning that justifies the use of the adjective ‘double’ in
the terminology.
(ii) One then estimates 𝜏 simply by regressing the residuals of the dependent variable,
𝑌 − 𝜉ˆ0 (𝑋), on the residuals of the treatment dummy, 𝐷 − 𝑚ˆ 0 (𝑋).
There are several variants of the DML procedure outlined above depending on
how the available sample data is employed in executing stages (i) and (ii). In the early
literature (reviewed by Belloni et al., 2014a), the full sample is used in both steps,
i.e., the residuals 𝑌𝑖 − 𝜉ˆ0 (𝑋𝑖 ) and 𝐷 𝑖 − 𝑚ˆ 0 (𝑋𝑖 ), are constructed ‘in-sample,’ for each
observation used in estimating 𝜉0 and 𝑚 0 .
By contrast, the more recent practice involves employing different subsamples
in stages (i) and (ii) as in Chernozhukov et al. (2018). More specifically, we can
partition the full set of observations 𝐼 = {1, . . . , 𝑛} into 𝐾 folds (subsamples) of size
𝑛/𝐾 each, denoted 𝐼 𝑘 , 𝑘 = 1, . . . 𝐾. Furthermore, let 𝐼 𝑘𝑐 = 𝐼 \ 𝐼 𝑘 , i.e., 𝐼 𝑘𝑐 is the set of
86 Lieli at al.
all observations not contained in 𝐼 𝑘 . Setting aside 𝐼1 , one can execute stage (i), the
machine learning estimators, on 𝐼1𝑐 . The resulting estimates are denoted as 𝑚ˆ 0,𝐼1𝑐 and
𝜉ˆ0,𝐼1𝑐 , respectively.
In stage (ii), the residuals are then constructed for the observations in 𝐼1 , i.e., one
computes 𝑌𝑖 − 𝜉ˆ0,𝐼1𝑐 (𝑋𝑖 ) and 𝐷 𝑖 − 𝑚ˆ 0,𝐼1𝑐 (𝑋𝑖 ) for 𝑖 ∈ 𝐼1 . Instead of estimating 𝜏 right
away, steps (i) and (ii) are repeated with 𝐼2 taking over the role of 𝐼1 and 𝐼2𝑐 taking over
the role of 𝐼1𝑐 . Thus, we obtain another set of ‘out-of-sample’ residuals 𝑌𝑖 − 𝜉ˆ0,𝐼2𝑐 (𝑋𝑖 )
and 𝐷 𝑖 − 𝑚ˆ 0,𝐼2𝑐 (𝑋𝑖 ) for 𝑖 ∈ 𝐼2 . We iterate in this way until all folds 𝐼1 , 𝐼2 , . . . , 𝐼 𝐾 are
used up and each observation 𝑖 ∈ 𝐼 has a 𝑌 -residual and a 𝐷-residual associated
with it. Stage (ii) is then completed by running an OLS regression of the full set of
𝑌 -residuals on the full set of 𝐷-residuals to obtain the DML estimate 𝜏ˆ 𝐷 𝑀 𝐿 . This
sample splitting procedure is called the cross-fitting approach to double machine
learning (see Chernozhukov et al., 2017, Chernozhukov et al., 2018).6
Regardless of which variant of 𝜏ˆ 𝐷 𝑀 𝐿 is used, the estimator has well-behaved
statistical properties: it is root-𝑛 consistent and asymptotically normal provided
that the nuisance functions 𝜉0 and 𝑚 0 satisfy some additional regularity conditions.
Thus, one can use 𝜏ˆ 𝐷 𝑀 𝐿 and its OLS standard error to conduct inference about 𝜏
in an entirely standard way. This is a remarkable result because machine learning
algorithms involve a thorough search for the proper specification of the nuisance
function estimators 𝑚ˆ 0 and 𝜉ˆ0 . Despite this feature of machine learning estimators,
post-selection inference is still possible.
As mentioned above, the nuisance functions 𝜉0 and 𝑚 0 need to satisfy some
additional regularity conditions for standard inference based on 𝜏ˆ 𝐷 𝑀 𝐿 to be valid.
Essentially, what needs to be ensured is that the first stage ML estimators converge
sufficiently fast—faster than the rate 𝑛−1/4 . When this estimator is the Lasso, the
required conditions are called sparsity assumptions. These assumptions describe how
‘efficiently’ one can approximate 𝑚 0 and 𝜉0 using linear combinations of the form
𝑏(𝑋) ′ 𝛽. In particular, the assumption is that one can achieve a small approximation
error just by using a few important terms, i.e., a ‘sparse’ coefficient vector 𝛽 (see also
Chapter 1 of this volume for a discussion).
More technically, in the absence of strong functional form assumptions, uniformly
consistent estimation of 𝜉0 and 𝑚 0 requires the inclusion of higher and higher order
polynomial terms into 𝑏(𝑋); hence, in theory, the dimension of 𝑏(𝑋) expands with
the sample size 𝑛. For a corresponding coefficient vector 𝛽 = 𝛽𝑛 , let the ‘sparsity
index’ 𝑠 𝑛 be the number of nonzero components. The sparsity assumption then states
that there exists a sequence of approximations 𝑏(𝑋) ′ 𝛽𝑛 to 𝜉0 and 𝑚 0 such that 𝑠 𝑛
increases slowly but the approximation error still vanishes at a sufficiently fast rate
(and is negligible relative to the estimation error).7 For example, Belloni et al. (2017)
6 Another variant of cross-fitting involves running 𝐾 separate regressions, one over each fold 𝐼𝑘 ,
using the nuisance function estimators 𝑚 ˆ 0,𝐼𝑘𝑐 and 𝜉ˆ0,𝐼𝑘𝑐 . There are 𝐾 resulting estimates of 𝜏, which
can be averaged to obtain the final estimate.
7 A special case of the sparsity assumption is that the functions 𝑚0 and 𝜉0 obey parametric models
linear in the coefficients, i.e., the approximation error 𝑚0 (𝑋) − 𝑏𝑛 (𝑋) ′ 𝛽𝑛 can be made identically
zero for a finite value of 𝑠𝑛 .
specify rigorously the sparsity conditions needed for the full-sample DML estimation
of 𝜏.
The required sparsity assumptions also provide theoretical motivation for the cross-
fitted version of the DML estimator. As pointed out by Chernozhukov et al. (2017), the
sparsity assumptions required of the nuisance functions to estimate ATE are milder
when the first stage estimates are constructed over an independent subsample. In
particular, this induces a tradeoff between how strict the sparsity assumptions imposed
on 𝑚 0 and 𝜉0 need to be — if, say, 𝑚 0 is easy to approximate, then the estimation of
𝜉0 can use a larger number of terms and vice versa. In the full-sample case the sparsity
indices for estimating 𝑚 0 and 𝜉0 have to satisfy strong restrictions individually.8
It is generally true that the split sample approach naturally mitigates biases due to
overfitting (estimating the nuisance functions and the parameter of interest using the
same data), whereas the full sample approach requires more stringent complexity
restrictions, such as entropy conditions, to do so (Chernozhukov et al., 2018). Despite
the theoretical appeal of cross-fitting, we are at present not aware of applications
or simulation studies, apart from illustrative examples in 𝑖𝑏𝑖𝑑., where a substantial
difference arises between the results delivered by the two DML approaches.
3.3.2 Why Does Double Machine Learning Work and Direct Machine
Learning Does Not?
Loosely speaking, direct estimation of (3.3) by Lasso (or some other ML estimator)
fails because it stretches the method beyond its intended use – prediction – and
expects it to produce an interpretable, ‘structural’ coefficient estimate. Mullainathan
and Spiess (2017) provide several insights into why this expectation is misguided in
general. By contrast, the double machine learning approach uses these methods for
their intended purpose only — to approximate conditional expectation functions in a
flexible way. It is however possible to provide a more formal explanation.
For the sake of argument we start by assuming that the control function 𝑔0 (𝑋)
is known. In this case 𝜏 can be consistently estimated by an OLS regression of
𝑌 − 𝑔0 (𝑋) on 𝐷 (and a constant). This estimation procedure is of course equivalent
to the moment condition

𝐸 [𝑈𝐷] = 𝐸 [𝑌 − 𝑔0 (𝑋) − 𝜏𝐷]𝐷 = 0.
Now suppose that instead of the true control function 𝑔0 (𝑋), we are presented
with a somewhat perturbed version, 𝑔0 (𝑋) + 𝑡 [𝑔(𝑋) − 𝑔0 (𝑋)], where 𝑔(𝑋) − 𝑔0 (𝑋)
is the direction of perturbation and 𝑡 > 0 is a scaling factor. When the scalar 𝑡
is sufficiently small, the deviation of the perturbed control function from 𝑔0 is
8 More formally, let 𝑠𝑚,𝑛 and 𝑠 𝜉 ,𝑛 denote the sparsity indices of 𝑚0 and 𝜉0√, respectively. The
full-sample variant of DML requires that both 𝑠𝑚,𝑛 and 𝑠 𝜉 ,𝑛 grow slower than 𝑛. The cross-fitted
variant only requires that 𝑠𝑚,𝑛 · 𝑠 𝜉 ,𝑛 grow slower than 𝑛.
88 Lieli at al.
(uniformly) small. In practice, the perturbed function is the Lasso estimate, which is
subject to approximation error (selection mistakes) as well as estimation error.
How does working with the perturbed control function affect our ability to estimate
𝜏? To answer this question, we can compute the derivative

𝜕𝑡 𝐸 [𝑌 − 𝑔0 (𝑋) − 𝑡 · ℎ(𝑋) − 𝜏𝐷]𝐷 𝑡=0 (3.6)
where ℎ(𝑋) = 𝑔(𝑋) − 𝑔0 (𝑋). This derivative expresses the change in the moment
condition used to estimate 𝜏 as one perturbs 𝑔0 (𝑋) in the direction ℎ(𝑋) by a small
amount. It is easy to verify that (3.6) is equal to −𝐸 [ℎ(𝑋)𝐷], which is generally
non-zero, given that the covariates 𝑋 are predictive of treatment status.
Intuitively, we can interpret this result in the following way. When equation (3.3)
is estimated by Lasso directly, the implicit perturbation to 𝑔0 (𝑋) is the combined
estimation and approximation error ℎ(𝑋) = 𝑏(𝑋) ′ 𝛽ˆ 𝑑𝑖𝑟 − 𝑔0 (𝑋). Due to the impact
of regularization (using a nonzero value of 𝜆), this error can be rather large in finite
samples, even though it vanishes asymptotically under sparseness conditions. More
specifically, Lasso may mistakenly drop components of 𝑋 from the regression that
enter 𝑔0 (𝑋) nontrivially and are also correlated with 𝐷. This causes the derivative
(3.6) to differ from zero and hence induces first-order omitted variable bias in the
estimation of 𝜏.
The DML procedure guards against large biases by using a moment condition with
more favorable properties to estimate 𝜏. As explained in Section 3.3.1, DML first
estimates the conditional mean functions 𝜉0 (𝑋) = 𝐸 (𝑌 |𝑋) and 𝑚 0 (𝑋) = 𝐸 (𝐷 |𝑋), and
then the estimated value of 𝜏 is obtained by a regression of 𝑌 − 𝜉0 (𝑋) on 𝐷 − 𝑚 0 (𝑋).
The second step is equivalent to estimating 𝜏 based on the moment condition
n 2o
𝐸 𝑌 − 𝜉0 (𝑋) 𝐷 − 𝑚 0 (𝑋) − 𝜏 𝐷 − 𝑚 0 (𝑋) = 0. (3.7)
Once again, consider a thought experiment where we replace 𝜉0 and 𝑚 0 with

perturbed versions 𝜉0 + 𝑡 (𝜉 − 𝜉0 ) and 𝑚 0 + 𝑡 (𝑚 − 𝑚 0 ), respectively. To gauge how
these perturbations affect the moment condition used to estimate 𝜏, we can compute
the derivative
n 2 o
𝜕𝑡 𝐸 𝑌 − 𝜉0 − 𝑡 · ℎ 𝜉 𝐷 − 𝑚 0 − 𝑡 · ℎ 𝑚 − 𝜏 𝐷 − 𝑚 0 − 𝑡 · ℎ 𝑚 (3.8)
𝑡=0
where ℎ 𝑚 (𝑋) = 𝑚(𝑋) − 𝑚 0 (𝑋) and ℎ 𝜉 (𝑋) = 𝜉 (𝑋) − 𝜉0 (𝑋). Given that the residuals
𝑌 − 𝜉0 (𝑋) and 𝐷 − 𝑚 0 (𝑋) are uncorrelated with all functions of 𝑋, including the
deviations ℎ 𝑚 (𝑋) and ℎ 𝜉 (𝑋), it is straightforward to verify that
n 2o
𝐸 𝑌 − 𝜉0 − 𝑡 · ℎ 𝜉 𝐷 − 𝑚 0 − 𝑡 · ℎ 𝑚 − 𝜏 𝐷 − 𝑚 0 − 𝑡 · ℎ 𝑚
= 𝐸 (𝑌 − 𝜉0 ) (𝐷 − 𝑚 0 ) + 𝑡 2 𝐸 ℎ 𝑚 (𝑋)ℎ 𝜉 (𝑋) − 𝜏 · 𝑡 2 𝐸 ℎ 𝑚 (𝑋) 2

so that the derivative (3.8) evaluated at 𝑡 = 0 is clearly zero.

This result means that the moment condition (3.7) is robust to small perturbations
around the true nuisance functions 𝜉0 (𝑋) and 𝑚 0 (𝑋). For example, even if Lasso
drops some relevant components of 𝑋, causing a large error in the approximation to

𝑔0 (𝑋), the resulting bias in 𝜏ˆ 𝐷 𝑀 𝐿 is an order smaller. This property is a key reason
why DML works for estimating 𝜏.
3.3.3 DML in a Method of Moments Framework
We can now give a more general perspective on double machine learning that is
not directly tied to a regression framework. In particular, stage (ii) of the procedure
described in Section 3.3.1 can be replaced by any moment condition that identifies
the parameter of interest and satisfies an orthogonality condition analogous to (3.8)
with respect to the unknown nuisance functions involved. (The nuisance functions are
typically conditional means such as the expectation of the outcome or the treatment
indicator conditional on a large set of covariates.) Stage (i) of the general DML
framework consists of the machine learning estimation of the nuisance functions
using a suitable method (such as the Lasso, random forest, etc.). The cross-fitted
version of the estimator is recommended for practical use in general. That is, the first
and the second stages should be conducted over non-overlapping subsamples and the
roles of the subsamples should be rotated.
Chernozhukov et al. (2018) provides a detailed exposition of the general moment
based framework with several applications. The general theory involves stating high
level conditions on the nuisance functions and their first stage machine learning
estimators so that the selection and estimation errors have a negligible impact on the
second stage moment condition. In typical problems, the minimum convergence rate
that is required of the first stage ML estimators is faster than 𝑛−1/4 .
As an example, we follow Chernozhukov et al. (2017) and consider the estimation
of 𝜏 using a moment condition that is, in a sense, even more robust than (3.7). In
particular, let us define 𝜉0 |𝐷=0 (𝑋) = 𝐸 (𝑌 |𝐷 = 0, 𝑋), 𝜉0 |𝐷=1 (𝑋) = 𝐸 (𝑌 |𝐷 = 1, 𝑋),
𝜉0 = (𝜉0 |𝐷=0 , 𝜉0 |𝐷=1 ), and
𝐷 (𝑌 − 𝜉0 |𝐷=1 (𝑋))
𝜓(𝑊, 𝑚 0 , 𝜉0 ) = + 𝜉0|𝐷=1 (𝑋)
𝑚 0 (𝑋)
(1 − 𝐷)(𝑌 − 𝜉0 |𝐷=0 (𝑋))
− − 𝜉0 |𝐷=0 (𝑋),
1 − 𝑚 0 (𝑋)
where 𝑊 = (𝑌 , 𝐷, 𝑋). The true value of 𝜏 is identified by the orthogonal moment

condition
𝐸 [𝜓(𝑊, 𝑚 0 , 𝜉0 ) − 𝜏] = 0. (3.9)
Splitting the sample into, say, two parts 𝐼1 and 𝐼2 , let 𝑚ˆ 0,𝐼𝑘 and 𝜉ˆ0,𝐼𝑘 denote the
first stage ML estimators over the subsample 𝐼 𝑘 . Define 𝜓ˆ 𝑖 as 𝜓(𝑊𝑖 , 𝑚ˆ 0,𝐼2 , 𝜉ˆ0,𝐼2 ) for
𝑖 ∈ 𝐼1 and as 𝜓(𝑊𝑖 , 𝑚ˆ 0,𝐼1 , 𝜉ˆ0,𝐼1 ) for 𝑖 ∈ 𝐼2 . Then the DML estimator of 𝜏 is simply
given by the sample average
90 Lieli at al.
𝑛
1 ∑︁ ˆ
𝜏ˆ 𝐷 𝑀 𝐿 = 𝜓𝑖 .
𝑛 𝑖=1
√
Under regularity conditions mentioned above, the distribution of 𝑛( 𝜏ˆ 𝐷 𝑀 𝐿 − 𝜏) is
asymptotically normal 2
zero and variance 𝐸 [(𝜓 − 𝜏) ], which is consistently
Í𝑛 with mean
estimated by 𝑛−1 𝑖=1 ( 𝜓ˆ 𝑖 − 𝜏ˆ 𝐷 𝑀 𝐿 ) 2 .
Equation (3.9) satisfies the previously introduced orthogonality condition

𝜕𝑡 𝐸 𝜓(𝑊, 𝑚 0 + 𝑡ℎ 𝑚 , 𝜉0 + 𝑡ℎ 𝜉 ) − 𝜏 𝑡=0 = 0,
where ℎ 𝑚 (𝑋) and ℎ 𝜉 (𝑋) are perturbations to 𝑚 0 and 𝜉0 . However, an even stronger
property is true. It is easy to verify that

𝐸 𝜓(𝑊, 𝑚 0 , 𝜉0 + 𝑡ℎ 𝜉 ) − 𝜏 = 0 ∀𝑡 and 𝐸 𝜓(𝑊, 𝑚 0 + 𝑡ℎ 𝑚 , 𝜉0 ) − 𝜏 = 0 ∀𝑡.
(3.10)
This means that if 𝑚 0 is reasonably well approximated, the moment condition for
estimating 𝜏 is completely robust to specification (selection) errors in modeling 𝜉0
and, conversely, if 𝜉0 is well approximated, then specification errors in 𝑚 0 do not
affect the estimation of 𝜏. For example, if Lasso mistakenly drops an important
component of 𝑋 in estimating 𝜉0|𝐷=1 or 𝜉0 |𝐷=0 , this error is inconsequential as long
as this variable (and all other relevant variables) are included in the approximation of
𝑚0.
In settings in which the nuisance functions 𝜉0 and 𝑚 0 are estimated based on
finite dimensional parametric models, the moment condition (3.9) is often said to
be ‘doubly robust.’ This is because property (3.10) implies that if fixed dimensional
parametric models are postulated for 𝜉0 and 𝑚 0 , then consistent estimation of 𝜏 is
still possible even when one of the models is misspecified.
The notion of an orthogonal moment condition goes back to Neyman (1959).
There is an extensive statistical literature spanning several decades on the use of
such moment conditions in various estimation problems. We cannot possibly do
justice to this literature in this space; the interested reader could consult references in
Chernozhukov et al. (2018) and Fan et al. (2020) for some guidance in this direction.
3.3.4 Extensions and Recent Developments in DML
There are practically important recent extensions of the DML framework that consider
treatment effect estimation with continuous treatments or in the presence of mediating
variables. For example, Colangelo and Lee (2022) study the average dose-response
function (the mean of the potential outcome) as a function of treatment intensity.
Let 𝑇 denote a continuous treatment taking values in the set T and let 𝑌 (𝑡) be
the potential outcome associated with treatment intensity 𝑡 ∈ T . The average dose-
response function is given by 𝜇𝑡 = 𝐸 [𝑌 (𝑡)]. Under the weak unconfoundedness
assumption 𝑌 (𝑡) ⊥ 𝑇 | 𝑋 and suitable continuity conditions, 𝜇𝑡 is identified by
" #
1 𝑇 −𝑡 𝑌
𝜇𝑡 = lim 𝐸 𝐾 ,
ℎ→0 ℎ ℎ 𝑓𝑇 |𝑋 (𝑡|𝑋)
where 𝑓𝑇 |𝑋 (𝑡|𝑋) is the conditional density function of 𝑇 given 𝑋 (also called the
generalized propensity score), 𝐾 (·) is a kernel function and ℎ is a bandwidth parameter.
Colangelo and Lee (2022) show that the associated orthogonal moment condition is
" #
1 𝑇 − 𝑡 𝑌 − 𝛾(𝑡, 𝑋)
𝜇𝑡 = lim 𝐸 𝛾(𝑡, 𝑋) + 𝐾 , (3.11)
ℎ→0 ℎ ℎ 𝑓𝑇 |𝑋 (𝑡|𝑋)
where 𝛾(𝑡, 𝑥) = 𝐸 [𝑌 |𝑇 = 𝑡, 𝑋 = 𝑥]. They apply the DML method to estimate 𝜇𝑡 based
on (3.11) and show that their kernel-based estimator is asymptotically normal but
converges at a nonparametric rate.
Using the same framework, Hsu, Huber, Lee and Liu (2022) propose a Cramer-von
Mises-type test for testing whether 𝜇𝑡 has a weakly monotonic relationship with the
treatment dose 𝑡. They first transform the null hypothesis of a monotonic relationship
to countably many moment inequalities where each of the moments can be identified
by an orthogonal moment condition, and√ the DML method can be applied to obtain
estimators converging at the parametric ( 𝑛) rate. They propose a multiplier bootstrap
procedure to construct critical values and show that their test controls asymptotic
size and is consistent against any fixed alternative.
Regarding causal mediation, Farbmacher, Huber, Laffers, Langen and Spindler
(2022) study the average direct and indirect effect of a binary treatment operating
through an intermediate variable that lies on the causal path between the treatment and
the outcome. They provide orthogonal moment conditions for these quantities under
the
√ unconfoundedness assumption and show that the associated DML estimators are
𝑛-consistent and asymptotically normal (see 𝑖𝑏𝑖𝑑. for further details).
Nevertheless, the use of DML extends beyond the estimation of various average
treatment effects and is also applicable to related problems such as learning the
optimal policy (assignment rule) from observational data (Athey & Wager, 2021).
For example, officials may need to choose who should be assigned to a job training
program using a set of observed characteristics and data on past programs. At the
same time, the program may need to operate within a budget constraint and/or the
assignment rule may need to satisfy other restrictions.
To set up the problem more formally, let 𝜋 be a policy that maps a subject’s
characteristics to a 0-1 binary decision (such as admission to a program). The policy
𝜋 is assumed to belong to a class of policies Π, which incorporates problem-specific
constraints pertaining to budget, functional form, fairness, etc. The utilitarian welfare
of deploying the policy 𝜋 relative to treating no one is defined as
𝑉 (𝜋) = 𝐸 [𝑌 (1)𝜋(𝑋) +𝑌 (0) (1 − 𝜋(𝑋))] − 𝐸 [𝑌 (0)] = 𝐸 [𝜋(𝑋) (𝑌 (1) −𝑌 (0))].
The policy maker is interested in finding the treatment rule 𝜋 with the highest
welfare in the class Π, i.e., solving 𝑉 ∗ = max 𝜋 ∈Π 𝑉 (𝜋). A 𝜋 ∗ satisfying 𝑉 (𝜋 ∗ ) = 𝑉 ∗ or
equivalently 𝜋 ∗ ∈ arg max 𝜋 ∈Π 𝑉 (𝜋) is called an optimal treatment rule (the optimal
92 Lieli at al.
treatment rule might not be unique). Under the unconfoundedness assumption, 𝑉 (𝜋)
is identified as
" #
𝐷𝑌 (1 − 𝐷)𝑌
𝑉 (𝜋) = 𝐸 𝜋(𝑋) − , (3.12)
𝑚 0 (𝑋) 1 − 𝑚 0 (𝑋)
or, alternatively, as
"
𝐷 (𝑌 − 𝜉
0 |𝐷=1 (𝑋))
𝑉 (𝜋) = 𝐸 𝜋(𝑋) + 𝜉0|𝐷=1 (𝑋)
𝑚 0 (𝑋)
#
(1 − 𝐷)(𝑌 − 𝜉0|𝐷=0 (𝑋))
− − 𝜉0|𝐷=0 (𝑋) , (3.13)
1 − 𝑚 0 (𝑋)
where 𝜉0 |𝐷=0 , 𝜉0|𝐷=1 and 𝑚 0 are defined as in Section 3.3.3. In studying the policy
learning problem, Kitagawa and Tetenov (2018) assume that the propensity score
function 𝑚 0 is known and estimate 𝑉 (𝜋) based on (3.12) using an inverse probability
weighted estimator. They show that the difference between √ the estimated optimal
welfare and the true optimal welfare decays at the rate of 1/ 𝑛 under suitable control
over the complexity of the class of decision rules. Athey and Wager (2021) extend
Kitagawa and Tetenov (2018) in two aspects. First, Athey and Wager (2021) allow
for the case in which the propensity score is unknown and estimate 𝑉 (𝜋) based on
the orthogonal moment condition (3.13) using the DML approach. They show that
the difference between the estimated optimal√welfare and the true optimal welfare
continues to converge to zero at the rate of 1/ 𝑛 under suitable conditions. Second,
in addition to binary treatments with unconfounded assignment, Athey and Wager
(2021) also allow for endogenous and continuous treatments.
3.4 Using Machine Learning to Discover Treatment Effect

Heterogeneity
3.4.1 The Problem of Estimating the CATE Function
Under the unconfoundedness assumption (3.1), it is the full dimensional conditional

average treatment effect (CATE) function that provides the finest breakdown of the
average treatment effect across all the subpopulations defined by the possible values
of 𝑋. Without any assumptions restricting individual treatment effect heterogeneity,
it is given by
𝜏(𝑋) = 𝐸 [𝑌 (1) −𝑌 (0)| 𝑋] = 𝜏 + 𝑔1 (𝑋) − 𝑔0 (𝑋).
For example, in the context of the empirical application presented in Section 3.5,
𝜏(𝑋) describes the average effect on birthweight of smoking during pregnancy given
the mother’s age, education, various medical conditions, the pattern of prenatal care
received, etc. Of course, some of these variables may be more important in capturing
heterogeneity than others.
While under the treatment effect homogeneity assumption 𝑔0 (𝑋) was simply a
nuisance object, now 𝜏 + 𝑔1 (𝑋) − 𝑔0 (𝑋) has become an infinite dimensional parameter
of interest. Following the same steps as in Section 3.3.1, equation (3.5) generalizes to
𝑌 − 𝜉0 (𝑋) = 𝜏(𝑋)(𝐷 − 𝑚 0 (𝑋)) + 𝑈
In addition to pre-estimating 𝜉0 and 𝑚 0 , the generalization of the DML approach

requires the introduction of some type of structured approximation to 𝜏(𝑋) to
make this equation estimable. For example, one could specify a parametric model
𝜏(𝑋) = 𝑏(𝑋) ′ 𝛽, where the dimension of 𝑏(𝑋) is fixed and its relevant components are
selected based on domain-specific theory and expert judgement. Then 𝛽 can still be
estimated by OLS, but the procedure is subject to misspecification bias. Using instead
the Lasso to estimate 𝛽 brings back the previously described problems associated
with direct machine learning estimation of causal parameters.
But there is an even more fundamental problem—when 𝑋 is high dimensional,
the CATE function may be quite complex and hard to describe, let alone visualize.
While 𝜏(𝑋) could be completely flat along some coordinates of 𝑋, it can be highly
nonlinear in others with complex interactions between multiple components. Suppose,
for example, that one could somehow obtain a debiased series approximation to 𝜏(𝑋).
It might contain terms such as 2.4𝑋3 − 1.45𝑋22 𝑋52 + 0.32𝑋1 𝑋32 𝑋5 + .... Understanding
and analyzing this estimate is already a rather formidable task.
There are two strands of the econometrics literature offering different solutions to
this problem. The first does not give up the goal of discovering 𝜏(𝑋) in its entirety
and uses a suitably modified regression tree algorithm to provide a step-function
approximation to 𝜏(𝑋). Estimated regression trees are capable of capturing and
presenting complex interactions in a relatively straightforward way that is often
amenable to interpretation. A pioneering paper of this approach is Athey and Imbens
(2016) with several follow-ups and extensions such as Wager and Athey (2018) and
Athey et al. (2019). The second idea is to reduce the dimensionality of 𝜏(𝑋) to a
coordinate (or a handful of coordinates) of interest and integrate out the ‘unneeded’
components. More formally, let 𝑋1 be a component or a small subvector of 𝑋.
Abrevaya, Hsu and Lieli (2015) introduce the reduced dimensional CATE function as
𝜏(𝑋1 ) = 𝐸 [𝑌 (1) −𝑌 (0)| 𝑋1 ] = 𝐸 [𝜏(𝑋)| 𝑋1 ],
where the second equality follows from the law of iterated expectations.9 As proposed
by Fan et al. (2020) and Semenova and Chernozhukov (2021), the reduced dimensional
CATE function can be estimated using an extension of the moment-based DML
method, where the second stage of the procedure, which is no longer a high dimensional
problem, employs a traditional nonparametric estimator.
9 It is a slight abuse of notation to denote the functions 𝜏 (𝑋) and 𝜏 (𝑋1 ) with the same letter. We
do so for simplicity; the arguments of the two function will distinguish between them.
94 Lieli at al.
The two approaches to estimating heterogeneous average treatment effects com-

plement each other. The regression tree algorithm allows the researcher to discover
in an automatic and data-driven way which variables, if any, are the most relevant
drivers of treatment effect heterogeneity. The downside is that despite the use of
regression trees, the results can still be too detailed and hard to present and interpret.
By contrast, the dimension-reduction methods ask the researcher to pre-specify a
variable of interest and the heterogeneity of the average treatment effect is explored
in a flexible way focusing on this direction. The downside, of course, is that other
relevant predictors of treatment effect heterogeneity may remain undiscovered.
3.4.2 The Causal Tree Approach
Regression tree basics. Regression trees are algorithmically constructed step func-
tions used to approximate conditional expectations such as 𝐸 (𝑌 |𝑋). More specifically,
let Π = {ℓ1 , ℓ2 , . . . , ℓ#Π } be a partition of X = 𝑠𝑢 𝑝 𝑝𝑜𝑟𝑡 (𝑋) and define
Í𝑛
𝑖=1 𝑌𝑖 1ℓ 𝑗 (𝑋𝑖 )
𝑌¯ℓ 𝑗 = Í𝑛 , 𝑗 = 1, . . . , #Π,
𝑖=1 1ℓ 𝑗 (𝑋𝑖 )
to be the average outcome for those observations 𝑖 for which 𝑋𝑖 ∈ ℓ 𝑗 . A regression

tree estimates 𝐸 (𝑌 | 𝑋 = 𝑥) using a step function
#Π
∑︁
ˆ Π) =
𝜇(𝑥; 𝑌¯ℓ 𝑗 1ℓ 𝑗 (𝑥). (3.14)
𝑗=1
The regression tree algorithm considers partitions Π that are constructed based on
recursive splits of the support of the components of 𝑋. Thus, the subsets ℓ 𝑗 , which are
called the leaves of the tree, are given by intersections of sets of the form {𝑋 𝑘 ≤ 𝑐} or
{𝑋 𝑘 > 𝑐}, where 𝑋 𝑘 denotes the 𝑘th component of 𝑋. In building the regression tree,
candidate partitions are evaluated through a mean squared error (MSE) criterion with
an added term that penalizes the number of splits to avoid overfitting. Chapter 2 of
this volume provides a more in-depth look at classification and regression trees.
From a regression tree to a causal tree. Under the unconfoundendess assumption
one way to estimate ATE consistently is to use the inverse probability weighted
estimator proposed by Hirano, Imbens and Ridder (2003):

1 ∑︁ 𝑌𝑖 𝐷 𝑖 𝑌𝑖 (1 − 𝐷 𝑖 )
𝜏ˆ = − , (3.15)
#S 𝑚(𝑋𝑖 ) 1 − 𝑚(𝑋𝑖 )
𝑖 ∈S
where S is the sample of observations on (𝑌𝑖 , 𝐷 𝑖 , 𝑋𝑖 ), #S is the sample size, and 𝑚(·)
is the propensity score function, i.e., the conditional probability 𝑚(𝑋) = 𝑃(𝐷 = 1| 𝑋).
To simplify the exposition, we will assume that the function 𝑚(𝑋) is known.10
Given a subset ℓ ⊂ X, one can also implement the estimator (3.15) in the subsample
of observations for which 𝑋𝑖 ∈ ℓ, yielding an estimate of the conditional average
treatment effect 𝜏(ℓ) = 𝐸 [𝑌 (1) −𝑌 (0)| 𝑋𝑖 ∈ ℓ]. More specifically, we define

1 ∑︁ 𝑌𝑖 𝐷 𝑖 𝑌𝑖 (1 − 𝐷 𝑖 )
𝜏ˆS (ℓ) = − ,
#ℓ 𝑚(𝑋𝑖 ) 1 − 𝑚(𝑋𝑖 )
𝑖 ∈S,𝑋𝑖 ∈ℓ
where #ℓ is the number of observations that fall in ℓ.

Computing 𝜏ˆS (ℓ) for various choices of ℓ is called subgroup analysis. Based on
subject matter theory or policy considerations, a researcher may pre-specify some
subgroups of interest. For example, theory may predict that the effect changes as
a function of age or income. Nevertheless, theory is rarely detailed enough to say
exactly how to specify the relevant age or income groups and may not be able to
say whether the two variables interact with each other or other variables. There may
be situations, especially if the dimension of 𝑋 is high, where the relevant subsets
need to be discovered completely empirically by conducting as detailed a search
as possible. However, it is now well understood that mining the data ‘by hand’ for
relevant subgroups is problematic for two reasons. First, some of these groups may
be complex and hard to discover, e.g., they may involve interactions between several
variables. Moreover, if 𝑋 is large there are simply too many possibilities to consider.
Second, it is not clear how to conduct inference for groups uncovered by data mining.
The (asymptotic) distribution of 𝜏(ℓ)
ˆ is well understood for fixed ℓ. But if the search
procedure picks a group ℓˆ because, say, 𝜏( ˆ − 𝜏ˆ is large, then the distribution of 𝜏(
ˆ ℓ) ˆ
ˆ ℓ)
will of course differ from the fixed ℓ case.
In their influential paper Athey and Imbens (2016) propose the use of the regression
tree algorithm to search for treatment effect heterogeneity, i.e., to discover the relevant
subsets ℓˆ from the data itself. The resulting partition Π̂ and the estimates 𝜏( ˆ ℓˆ ∈ Π̂
ˆ ℓ),
are called a causal tree. Their key contribution is to modify to the standard regression
tree algorithm in a way that accommodates treatment effect estimation (as opposed to
prediction) and addresses the statistical inference problem discussed above.
The first proposed modification is what Athey and Imbens (2016) call an ‘honest’
approach. This consist of partitioning the available data S = {(𝑌𝑖 , 𝐷 𝑖 , 𝑋𝑖 )}𝑖=1
𝑛 into an
estimation sample S and a training sample S . The search for hetereogeneity, i.e.,
𝑒𝑠𝑡 𝑡𝑟
the partitioning of X into relevant subsets ℓ, ˆ is conducted entirely over the training
sample S 𝑡𝑟 . Once a suitable partition of X is identified, it is taken as given and the
group-specific treatment effects are re-estimated over the independent estimation
sample that has been completely set aside up to that point. More formally, the eventual
conditional average treatment effect estimates are of the form 𝜏ˆS 𝑒𝑠𝑡 ( ℓˆS 𝑡𝑟 ), where the
notation emphasizes the use of the two samples for different purposes. Inference based
10 It is not hard to extend the following discussions to the more realistic case in which 𝑚(𝑋) needs
to be estimated. We also note that even when 𝑚(𝑋) is known, it is more efficient to work with an
estimated counterpart; see Hirano et al. (2003).
96 Lieli at al.
on 𝜏ˆS 𝑒𝑠𝑡 ( ℓˆS 𝑡𝑟 ) can then proceed as if ℓˆS 𝑡𝑟 were fixed. While in the DML literature
sample splitting is an enhancement, it is absolutely crucial in this setting.
The second modification concerns the MSE criterion used in the tree-building
algorithm. The criterion function is used to compare candidate partitions, i.e., it is
used to decide whether it is worth imposing additional splits on the data to estimate
the CATE function in more detail. The proposed changes account for the fact that (i)
instead of approximating a conditional expectation function the goal is to estimate
treatment effects; and (ii) for any partition Π̂ constructed from S 𝑡𝑟 , the corresponding
conditional average treatment effects will be re-estimated using S 𝑒𝑠𝑡 .
Technical discussion of the modified criterion. We now formally describe the
proposed criterion function. Given a partition Π = {ℓ1 , . . . , ℓ#Π } of X and a sample S,
let
#Π
∑︁
𝜏ˆS (𝑥; Π) = 𝜏ˆS (ℓ 𝑗 )1ℓ 𝑗 (𝑥)
𝑗=1
be the corresponding step function estimator of the CATE function 𝜏(𝑥), where the
value of 𝜏ˆS (𝑥; Π) is the constant 𝜏ˆS (ℓ 𝑗 ) for 𝑥 ∈ ℓ 𝑗 . For a given 𝑥 ∈ X, the MSE of the
CATE estimator is 𝐸 [(𝜏(𝑥) − 𝜏ˆS (𝑥; Π)) 2 ]; the proposed criterion function is based
on the expected (average) MSE
h 2i
EMSE(Π) = 𝐸 𝑋𝑡 , S 𝑒𝑠𝑡 𝜏(𝑋𝑡 ) − 𝜏ˆS 𝑒𝑠𝑡 (𝑋𝑡 ; Π) ,
where 𝑋𝑡 is a new, independently drawn,‘test’ observation. Thus, the goal is to choose

the partition Π in a way so that 𝜏ˆS 𝑒𝑠𝑡 (𝑥; Π) provides a good approximation to 𝜏(𝑥)
on average, where the averaging is with respect to the marginal distribution of 𝑋.
While EMSE(Π) cannot be evaluated analytically, it can still be estimated. To this
end, one can rewrite EMSE(Π) as11
n o
EMSE(Π) = 𝐸 𝑋𝑡 𝑉S 𝑒𝑠𝑡 [ 𝜏ˆS 𝑒𝑠𝑡 (𝑋𝑡 ; Π)] − 𝐸 [𝜏(𝑋𝑡 ; Π) 2 ] + 𝐸 [𝜏(𝑋𝑡 ) 2 ], (3.16)
where 𝑉S 𝑒𝑠𝑡 (·) denotes the variance operator with respect to the distribution of the
sample S 𝑒𝑠𝑡 and
#Π
∑︁
𝜏(𝑥; Π) = 𝜏(ℓ 𝑗 )1ℓ 𝑗 (𝑥).
𝑗=1
As the last term in (3.16) does not depend on Π, it does not affect the choice of
the optimal partition. We will henceforth drop this term from (3.16) and denote the
remaining two terms as EMSE—a convenient and inconsequential abuse of notation.
Recall that the key idea is to complete the tree building process (i.e., the choice of the
partition Π) on the basis of the training sample alone. Therefore, EMSE(Π) will be
11 Equation (3.16) is derived in the Electronic Online Supplement, Section 3.1. The derivations
assume that 𝐸 ( 𝜏ˆS (ℓ 𝑗 )) = 𝜏 (ℓ 𝑗 ), i.e., that the leaf-specific average treatment effect estimator is
unbiased. This is true if the propensity score function is known but only approximately true otherwise.
estimated using S 𝑡𝑟 ; the only information used from S 𝑒𝑠𝑡 is the sample size, denoted
as #S 𝑒𝑠𝑡 .
We start with the expected variance term. It is given by
n #Π
o ∑︁
𝐸 𝑋𝑡 𝑉S 𝑒𝑠𝑡 [ 𝜏ˆS 𝑒𝑠𝑡 (𝑋𝑡 ; Π)] = 𝑉S 𝑒𝑠𝑡 [ 𝜏ˆS 𝑒𝑠𝑡 (ℓ 𝑗 )]𝑃(𝑋𝑡 ∈ ℓ 𝑗 ).
𝑗=1
The variance of 𝜏ˆS 𝑒𝑠𝑡 (ℓ 𝑗 ) is of the form 𝜎 2𝑗 /#ℓ 𝑒𝑠𝑡

𝑗 , where #ℓ 𝑗
𝑒𝑠𝑡
is the number of
observations in the estimation sample falling in leaf ℓ 𝑗 and
𝜎 2𝑗 = 𝑉 𝑌𝑖 𝐷 𝑖 /𝑝(𝑋𝑖 ) −𝑌𝑖 (1 − 𝐷 𝑖 )/(1 − 𝑝(𝑋𝑖 )) 𝑋𝑖 ∈ ℓ 𝑗 .

Substituting 𝑉S 𝑒𝑠𝑡 [ 𝜏ˆS 𝑒𝑠𝑡 (ℓ 𝑗 )] = 𝜎 2𝑗 /#ℓ 𝑒𝑠𝑡

𝑗 into the expected variance equation yields
n #Π 𝜎 2
o ∑︁ 𝑗
𝐸 𝑋𝑡 𝑉S 𝑒𝑠𝑡 [ 𝜏ˆS 𝑒𝑠𝑡 (𝑋𝑡 ; Π)] = 𝑒𝑠𝑡 𝑃(𝑋𝑡 ∈ ℓ 𝑗 )
𝑗=1
#ℓ 𝑗
#Π
1 ∑︁ 2 #S 𝑒𝑠𝑡
= 𝜎 𝑃(𝑋 𝑡 ∈ ℓ 𝑗 ) .
#S 𝑒𝑠𝑡 𝑗=1 𝑗 #ℓ 𝑒𝑠𝑡
𝑗
As #S 𝑒𝑠𝑡 /#ℓ 𝑒𝑠𝑡

𝑗 ≈ 1/𝑃(𝑋𝑖 ∈ ℓ 𝑗 ) we can simply estimate the expected variance term
by
#Π
n o 1 ∑︁ 2
𝐸ˆ 𝑋𝑡 𝑉S 𝑒𝑠𝑡 [ 𝜏ˆS 𝑒𝑠𝑡 (𝑋𝑡 ; Π)] = ˆ 𝑡𝑟 ,
𝜎 (3.17)
#S 𝑒𝑠𝑡 𝑗=1 𝑗,S
ˆ 2𝑗, S 𝑡𝑟 is a suitable (approximately unbiased) estimator of 𝜎 2𝑗 over the training

where 𝜎
sample.
Turning to the second moment term in (3.16), note that for any sample S and an
independent observation 𝑋𝑡 ,
𝐸 𝑋 [𝜏(𝑋𝑡 ; Π) 2 ] = 𝐸 𝑋 𝐸 S [ 𝜏ˆS (𝑋𝑡 ; Π) 2 ] − 𝐸 𝑋 {𝑉S [ 𝜏ˆS (𝑋𝑡 ; Π)]}
because 𝐸 S [ 𝜏ˆS (𝑥; Π)] = 𝜏(𝑥; Π) for any fixed point 𝑥. Thus, an unbiased estimator
of 𝐸 𝑋 [𝜏(𝑋𝑡 ; Π) 2 ] can be constructed from the training sample S 𝑡𝑟 as
#Π
1 ∑︁ 1 ∑︁ 2
𝐸ˆ 𝑋 [𝜏(𝑋𝑡 ; Π) 2 ] = ˆ
𝜏S
2
𝑡𝑟 ,−𝑖 (𝑋𝑖 ; Π) − ˆ 𝑡𝑟 ,
𝜎 (3.18)
#S 𝑡𝑟 #S 𝑡𝑟 𝑗=1 𝑗, S
𝑖 ∈S
𝑡𝑟
where 𝜏ˆS 𝑡𝑟 ,−𝑖 is the leave-one-out version of 𝜏ˆS 𝑡𝑟 and we use the analog of (3.17) to
estimate the expected variance of 𝜏ˆS 𝑡𝑟 (𝑋𝑡 ; Π). Combining the estimators (3.17) and
(3.18) with the decomposition (3.16) gives the estimated EMSE criterion function
98 Lieli at al.
#Π
1 1 ∑︁ 2 1 ∑︁
EMSE(Π)
= + ˆ
𝜎 𝑡𝑟 − 𝜏ˆS 𝑡𝑟 ,−𝑖 (𝑋𝑖 ; Π) 2 . (3.19)
#S 𝑒𝑠𝑡 #S 𝑡𝑟 𝑗=1 𝑗,S #S 𝑡𝑟 𝑡𝑟 𝑖 ∈S
The criterion function (3.19) has almost exactly the same form as in Athey and Imbens
(2016), except that they do not use a leave-one-out estimator in the second term.
The intuition about how the criterion (3.19) works is straightforward. Say that Π
and Π ′ are two partitions where Π ′ is finer in the sense that there is an additional split
along a given 𝑋 coordinate. If the two estimated treatment effects are not equal across
this extra split, then the second term in the difference will decrease in absolute value,
making EMSE(Π ′) ceteris paribus lower. However, the first term of the criterion also
takes into account the fact that an additional split will result in leaves with fewer
observations, increasing the variance of the ultimate CATE estimate. In other words,
the first term will generally increase with an additional split and it is the net effect
that determines whether Π or Π ′ is deemed as a better fit to the data.
The criterion function (3.19) is generalizable to other estimators. In fact, the
precise structure of 𝜏ˆ𝑆 (ℓ) played almost no role in deriving (3.19); the only properties
we made use of was unbiasedness (𝐸 𝑆 ( 𝜏ˆ𝑆 (ℓ 𝑗 )) = 𝜏(ℓ 𝑗 )) and that the variance of
𝜏ˆ𝑆 (ℓ 𝑗 ) is of the form 𝜎 2𝑗 /#ℓ 𝑗 . Hence, other types of estimators could be implemented
in each leaf; for example, Reguly (2021) uses this insight to extend the framework to
the parametric sharp regression discontinuity design.
3.4.3 Extensions and Technical Variations on the Causal Tree Approach
Wager and Athey (2018) extend the causal tree approach of Athey and Imbens (2016)
to causal forests, which are composed of causal trees with a conditional average
treatment effect estimate in each leaf. To build a causal forest estimator, one first
generates a large number of random subsamples and grows a causal tree on each
subsample using a given procedure. The causal forest estimate of the CATE function is
then obtained by averaging the estimates produced by the individual trees. Practically,
a forest approach can reduce variance and smooth the estimate.
Wager and Athey (2018) propose two honest causal forest algorithms. One is
based on double-sample trees and the other is based on propensity score trees. The
construction of a double-sample tree involves a procedure similar to the one described
in Section 4.2. First one draws without replacement 𝐵 subsamples of size 𝑠 = 𝑂 (𝑛𝜌 )
for some 0 < 𝜌 < 1 from the original data. Let these artificial samples be denoted as S𝑏 ,
𝑏 = 1, . . . , 𝐵. Then one splits each S𝑏 into two parts with size #S𝑏𝑡𝑟 = #S𝑏𝑒𝑠𝑡 = 𝑠/2.12
The tree is grown using the S𝑏𝑡𝑟 data and the leafwise treatment effects are estimated
using the S𝑏𝑒𝑠𝑡 data, generating an individual estimator 𝜏ˆ𝑏 (𝑥). The splits of the tree
are chosen by minimizing an expected MSE criterion analogous to (3.19), where
S𝑏𝑡𝑟 takes the role of S 𝑡𝑟 and S𝑏𝑒𝑠𝑡 takes the role of S 𝑒𝑠𝑡 . For any fixed value 𝑥, the
causal forest CATE estimator 𝜏(𝑥) ˆ is obtained by averaging the individual estimates
12 We assume 𝑠 is an even number.

= 𝐵−1 𝑏=1
Í𝐵
𝜏ˆ𝑏 (𝑥), i.e., 𝜏(𝑥)
ˆ 𝜏ˆ𝑏 (𝑥). The propensity score tree procedure is similar
except that one grows the tree using the whole subsample S𝑏 and uses the treatment
assignment as the outcome variable. Wager and Athey (2018) show that the random
forest estimates are asymptotically normally distributed and the asymptotic variance
can be consistently estimated by infinitesimal jackknife so valid statistical inference
is available.
Athey et al. (2019) propose a generalized random forest method that transforms
the traditional random forest into a flexible procedure for estimating any unknown
parameter identified via local moment conditions. The main idea is to use forest-based
algorithms to learn the problem specific weights so as to be able to solve for the
parameter of interest via a weighted local M-estimator. For each tree, the splitting
decisions are based on a gradient tree algorithm; see Athey et al. (2019) for further
details.
There are other methods proposed in the literature for the estimation of treatment
effect heterogeneity such as the neural network based approaches by Yao et al. (2018)
and Alaa, Weisz and van der Schaar (2017). These papers do not provide asymptotic
theory and valid statistical inference is not available at the moment, so this is an
interesting direction for future research.
3.4.4 The Dimension Reduction Approach
As discussed in Section 3.4.1, the target of this approach is the reduced dimensional
CATE function 𝜏(𝑋1 ) = 𝐸 [𝑌 (1) −𝑌 (0)| 𝑋1 ], where 𝑋1 is a component of 𝑋.13 This
variable is a priori chosen (rather than discovered from the data) and the goal is to
estimate 𝜏(𝑋1 ) in a flexible way. This can be accomplished by an extension of the
DML framework.
Let 𝜓(𝑊, 𝑚 0 , 𝜉0 ) be defined as in Section 3.4.1. Then it is not hard to show the
reduced dimensional CATE function is identified by the orthogonal moment condition
𝐸 [𝜓(𝑊, 𝑚 0 , 𝜉0 ) − 𝜏(𝑋1 )| 𝑋1 ] = 0 ⇐⇒ 𝜏(𝑋1 ) = 𝐸 [𝜓(𝑊, 𝑚 0 , 𝜉0 )|𝑋1 ].
While ATE is given by the unconditional expectation of the 𝜓 function, the reduced
dimensional CATE function 𝜏(𝑋1 ) is the conditional expectation 𝐸 (𝜓|𝑋1 ). As 𝑋1 is
a scalar, it is perfectly feasible to estimate this regression function by a traditional
kernel-based nonparametric method such as a local constant (Nadaraya-Watson) or a
local linear regression estimator.
To be more concrete, let us consider the same illustrative setup as in Section 3.3.3.
Splitting the sample into, say, two parts 𝐼1 and 𝐼2 , let 𝑚ˆ 0,𝐼𝑘 and 𝜉ˆ0,𝐼𝑘 denote the first
stage ML estimators over the subsample 𝐼 𝑘 . Define 𝜓ˆ 𝑖 as 𝜓(𝑊𝑖 , 𝑚ˆ 0,𝐼2 , 𝜉ˆ0,𝐼2 ) for 𝑖 ∈ 𝐼1
and as 𝜓(𝑊𝑖 , 𝑚ˆ 0,𝐼1 , 𝜉ˆ0,𝐼1 ) for 𝑖 ∈ 𝐼2 . Then, for a given 𝑥 1 ∈ 𝑠𝑢 𝑝 𝑝𝑜𝑟𝑡 (𝑋1 ), the DML
estimator of 𝜏(𝑥1 ) with a Nadaraya-Watson second stage is given by
13 The theory allows for 𝑋1 to be a small vector, but 𝜏 (𝑋1 ) is easiest to visualize, which is perhaps
its main attraction, when 𝑋1 is a scalar. So this is the case we will focus on here.
100 Lieli at al.

Í𝑛
ˆ 𝑋1𝑖 −𝑥1
𝑖=1 𝜓𝑖 𝐾 ℎ
𝐷𝑀𝐿
𝜏ˆ (𝑥 1 ) = Í ,
𝑛 𝑋1𝑖 −𝑥1
𝑖=1 𝐾 ℎ
where ℎ = ℎ 𝑛 is a bandwidth sequence (satisfying ℎ → 0 and 𝑛ℎ → ∞) and 𝐾 (·) is

a kernel function. Thus, instead of simply taking the average of the 𝜓ˆ 𝑖 values, one
performs a nonparametric regression of 𝜓ˆ 𝑖 on 𝑋1𝑖 . An estimator of this form was
already proposed by Abrevaya et al. (2015), except their identification was based
on a non-orthogonal inverse probability weighted moment condition, and their first
stage estimator of the propensity score was a traditional parametric or nonparametric
estimator.
The asymptotic theory of 𝜏ˆ 𝐷 𝑀 𝐿 (𝑥1 ) is developed by Fan et al. (2020).14 The
central econometric issue is again finding the appropriate conditions under which
the first stage model selection and estimation leaves the asymptotic distribution
of the second stage estimator unchanged in the sense that 𝜏ˆ 𝐷 𝑀 𝐿 (𝑥1 ) is first order
asymptotically equivalent to the infeasible estimator in which 𝜓ˆ 𝑖 is replaced by 𝜓𝑖 . In
this case
√
𝑛ℎ[ 𝜏ˆ 𝐷 𝑀 𝐿 (𝑥1 ) − 𝜏(𝑥 1 )] →𝑑 𝑁 (0, 𝜎 2 (𝑥 1 )), for all 𝑥1 ∈ 𝑠𝑢 𝑝 𝑝𝑜𝑟𝑡 (𝑋1 ), (3.20)
provided that the undersmoothing condition 𝑛ℎ5 → 0 holds to eliminate asymptotic

bias. However, one is often interested in the properties of the entire function 𝑥1 ↦→ 𝜏(𝑥1 )
rather than just its value evaluated at a fixed point. 𝐼𝑏𝑖𝑑. state a uniform representation
result for 𝜏ˆ 𝐷 𝑀 𝐿 (𝑥1 ) that implies (3.20) and but also permits the construction of
uniform confidence bands that contain the whole function 𝑥1 ↦→ 𝜏(𝑥 1 ) with a pre-
specified probability. They propose the use of a multiplier bootstrap procedure for
this purpose, which requires only a single estimate 𝜏ˆ 𝐷 𝑀 𝐿 (𝑥 1 ).
The high level conditions used by Fan et al. (2020) to derive these results account for
the interaction between the required convergence rate of the first stage ML estimators
and the second stage bandwidth ℎ. (A simplified general sufficient condition is that
the ML estimators converge faster than ℎ1/4 𝑛−1/4 .) They also show that the high
level conditions are satisfied in case of a Lasso first stage and strengthened sparsity
assumptions on the nuisance functions relative to ATE estimation; see 𝑖𝑏𝑖𝑑. for
details.15
The paper by Semenova and Chernozhukov (2020) extends the general DML
framework of Section 3.3.3 to cases in which the parameter of interest is a function
identified by a low-dimensional conditional moment condition. This includes the
reduced dimensional CATE function discussed thus far but other estimands as well.
They use series estimation in the second stage and provide asymptotic results that
14 To be exact, ibid. use local linear regression in the second stage; the local constant version of the
estimator is considered in an earlier working paper. The choice makes no difference in the main
elements of the theory and the results.
15 Using the notation of footnote 8, a sufficient sparsity assumption for the cross-fitted case considered
here is that 𝑠𝑚 · 𝑠 𝜉 grows slower than 𝑛ℎ, where 𝑠 𝜉 is the larger sparsity index among the two
components of 𝜉 .
can be used for inference under general high-level conditions on the first stage ML
estimators.
3.5 Empirical Illustration
We revisit the application in Abrevaya et al. (2015) and Fan et al. (2020) and study
the effect of maternal smoking during pregnancy on the baby’s birthweight. The data
set comes from vital statistics records in North Carolina between 1988 and 2002 and
is large both in terms of the number of observations and the available covariates.
We restrict the sample to first time mothers and, following the literature, analyze
the black and caucasian (white) subsamples separately.16 The two sample sizes are
157,989 and 433,558, respectively.
The data set contains a rich set of individual-level covariates describing the
mother’s socioeconomic status and medical history, including the progress of the
pregnancy. This is supplemented by zip-code level information corresponding to
the mother’s residence (e.g., per capita income, population density, etc.). Table 3.1
summarizes the definition of the dependent variable (𝑌 ), the treatment variable (𝐷)
and the control variables used in the analysis. We divide the latter variables into
a primary set 𝑋1 , which includes the more important covariates and some of their
powers and interactions, and a secondary set 𝑋2 , which includes auxiliary controls
and their transformations.17 Altogether, 𝑋1 includes 26 variables while the union of
𝑋1 and 𝑋2 has 679 components for the black subsample and 743 for the caucasian.
16 The reason for focusing on first time mothers is that in case of previous deliveries we cannot
identify the data point that corresponds to it. Birth outcomes across the same mother at different
points in time are more likely to be affected by unobserved heterogeneity (see Abrevaya et al., 2015
for a more detailed discussion).
17 As 𝑋1 and 𝑋2 already include transformations of the raw covariates, these vectors correspond to
the dictionary 𝑏 (𝑋) in the notation of Section 3.3.
102 Lieli at al.
Table 3.1: The variables used in the empirical exercise
𝑌 bweight birth weight of the baby (in grams)

𝐷 smoke if mother smoked during pregnancy (1 if yes)
mage mother’s age (in years)
meduc mother’s education (in years)
prenatal month of first prenatal visit
prenatal_visits number of prenatal visits
male baby’s gender (1 if male)
married if mother is married (1 if yes)
drink if mother used alcohol during pregnancy (1 if yes)
diabetes if mother has diabetes (1 if yes)
hyperpr if mother has high blood pressure (1 if yes)
amnio if amniocentesis test (1 if yes)
𝑋1
ultra if ultrasound during pregnancy (1 if yes)
dterms previous terminated pregnancies (1 if yes)
fagemiss if father’s age is missing (1 if yes)
polynomials: mage2
meduc, prenatal, prenatal_visits, male,
interactions: mage × married, drink, diabetes, hyperpr, amnio,
ultra, dterms, fagemiss
mom_zip zip code of mother’s residence (as a series of dummies)
byear birth year 1988-2002 (as series of dummies)
anemia if mother had anemia (1 if yes)
med_inc median income in mother’s zip code
pc_inc per capita income in mother’s zip code
popdens population density in mother’s zip code
𝑋2 fage father’s age (in years)
feduc father’s education (in years)
feducmisss if father’s education is missing (1 if yes)
polynomials: fage2 , fage3 , mage3
We present two exercises. First, we use the DML approach to estimate 𝜏, i.e.,
the average effect (ATE) of smoking. This is a well-studied problem with several
estimates available in the literature using various data sets and methods (see e.g.,
Abrevaya, 2006, Da Veiga & Wilder, 2008 and Walker, Tekin & Wallace, 2009). The
point estimates range from about −120 to −250 grams with the magnitude of the
effect being smaller for blacks than whites. Second, we use the causal tree approach
to search for covariates that drive heterogeneity in the treatment effect and explore
the full-dimensional CATE function 𝜏(𝑋). To our knowledge, this is a new exercise;
both Abrevaya et al. (2015) and Fan et al. (2020) focus on the reduced dimensional
CATE function with mother’s age as the pre-specified variable of interest.
The findings from the first exercise for black mothers are presented in Table 3.2;
the corresponding results for white mothers can be found in the Electronic Online
Supplement, Table 3.1. The left panel shows a simple setup where we use 𝑋1 as the
Table 3.2: Estimates of 𝜏 for black mothers
Basic setup Extended setup

Point-estimate SE Point-estimate SE
OLS -132.3635 6.3348 -130.2478 6.3603
Naive Lasso (𝜆∗ ) -131.6438 - -132.3003 -
Naive Lasso (0.5𝜆∗ ) -131.6225 - -130.7610 -
Naive Lasso (2𝜆∗ ) -131.4569 - -134.6669 -
Post-naive-Lasso (𝜆∗ ) -132.3635 6.3348 -128.6592 6.2784
Post-naive-Lasso (0.5𝜆∗ ) -132.3635 6.3348 -130.7227 6.2686
Post-naive-Lasso (2𝜆∗ ) -132.3635 6.3348 -130.2032 6.3288
DML (𝜆∗ ) -132.0897 6.3345 -129.9787 6.3439
DML (0.5𝜆∗ ) -132.1474 6.3352 -128.8866 6.3413
DML (2𝜆∗ ) -132.1361 6.3344 -131.5680 6.3436
DML-package -132.0311 6.3348 -131.0801 6.3421
Notes: Sample size= 157, 989. 𝜆∗ denotes Lasso penalties obtained by 5-fold cross validation. The
DML estimators are implemented by 2-fold cross-fitting. The row titled ‘DML-package’ contains the
estimate obtained by using the ‘official’ DML code (dml2) available at https://docs.doubleml.org/r/
stable/. All other estimators are programmed directly by the authors.
vector of controls, whereas the right panel works with the extended set 𝑋1 ∪ 𝑋2 . In
addition to the DML estimates, we report several benchmarks: (i) the OLS estimate
of ATE from a regression of 𝑌 on 𝐷 and the set of controls; (ii) the corresponding
direct/naive Lasso estimates of ATE (with the treatment dummy excluded from the
penalty); and (iii) post-Lasso estimates where the model selected by the Lasso is
re-estimated by OLS. The symbol 𝜆∗ denotes Lasso penalty levels chosen by 5-fold
cross validation; we also report the results for the levels 𝜆∗ /2 and 2𝜆∗ . The DML
estimators are implemented using 2-fold cross-fitting.
The two main takeaways from Table 3.2 are that the estimated ‘smoking effect’
is, on average, −130 grams for first time black mothers (which is consistent with
previous estimates in the literature), and that this estimate is remarkably stable across
the various methods. This includes the OLS benchmark with the basic and extended
set of controls as well as the corresponding naive Lasso and post-Lasso estimators.
We find that even at the penalty level 2𝜆∗ the naive Lasso keeps most of the covariates
and so do the first stage Lasso estimates from the DML procedure. This is due to
the fact that even small coefficients are precisely estimated in such large samples,
so one does not achieve a meaningful reduction in variance by setting them to zero.
As the addition of virtually any covariate improves the cross-validated MSE by a
small amount, the optimal penalty 𝜆∗ is small. With close to the full set of covariates
utilized in either setup, and a limited amount of shrinkage due to 𝜆∗ being small, there
104 Lieli at al.
is little difference across the methods.18 In addition, using 𝑋1 ∪ 𝑋2 versus 𝑋1 alone

affects the size of the point estimates only to a small degree—the post-Lasso and
DML estimates based on the extended setup are 1 to 3 grams smaller in magnitude.
Physically this is a very small difference, though still a considerable fraction of the
standard error.
It is also noteworthy that the standard errors provided by the post-Lasso estimator
are very similar to the DML standard errors. Hence, in the present situation naively
conducting inference using the post-Lasso estimator would lead to the proper
conclusions. Of course, relying on this estimator for inference is still bad practice as
there are no a priori theoretical guarantees for it to be unbiased or asymptotically
normal.
The results for white mothers (see the Electronic Online Supplement, Table 3.1.)
follow precisely the same patterns as discussed above—the various methods are in
agreement and the basic and extended model setups deliver the same results. The
only difference is that the point estimate of the average smoking effect is about −208
grams.
The output from the second exercise is displayed in Figure 3.1, which shows the
estimated CATE function for black mothers represented as a tree (the corresponding
figure for white mothers is in the Electronic Online Supplement, Figure 3.1). The
two most important leaves are in the bottom right corner, containing about 92% of
the observations in total. These leaves originate from a parent node that splits the
sample by mother’s age falling below or above 19 years; the average smoking effect
is then significantly larger in absolute value for the older group (−221.7 grams) than
for the younger group (−120.1 grams).
This result is fully consistent with Abrevaya et al. (2015) and Fan et al. (2020),
who estimate the reduced dimensional CATE function for mother’s age and find that
the smoking effect becomes more detrimental for older mothers. The conceptual
difference between these two papers and the tree in Figure 3.1 is twofold. First, in
building the tree, age is not designated a priori as a variable of interest; it is the tree
building algorithm that recognized its relevance for treatment effect heterogeneity.
Second, not all other control variables are averaged out in the two leaves highlighted
above; in fact, the preceding two splits condition on normal blood pressure as well as
the mother being older than 15. However, given the small size of the complementing
leaves, we would caution against over-interpreting the rest of the tree. Whether or
not blood pressure is related to the smoking effect in a stable way would require
some further robustness checks. We nevertheless note that even within the high blood
pressure group the age-related pattern is qualitatively similar—smoking has a larger
negative effect for older mothers and is in fact statistically insignificant below age 22.
The results for white mothers (see the Electronic Online Supplement, Figure 3.1.)
are qualitatively similar in that they confirm that the smoking effect becomes more
18 Another factor that contributes to this result is that no individual covariate is strongly correlated
with the treatment dummy. Hence, even if one were mistakenly dropped from the naive Lasso
regression, it would not create substantial bias. Indeed, in exercises run on small subsamples (so that
the 𝑑𝑖𝑚(𝑋)/𝑛 ratio is much larger) we still find little difference between the naive (post) Lasso and
the DML method in this application.
hyperpr = 1
yes no
mage < 22 mage < 15

yes no yes no
85.2 [60.1], (3%) −163.0 [54.5], (3%) 6.4 [101.1], (2%) mage ≥ 19
yes no
−221.7 [9.94] (64%) −120.1 [19.2], (27%)
Fig. 3.1: A causal tree for the effect of smoking on birth weight (first time black
mothers).
Notes: standard errors are in brackets and the percentages in parenthesis denote share
of group in the sample. The total number of observations is 𝑁 = 157, 989 and the
covariates used are 𝑋1 , except for the polynomial and interaction terms. To obtain a
simpler model, we choose the pruning parameter using the 1SE-rule rather than the
minimum cross-validated MSE.
negative with age (compare the two largest leaves on the 𝑚𝑎𝑔𝑒 ≥ 28 and 𝑚𝑎𝑔𝑒 < 28
branches). Hypertension appears again as a potentially relevant variable, but the share
of young women affected by this problem is small. We also note that the results are
obtained over a subsample of 𝑁 = 150, 000 and robustness checks show sensitivity to
the choice of the subsample. Nonetheless, the age pattern is qualitatively stable.
The preceding results illustrate both the strengths and weaknesses of using a
causal tree for heterogeneity analysis. On the one hand, letting the data speak for
itself is philosophically attractive and can certainly be useful. On the other hand, the
estimation results may appear to be too complex or arbitrary and can be challenging
to interpret. This problem is exacerbated by the fact that trees grown on different
subsamples can show different patterns — suggesting that in practice it is prudent to
construct a causal forest, i.e., use the average of multiple trees.
3.6 Conclusion
Until about the early 2010s econometrics and machine learning developed on
parallel paths with only limited interaction between the two fields. This has changed
considerably over the last decade and ML methods are now widely used in economic
applications. In this chapter we reviewed their use in treatment effect estimation,
focusing on two strands of the literature.
In applications of the double or debiased machine learning (DML) approach, the
parameter of interest is typically some average treatment effect, and ML methods
are employed in the first stage to estimate the unknown nuisance functions necessary
for identification (such as the propensity score). A key insight that emerges from
106 Lieli at al.
this literature is that machine learning can be fruitfully applied for this purpose if
the nuisance functions enter the second stage estimating equations (more precisely,
moment conditions) in a way that satisfies an orthogonality condition. This condition
ensures that the parameter of interest is consistently estimable despite the selection
and approximation errors introduced by the first stage machine learning procedure.
Inference in the second stage can then proceed as usual. In practice a cross-fitting
procedure is recommended, which involves splitting the sample between the first and
second stages.
In applications of the causal tree or forest methodology, the parameter of interest
is the full dimensional conditional average treatment effect (CATE) function, i.e.,
the focus is on treatment effect heterogeneity. The method permits near automatic
and data-driven discovery of this function, and if the selected approximation is
re-estimated on an independent subsample (the ‘honest approach’) then inference
about group-specific effects can proceed as usual. Another strand of the heterogeneity
literature estimates the projection of the CATE function on a given, pre-specified
coordinate to facilitate presentation and interpretation. This can be accomplished by
an extension of the DML framework where in the first stage the nuisance functions
are estimated by an ML method and in the second stage a traditional nonparametric
estimator is used (e.g., kernel-based or series regression).
In our empirical application (the effects of maternal smoking during pregnancy on
the baby’s birthweight) we illustrate the use of the DML estimator as well as causal
trees. While the results confirm previous findings in the literature, they also highlight
some limitations of these methods. In particular, with the number of observations
orders of magnitude larger than the number of covariates, and the covariates not
being very strong predictors of the treatment, DML virtually coincides with OLS
and even the naive (direct) Lasso estimator. The causal tree approach successfully
uncovers important patterns in treatment effect heterogeneity but also some that seem
somewhat incidental and/or less straightforward to interpret. This suggests that in
practice a causal forest should be used unless the computational cost is prohibitive.
In sum, machine learning methods, while geared toward prediction tasks in
themselves, can be used to enhance treatment effect estimation in various ways. This
is an active research area in econometrics at the moment, with a promise to supply
exciting theoretical developments and a large number of empirical applications for
years to come.
Acknowledgements We thank Qingliang Fan for his help in collecting literature. We are also
grateful to Alice Kuegler, Henrika Langen and the editors for their constructive comments, which
led to noticeable improvements in the exposition. The usual disclaimer applies.
References
Abrevaya, J. (2006). Estimating the effect of smoking on birth outcomes using

a matched panel data approach. Journal of Applied Econometrics, 21(4),
References 107
489–519.
Abrevaya, J., Hsu, Y.-C. & Lieli, R. P. (2015). Estimating conditional average
treatment effects. Journal of Business & Economic Statistics, 33(4), 485-505.
Alaa, A. M., Weisz, M. & van der Schaar, M. (2017). Deep counterfactual networks
with propensity-dropout. arXiv preprint arXiv:1706.05966, https:/ / arxiv.org/
abs/ 1706.05966.
Athey, S. & Imbens, G. (2016). Recursive partitioning for heterogeneous causal
Athey, S. & Imbens, G. W. (2019). Machine learning methods that economists should
know about. Annual Review of Economics, 11(1), 685-725.
Athey, S., Tibshirani, J. & Wager, S. (2019). Generalized random forests. Annals of
Statistics, 47(2), 1148–1178.
Athey, S. & Wager, S. (2021). Policy learning with observational data. Econometrica,
89(1), 133-161.
Belloni, A., Chen, D., Chernozhukov, V. & Hansen, C. (2012). Sparse models
and methods for optimal instruments with an application to eminent domain.
Econometrica, 80(6), 2369–2429.
Belloni, A., Chernozhukov, V., Fernández-Val, I. & Hansen, C. (2017). Program
evaluation and causal inference with high-dimensional data. Econometrica,
85(1), 233–298.
Belloni, A., Chernozhukov, V. & Hansen, C. (2013). Inference for high-dimensional
sparse econometric models. In M. A. D. Acemoglu & E. Dekel (Eds.), Advances
in economics and econometrics. 10th world congress, vol. 3. (pp. 245–95).
Cambridge University Press.
Belloni, A., Chernozhukov, V. & Hansen, C. (2014a). High-dimensional methods and
inference on structural and treatment effects. Journal of Economic Perspectives,
28(2), 29-50.
Belloni, A., Chernozhukov, V. & Hansen, C. (2014b). Inference on treatment effects
after selection among high-dimensional controls. The Review of Economic
Studies, 81(2), 608–650.
Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C. & Newey, W.
(2017, May). Double/debiased/neyman machine learning of treatment effects.
American Economic Review, 107(5), 261-65.
Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W.
& Robins, J. (2018). Double/debiased machine learning for treatment and
structural parameters. The Econometrics Journal, 21(1), C1-C68.
Colangelo, K. & Lee, Y.-Y. (2022). Double debiased machine learning nonparametric
inference with continuous treatments. arXiv preprint arXiv:2004.03036, https:/ /
arxiv.org/ abs/ 2004.03036.
Da Veiga, P. V. & Wilder, R. P. (2008). Maternal smoking during pregnancy and
birthweight: a propensity score matching approach. Maternal and Child Health
Journal, 12(2), 194–203.
Donald, S. G., Hsu, Y.-C. & Lieli, R. P. (2014). Testing the unconfoundedness
assumption via inverse probability weighted estimators of (l)att. Journal of
Business & Economic Statistics, 32(3), 395-415.
108 Lieli at al.
Electronic Online Supplement. (2022). Online Supplement of the book Econometrics

with Machine Learning. https://sn.pub/0ObVSo.
Fan, Q., Hsu, Y.-C., Lieli, R. P. & Zhang, Y. (2020). Estimation of conditional
average treatment effects with high-dimensional data. Journal of Business and
Economic Statistics. (forthcoming)
Farbmacher, H., Huber, M., Laffers, L., Langen, H. & Spindler, M. (2022). Causal
mediation analysis with double machine learning. The Econometrics Journal.
(forthcoming)
Hirano, K., Imbens, G. W. & Ridder, G. (2003). Efficient estimation of average
treatment effects using the estimated propensity score. Econometrica, 71(4),
1161–1189.
Hsu, Y.-C., Huber, M., Lee, Y.-Y. & Liu, C.-A. (2022). Testing monotonicity of
mean potential outcomes in a continuous treatment with high-dimensional data.
arXiv preprint arXiv:2106.04237 ,https:/ / arxiv.org/ abs/ 2106.04237.
Huber, M. (2021). Causal analysis (Working Paper). Fribourg, Switzerland:
University of Fribourg.
Imbens, G. W. & Wooldridge, J. M. (2009). Recent developments in the econometrics
of program evaluation. Journal of economic literature, 47(1), 5–86.
Kitagawa, T. & Tetenov, A. (2018). Who should be treated? empirical welfare
maximization methods for treatment choice. Econometrica, 86(2), 591-616.
Knaus, M. (2021). Double machine learning based program evaluation under
unconfoundedness. arXiv preprint arXiv:2003.03191, https:/ / arxiv.org/ abs/
2003.03191.
Knaus, M., Lechner, M. & Strittmatter, A. (2021). Machine learning estimation of het-
erogeneous causal effects: Empirical monte carlo evidence. The Econometrics
Journal, 24(1), 134-161.
Kreif, N. & DiazOrdaz, K. (2019). Machine learning in policy evaluation: New tools
for causal inference. arXiv preprint arXiv:1903.00402, https:/ / arxiv.org/ abs/
1903.00402.
Kueck, J., Luo, Y., Spindler, M. & Wang, Z. (2022). Estimation and inference of
treatment effects with l2-boosting in high-dimensional settings. Journal of
Econometrics. (forthcoming)
Mullainathan, S. & Spiess, J. (2017). Machine learning: An applied econometric
approach. Journal of Economic Perspectives, 31(2), 87-106.
Neyman, J. (1959). Optimal asymptotic tests of composite statistical hypotheses. In
U. Grenander (Ed.), Probability and statistics (p. 416-44).
Pagan, A. & Ullah, A. (1999). Nonparametric econometrics. Cambridge University
Press.
Reguly, A. (2021). Heterogeneous treatment effects in regression discontinuity
designs. arXiv preprint arXiv:2106.11640, https:/ / arxiv.org/ abs/ 2106.11640.
Semenova, V. & Chernozhukov, V. (2020). Debiased machine learning of conditional
average treatment effects and other causal functions. The Econometrics Journal,
24(2), 264-289.
Semenova, V. & Chernozhukov, V. (2021). Debiased machine learning of conditional
average treatment effects and other causal functions. The Econometrics Journal,
References 109
24(2), 264–289.
Wager, S. & Athey, S. (2018). Estimation and inference of heterogeneous treatment
effects using random forests. Journal of the American Statistical Association,
113(523), 1228–1242.
Walker, M. B., Tekin, E. & Wallace, S. (2009). Teen smoking and birth outcomes.
Southern Economic Journal, 75(3), 892–907.
Yao, L., Li, S., Li, Y., Huai, M., Gao, J. & Zhang, A. (2018). Representation learning
for treatment effect estimation from observational data. Proceedings of the
32nd International Conference on Neural Information Processing Systems,
2638-2648.
Zimmert, M. & Lechner, M. (2019). Nonparametric estimation of causal heterogeneity
under high-dimensional confounding. arXiv preprint arXiv:1908.08779, https:/ /
arxiv.org/ abs/ 1908.08779.
Chapter 4
Forecasting with Machine Learning Methods
Marcelo C. Medeiros
Abstract This chapter surveys the use of supervised Machine Learning (ML) models
to forecast time-series data. Our focus is on covariance stationary dependent data
when a large set of predictors is available and the target variable is a scalar. We
start by defining the forecasting scheme setup as well as different approaches to
compare forecasts generated by different models/methods. More specifically, we
review three important techniques to compare forecasts: the Diebold-Mariano (DM)
and the Li-Liao-Quaedvlieg tests, and the Model Confidence Set (MCS) approach.
Second, we discuss several linear and nonlinear commonly used ML models. Among
linear models, we focus on factor (principal component)-based regressions, ensemble
methods (bagging and complete subset regression), and the combination of factor
models and penalized regression. With respect to nonlinear models, we pay special
attention to neural networks and autoenconders. Third, we discuss some hybrid
models where linear and nonlinear alternatives are combined.
4.1 Introduction
This chapter surveys the recent developments in the Machine Learning (ML) literature
to forecast time-series data. ML methods have become an important estimation,
model selection and forecasting tool for applied researchers in different areas, ranging
from epidemiology to marketing, economics and finance. With the availability of
vast datasets in the era of Big Data, producing reliable and robust forecasts is of great
importance.
ML has gained a lot of popularity during the last few years and there are
several definitions in the literature of what exactly the term means. Some of the
most popular definitions yield the misconception that ML is a sort of magical
framework where computers learn patterns in the data without being explicitly
Marcelo C. Medeiros B
Pontifical Catholic University of Rio de Janeiro, Rio de Janeiro, Brazil, e-mail: mcm@econ.puc-rio.br
112 Medeiros
programmed; see, for example, Samuel (1959). With respect to the framework
considered in this chapter, we define ML, or more specifically, Supervised ML,
as set of powerful statistical models/methods combined with automated computer
algorithms to learn hidden patterns between a target variable and a potentially very
large set of explanatory variables. More specifically, in the forecasting framework,
we want to learn the expectation of a random variable 𝑌 , which take values on R,
conditional on observations of a 𝑝-dimensional set of random variables 𝑿. Therefore,
our goal is to learn the function 𝑓 (𝒙) := E(𝑌 | 𝑿 = 𝒙), based on a sample of 𝑇
observations of {𝑌𝑡 , 𝑿 𝑡′ }. We have a particular interest in the Big Data scenario where
the number of predictors (𝑝) is much larger than the sample size (𝑇). We target
the conditional expectation due to the well-known result that it gives the optimal
prediction, in the mean-squared error sense, of 𝑌 based on observations of 𝑿.
The supervised ML methods presented here can be roughly divided in three groups.
The first one includes linear models, where the conditional expectation E(𝑌 | 𝑿 = 𝒙)
is assumed to be a linear function of the data. We consider three subclasses of linear
alternatives. We start by surveying models based on factors, where the large set
predictors 𝑿 is represented by a small number of factors 𝑭 taking values on R 𝑘 , where
𝑘 is much smaller than 𝑝; see for example, Stock and Watson (2002b, 2002a). We
continue by discussing the combination of factor models and penalized regression. For
a nice overview of penalized regression models see the Chapter 1. Finally, we present
methods based on ensemble of forecasts, where a potentially large number of very
simple models is estimated and the final forecast is a combination of the predictions
from individual models. Examples of such techniques are Bagging (Breiman, 1996)
and the Complete Subset Regression (Elliott, Gargano & Timmermann, 2013, 2015).
The second group of models consists of nonlinear approximations to E(𝑌 | 𝑿 = 𝒙).
In general, the models considered here assume that the unrestricted nonlinear relation
between 𝑌 and 𝑿 can be well approximated by a combination of simpler nonlinear
(basis) functions. We start by presenting an unified framework based on sieve
semiparametric approximation as in Grenander (1981). We continue by analysing
specific models as special cases of our general setup. More specifically, we cover
feedforward neural networks, both in their shallow and deep versions, and convolution
and recurrent neural networks. Neural Networks (NN) are probably one of the most
popular ML methods. The success is partly due to the, in our opinion, misguided
analogy to the functioning of the human brain. Contrary of what has been boasted in
the early literature, the empirical success of NN models comes from a mathematical
fact that a linear combination of sufficiently many simple basis functions is able to
approximate very complicated functions arbitrarily well in some specific choice of
metric.
Finally, we discuss several alternatives to combine linear and nonlinear ML
models. For example, we can write the conditional expectation as 𝑓 (𝑥) := E(𝑌 |𝑋 =
𝑥) = 𝒙 ′ 𝜷 + 𝑔(𝒙), where 𝑔(·) is a nonlinear function and 𝜷 is a parameter to be
estimated.
Before reviewing the ML methods discussed above, we start by defining the
estimation and forecasting framework that is considered throughout this chapter.
More specifically, we discuss the use of rolling versus expanding windows and direct
4 Forecasting with Machine Learning Methods 113
versus indirect construction of multi-step-ahead forecasts. In the sequence, we review

approaches to compare forecasts from different alternative models. Such tests are in
the class of equal or superior predictive ability.
4.1.1 Notation
A quick word on notation: an uppercase letter as in 𝑋 denotes a random quantity as

opposed to a lowercase letter 𝑥 which denotes a deterministic (non-random) quantity.
Bold letters as in 𝑿 and 𝒙 are reserved for multivariate objects such as vector and
matrices. The symbol ∥ · ∥ 𝑞 for 𝑞 ≥ 1 denotes the ℓ𝑞 norm of a vector. For a set 𝑆 we
use |𝑆| to denote its cardinality.
4.1.2 Organization
In addition to the Introduction, this Chapter is organized as follows. In Section 4.2 we

present the forecasting setup considered and we discuss the benefits of rolling window
versus expanding window approach. We also analyze the pros and cons of direct
forecasts as compared to indirect forecasts. Section 4.3 discuss different methods
to compare forecasts from different models/methods. In Section 4.4 we present the
linear ML models, while in Section 4.5 we consider the nonlinear alternatives. Hybrid
approaches are reviewed in Section 4.5.5.
4.2 Modeling Framework and Forecast Construction
A typical exercise to construct forecasting models involve many decisions from the
practitioner, which can be summarized as follows.
1. Definition of the type of data considered in the exercise. For example, is the target
variable covariance-stationary or non-stationary? And the predictors? In this
chapter we focus on covariance-stationary data. Moreover, our aim is to predict
the mean of the target variable in the future conditioned on the information
available at the moment the forecasts are made. We do not consider forecasts for
higher moments, such as the conditional variance of the target.
2. How are the forecasts computed for each forecasting horizon? Are the forecasts
direct or iterated?
3. Usually, to evaluate the generalization potential of a forecast model/method,
running some sort of backtesting exercise is strongly advisable. The idea of
backtesting models is to assess what the models’ performance would have been if
we had generated predictions over the past history of the data. This is also called
a pseudo-out-sample exercise. The forecasting models are estimated in several
114 Medeiros
points of the past history of the data and forecasts for the ‘future’ are generated
and compared to the realized values of the target variable. The question is when
the models are estimated and based on which sample estimation is carried out.
For example, are the models estimated in a rolling window or expanding window
setup?
4. Which models are going to be considered? With the recent advances in the ML
litearure and the availability of large and new datasets, the number of potential
forecasting models has been increasing at a vertiginous pace.
In this section we review the points above and give some practical guidance to the
forecaster.
4.2.1 Setup
′
Given a sample with 𝑇 > 0 realizations of the random vector 𝑌𝑡 , 𝑾 𝑡′ , the goal is to
predict 𝑌𝑇+ℎ for horizons ℎ = 1, . . . , 𝐻. For example, 𝑌𝑡 may represent the daily sales
of a product and we want to forecast for each day of the upcoming week. Similarly,
𝑌𝑡 can be daily Covid-19 new cases or monthly inflation rates. In the later case, it is
of extreme importance to policy makers to have precise forecasts of monthly inflation
for the next 12 months, such that ℎ = 1, . . . , 12.
Throughout the chapter, we consider the following assumption:
Assumption Let {𝑫 𝑡 := (𝑌𝑡 , 𝑾 𝑡′ ) ′ }∞
𝑡=1 be a sequence of zero-mean covariance-
′
stationary stochastic process taking values on R𝑑+1 . Furthermore, E(𝑫 𝑡 𝑫 𝑡− 𝑗 ) −→ 0,
as | 𝑗 | −→ ∞. □
Therefore, we are excluding important processes that usually appear in time-series
applications. In particular, unit-roots and long-memory processes are excluded by
Assumption 1. The assumption of a zero mean is without lack of generality as we can
always consider the data to be demeaned.
4.2.2 Forecasting Equation
For (usually predetermined) integers 𝑟 ≥ 1 and 𝑠 ≥ 0 define the 𝑝-dimensional vector

′
′
of predictors 𝑿 𝑡 := 𝑌𝑡 , . . . ,𝑌𝑡−𝑟+1 , 𝑾 𝑡′ , . . . , 𝑾 𝑡−𝑠+1 where 𝑝 = 𝑟 + 𝑑𝑠 and consider
the following assumption on the data generating process (DGP):
Assumption (Data Generating Process)
𝑌𝑡+ℎ = 𝑓 ℎ ( 𝑿 𝑡 ) + 𝑈𝑡+ℎ , ℎ = 1, . . . , 𝐻, 𝑡 = 1, . . . ,𝑇 − ℎ, (4.1)
where 𝑓 ℎ (·) : R𝑛 → R is an unknown (measurable) function and {𝑈𝑡+ℎ }𝑇−ℎ

𝑡=1 is
a sequence of zero mean covariance-stationary stochastic process. In addition,
′ ) −→ ∞, as | 𝑗 | −→ 0.
E(𝑈𝑡 𝑈𝑡− □
𝑗
Model (4.1) is an example of a direct forecasting equation. In this case, the future
value of the target variable, 𝑌𝑡+ℎ , is explicitly modeled as function of the data in time
𝑡. We are going to adopt this forecasting approach in this chapter.
An alternative to direct forecast is to write a model for ℎ = 1 and iterate it in order
to produce forecasts for longer horizons. This is the iterated forecast approach. This
is trivial to be achieved for some linear specifications, but can be rather complicated
in general as it requires a forecasting model to all variables in 𝑾. Furthermore, for
nonlinear models, the construction of iterated multi-step forecasts requites numerical
evaluations of integrals via Monte Carlo techniques; see, for instance, Teräsvirta
(2006).
Example (Autoregressive Model) Consider tat the DGP is an autoregressive model
of order 1, AR(1), such that:
𝑌𝑡+1 = 𝜙𝑌𝑡 +𝑉𝑡+1 , |𝜙| < 1, (4.2)
where 𝜙 is an unknown parameter and 𝑉𝑡 is a uncorrelated zero-mean process.

By recursive iteration of (4.2), we can write:
𝑌𝑡+ℎ = 𝜙 ℎ𝑌𝑡 + 𝜙 ℎ−1𝑉𝑡+1 + · · · +𝑉𝑡+ℎ ,

(4.3)
= 𝜃𝑌𝑡 + 𝑈𝑡+ℎ , ℎ = 1, . . . , 𝐻.
where 𝜃 := 𝜙 ℎ and 𝑈𝑡+ℎ := 𝜙 ℎ−1𝑉𝑡+1 + · · · +𝑉𝑡+ℎ .

Note that the forecast for 𝑌𝑡+ℎ can be computed either by estimating model (4.2)
and iterating it ℎ-steps-ahead or by estimating (4.3) directly for each ℎ, ℎ = 1, . . . , 𝐻.□
4.2.3 Backtesting
The next choice faced the applied researcher is how the models are going to be
estimated and how they are going to be evaluated. Typically, in the time-series
literature, forecasting models are evaluated by their (pseudo) out-of-sample (OOS)
performance. Hence, the sample is commonly divided into two subsamples. The first
one is used to estimate the parameters of the model. This is known as the in-sample
(IS) estimation of the model. After estimation, forecasts are constructed for the OOS
period. However, the conclusions about the (relative) quality of a model can be
heavily dependent on how the sample is split, i.e., how many observations we use for
estimation and how many we leave to test the model. The set of observations used to
estimate the forecasting models is usually called the estimation window.
A common alternative is evaluate forecasts using different combinations of IS/OOS
periods constructed in different ways as, for example:
Expanding window: the forecaster chooses an initial window size to estimate the
models, say 𝑅. After estimation, the forecasts for 𝑡 = 𝑅 + 1, . . . , 𝑅 + ℎ are computed.
When a new observation arrives, the forecaster incorporate the new data in the
estimation window, such that 𝑅 is increased by one unit. The process is repeated until
116 Medeiros
we reach the end of the sample. See the upper panel in Figure 4.1. In this case, the
models are estimated with an increasing number of observations over time.
Rolling window: the forecaster chooses an initial window size to estimate the models,
say 𝑅. After estimation, the forecasts for 𝑡 = 𝑅 + 1, . . . , 𝑅 + ℎ are computed. When a
new observation arrives, the first data point is dropped and the forecaster incorporate
the new observation in the estimation window. Therefore, the window size is kept
constant. See lower panel in Figure 4.1. In this case, all models are estimated in a
sample with 𝑅 observations.
Expanding window
Rolling window
Estimation data point Forecasting data point Excluded data point
Fig. 4.1: Expanding versus rolling window framework
However, it is important to notice the actual number of observations used for

estimation depend on the maximum lag order in the construction of the predictor
vector 𝑿, on the choice of forecasting horizon, and on the moment the forecasts
are constructed. For example, suppose we are at the time period 𝑡 and we want to
estimate a model to forecast the target variable at 𝑡 + ℎ. If we consider an expanding
window framework, the effective number of observations used to estimate the model
is 𝑇 ∗ := 𝑇 ∗ (𝑡, 𝑟, 𝑠, ℎ) = 𝑡 − max(𝑠, 𝑟) − ℎ, for 𝑡 = 𝑅, 𝑅 + 1, . . . ,𝑇 − ℎ.1 For the rolling
window case we have
1 We drop the dependence of 𝑇 ∗ on 𝑡 , 𝑟 , 𝑠, ℎ in order to simplify notation.

(
𝑅 if 𝑡 > max(𝑟, 𝑠, ℎ)
𝑇∗ =
𝑅 − max(𝑟, 𝑠, ℎ) otherwise,
for 𝑡 = 𝑅, 𝑅 + 1, . . . ,𝑇 − ℎ. To avoid making the exercise too complicated, we can start

the backtesting of the model for 𝑡 > max(𝑟, 𝑠, ℎ).
The choice between expanding versus rolling window is not necessarily trivial.
On one hand, with expanding windows, the number of observations increase over
time, potentially yielding more precise estimators of the models. On the other hand,
estimation with expanding windows is more susceptible to the presence of structural
breaks and outliers, therefore, yielding less precise estimators. Forecasts based on
rolling windows are influenced less by structural breaks and outliers.
4.2.4 Model Choice and Estimation
𝑓 ℎ (𝒙 𝑡 ) be the forecast for 𝑌𝑡+ℎ based on information up to time 𝑡. In order

Let 𝑌b𝑡+ℎ |𝑡 := b
to estimate a given forecasting model, we must choose a loss function, L (𝑌𝑡+ℎ , 𝑌b𝑡+ℎ |ℎ ),
which measures the discrepancy h between 𝑌𝑡+ℎ iand 𝑌𝑡+ℎ |𝑡 . We define the risk function
b
as R (𝑌𝑡+ℎ , 𝑌b𝑡+ℎ |ℎ ) := E𝑌𝑡+ℎ | 𝒙𝑡 L (𝑌𝑡+ℎ , 𝑌b𝑡+ℎ |ℎ ) . The pseudo-true-model is given by
𝑓 ℎ∗ (𝒙 𝑡 ) = arg min R (𝑌𝑡+ℎ , 𝑌b𝑡+ℎ |ℎ ), (4.4)

𝑓 ℎ ( 𝒙𝑡 ) ∈ F
where F is a generic function space.

In this chapter we set 𝑓 ℎ as the conditional expectation function: 𝑓 ℎ (𝒙) :=
E(𝑌𝑡+ℎ | 𝑿 𝑡 = 𝒙).2 Therefore, the pseudo-true-model is the function 𝑓 ℎ∗ (𝒙 𝑡 ) that
minimizes the expected value of the quadratic loss:
𝑓 ℎ∗ (𝒙 𝑡 ) = arg min E𝑌𝑡+ℎ |𝑿 𝑡 [𝑌𝑡+ℎ − 𝑓 ℎ ( 𝑿 𝑡 )] 2 | 𝑿 𝑡 = 𝒙 𝑡 , ℎ = 1, . . . , 𝐻.

𝑓ℎ ( 𝑿 𝑡 ) ∈ F
In practice, the model should be estimated based on a sample of data points. Hence,
in the rolling window setup with length 𝑅 and for 𝑡 > max(𝑟, 𝑠, ℎ),
𝑡−ℎ
1 ∑︁
𝑓 ℎ ( 𝑿 𝑡−ℎ )] ′ = arg
𝑓 ℎ ( 𝑿 𝑡−𝑅−ℎ+1 ), . . . , b
[b min [𝑌𝜏+ℎ − 𝑓 ℎ ( 𝑿 𝜏 )] 2 .
𝑓ℎ ( 𝑿 𝜏 ) ∈ F 𝑅 𝜏=𝑡−𝑅−ℎ+1
However, the optimization problem stated above is infeasible when F is infinite

dimensional, as there is no efficient technique to search over all F . Of course,
one solution is to restrict the function space, as for instance, imposing linearity or
specific forms of parametric nonlinear models as in, for example, Teräsvirta (1994),
Suarez-Fariñas, Pedreira and Medeiros (2004) or McAleer and Medeiros (2008); see
also Teräsvirta, Tjøstheim and Granger (2010) for a recent review of such models.
b𝑡+ℎ|𝑡 in this Chapter, we mean an estimator of E(𝑌𝑡+ℎ |𝑿 𝑡 = 𝒙).

2 Therefore, whenever we write 𝑌
118 Medeiros
Alternatively, we can replace F by a simpler and finite dimensional F𝐷 . The idea

is to consider a sequence of finite dimensional spaces, the sieve spaces, F𝐷 , 𝐷 =
1, 2, 3, . . . , that converges to F in some norm. The approximating function 𝑓 ℎ,𝐷 ( 𝑿 𝑡 )
is written as
𝐽
∑︁
𝑓 ℎ,𝐷 ( 𝑿 𝑡 ) = 𝛽 𝑗 𝑔 ℎ, 𝑗 ( 𝑿 𝑡 ), (4.5)
𝑗=1
where 𝑔 ℎ, 𝑗 (·) is the 𝑗-th basis function for F𝐷 and can be either fully known or
indexed by a vector of parameters, such that: 𝑔 ℎ, 𝑗 ( 𝑿 𝑡 ) := 𝑔 ℎ ( 𝑿 𝑡 ; 𝜽 𝑗 ). The number
of basis functions 𝐽 := 𝐽𝑅 depends on the sample size 𝑅. 𝐷 is the dimension of the
space and it also depends on the sample size: 𝐷 := 𝐷 𝑅 .3 Therefore, the optimization
problem is then modified to
𝑓 ℎ,𝐷 ( 𝑿 𝑡 )] ′ =
𝑓 ℎ,𝐷 ( 𝑿 𝑡−𝑅−ℎ+1 ), . . . , b
[b
𝑡−ℎ
1 ∑︁ 2
arg min 𝑌𝜏+ℎ − 𝑓 ℎ,𝐷 ( 𝑿 𝜏 ) .
𝑓 ℎ,𝐷 ( 𝑿 𝑡 ) ∈ F𝐷 𝑅 𝜏=𝑡−𝑅−ℎ+1
In terms of parameters, set 𝑓 ℎ,𝐷 ( 𝑿 𝑡 ) := 𝑓 ℎ,𝐷 ( 𝑿 𝑡 ; 𝜽), where 𝜽 = (𝜽 1′ , . . . , 𝜽 ′𝐽 ) ′ :=

𝜽 ℎ,𝐷 . Therefore, the pseudo-true parameter 𝜽 ∗ is defined as
2
𝜽 ∗ := arg min E𝑌𝜏+ℎ |𝑿 𝜏 [ 𝑌𝜏+ℎ − 𝑓 ℎ,𝐷 ( 𝑿 𝜏 ; 𝜽 | 𝑿 𝜏 = 𝒙 𝜏 ]. (4.6)
𝜽 ∈R𝐷
As a consequence, the estimator for 𝜽 ∗ is

𝑡−ℎ
1 ∑︁ 2
𝜽 = arg min
b 𝑌𝜏+ℎ − 𝑓 ℎ,𝐷 ( 𝑿 𝜏 ; 𝜽) . (4.7)
𝜽 ∈R𝐷 𝑅
𝜏=𝑡−𝑅−ℎ+1
The sequence of approximating spaces F𝐷 is chosen by using the structure of the

original underlying space F and the fundamental concept of dense sets.
Definition (Dense sets) If we have two sets 𝐴 and 𝐵 ∈ X, X being a metric space, 𝐴
is dense in 𝐵 if for any 𝜖 > 0, ∈ R and 𝑥 ∈ 𝐵, there is a 𝑦 ∈ 𝐴 such that ∥𝑥 − 𝑦∥ X < 𝜖.□
The approach to approximate F by a sequence of simpler spaces is called the
method of sieves. For a comprehensive review of the method for time-series data,
see Chen (2007). There are many examples of sieves in the literature. Examples are:
polynomial series, Fourier series, trigonometric series, neural networks, etc.
When the basis functions are all known (linear sieves), the problem is linear in the
parameters and methods like ordinary least squares (when 𝐽 ≪ 𝑇 ∗ , where 𝑇 ∗ is the
size of the estimation sample) or penalized estimation can be used as we discuss later
in this chapter.
Example (Linear Sieves) From the theory of approximating functions we know that
the proper subset P ⊂ C of polynomials is dense in C, the space of continuous
3 We are assuming here that the models are estimated with a sample of 𝑅 observations.
functions. The set of polynomials is smaller and simpler than the set of all continuous
functions. In this case, it is natural to define the sequence of approximating spaces
F𝐷 , 𝐷 = 1, 2, 3, . . . by making F𝐷 the set of polynomials of degree smaller or equal
to 𝐷 − 1 (including a constant in the parameter space). Note that dim(F𝐷 ) = 𝐷 < ∞.
In the limit this sequence of finite dimensional spaces converges to the infinite
dimensional space of polynomials, which on its turn is dense in C.
Let 𝑝 = 1 and pick a polynomial basis such that
𝑓 𝐷 (𝑋𝑡 ) = 𝛽0 + 𝛽1 𝑋𝑡 + 𝛽2 𝑋𝑡2 + 𝛽3 𝑋𝑡3 + · · · + 𝛽 𝐽 𝑋𝑡𝐽 .
In this case, the dimension 𝐷 of F𝐷 is 𝐽 + 1, due to the presence of a constant term.

If 𝐽 << 𝑇 ∗ , the vector of parameters 𝜷 = (𝛽1 , . . . , 𝛽 𝐽 ) ′ can be estimated by
−1
𝜷 = 𝑿 ′𝐽 𝑿 𝐽
b 𝑿 ′𝐽 𝒀,
where 𝑿 𝐽 is the 𝑇 ∗ × (𝐽 + 1) design matrix and 𝒀 is a 𝑇 ∗ vector. When 𝐽 > 𝑇, 𝜷 can

be estimated by penalized regression:
1
𝜷 = arg min ∗ ∥𝒀 − 𝑿 𝜷∥ 22 + 𝜆𝑝( 𝜷),
b
𝜷 ∈R𝐷 𝑇
where 𝜆 > 0 and 𝑝( 𝜷) is a penalty function as discussed in Chapter 1 of this book.□

When the basis functions are also indexed by parameters (nonlinear sieves),
nonlinear least-squares methods should be used.
Example (Nonlinear Sieves) Let 𝑝 = 1 and consider the case where
𝐽
∑︁ 1
𝑓 𝐷 (𝑋𝑡 ) = 𝛽0 + 𝛽𝑓 .
𝑗=1
1 + exp −𝛾 𝑗 (𝑋𝑡 − 𝛾0 𝑗 )
This is an example of a single-hidden-layer feedforward neural network, one of the

most popular machine learning models. We discuss such models later in this chapter.
The vector of parameters is given by
𝜽 = (𝛽0 , 𝛽1 , . . . , 𝛽 𝐽 , 𝛾1 , . . . , 𝛾 𝑗 , 𝛾01 , . . . , 𝛾0𝐽 ) ′,
which should be estimated by nonlinear least squares. As the number of parameters

can be very large compared to the sample size, some sort of regularization is necessary
to estimate the model and avoid overfitting. □
The forecasting model to be estimated has the following general form:
𝑌𝑡+ℎ = 𝑓 ℎ,𝐷 ( 𝑿 𝑡 ) + 𝑍𝑡+ℎ ,

where 𝑍𝑡+ℎ = 𝑈𝑡+ℎ + 𝑓 ℎ ( 𝑿 𝑡 ) − 𝑓 ℎ,𝐷 ( 𝑿 𝑡 ) .
In a rolling window framework, the forecasts are computed as follows:
120 Medeiros
𝑓 ℎ,𝐷, (𝑡−𝑅+1:𝑡) (𝒙 𝑡 ),
𝑌b𝑡+ℎ |𝑡 = b 𝑡 = 𝑅, . . . ,𝑇 − ℎ, (4.8)
where b𝑓 ℎ,𝐷, (𝑡−𝑅ℎ +1:𝑡) (𝒙 𝑡 ) is the estimated approximating function based on data from
time 𝑡 − 𝑅 + 1 up to 𝑡 and 𝑅 is the window size. Therefore, for a given sample of 𝑇
observations, the pseudo OOS exercise results in 𝑃 ℎ := 𝑇 − ℎ − 𝑅 + 1 forecasts for
each horizon.
In practice, it is hardly the case where there is only one forecasting model. The
common scenario is to have a potentially large number of competing models. Even we
the forecaster is restricted to the linear setup there are multiple potential alternatives.
For example, different set of variables or different estimation techniques, specially in
the high-dimensional environment. Therefore, it is important to compare the forecasts
from set of alternatives available. This is what we review in the next section of this
chapter.
4.3 Forecast Evaluation and Model Comparison
Let M ℎ be a set of alternative forecasting models for the target variable 𝑌 at horizon
𝑡 + ℎ, ℎ = 1, 2, . . . , 𝐻 and 𝑡 > max(𝑟, 𝑠, ℎ) := 𝑇0 .4 Each model is defined in terms of a
vector of parameters 𝜽 𝑚 . Set 𝜽 ∗𝑚 the pseudo-true parameter and b 𝜽 𝑚 is its estimator.
In addition, set 𝑌𝑚,𝑡+ℎ |𝑡 := 𝑌𝑚(b𝜽 𝑚 ),𝑡+ℎ |𝑡 to be the forecast of 𝑌𝑡+ℎ produced by a given
b b
model 𝑚 ∈ M ℎ .
The loss function L 𝑚,𝑡 associated to the 𝑚th model is written as
L 𝑚(b𝜽),𝑡+ℎ := L (𝑌𝑡+ℎ , 𝑌b𝑚,𝑡+ℎ |𝑡 ), 𝑡 = 𝑅, . . . ,𝑇 − ℎ. (4.9)
The equality of forecasts produced by two different models, say 𝑚 1 and 𝑚 2 , can
be compared, for instance, by one of the following two hypotheses:
h i h i
(population) H0 : E L 𝑚1 (𝜽 ∗1 ),𝑡+ℎ = E L 𝑚2 (𝜽 ∗2 ),𝑡+ℎ (4.10)
h i h i
(sample) H0 : E L 𝑚1 (b𝜽 1 ),𝑡+ℎ = E L 𝑚2 (b𝜽 2 ),𝑡+ℎ . (4.11)
Although the difference between the hypotheses (4.10) and (4.11) seem minor,
there are important consequences about their properties in practice. Note that the
first one is testing equality of forecasts at population values of the parameters and
the main goal of testing such null hypotheses is to validate population models. On
the other hand, testing the second null hypothesis is the same as taken the forecasts
as given (model-free) and just test the equality of the forecasts with respect to some
expected loss function.
Testing (4.10) is way more complicated than testing (4.11). A nice discussion
about the differences can be found in West (2006), Diebold (2015), and Patton (2015).
In this chapter we focus on testing (4.11).
4 Again we are assuming that 𝑡 > max(𝑟 , 𝑠, ℎ) just to simplify notation.
4.3.1 The Diebold-Mariano Test
The Diebold-Mariano (DM) approach, proposed in Diebold and Mariano (1995),

considers the forecast errors as primitives and makes assumptions directly on those
errors. Therefore, the goal is to test the null (4.11).
As before, consider two models, 𝑚 1 and 𝑚 2 , which yield two sequence of forecasts
{𝑌b𝑚1 ,𝑡+ℎ |𝑡 }𝑇−ℎ
𝑡=𝑅 and {𝑌𝑚1 ,𝑡+ℎ |𝑡 } 𝑡=𝑅 and the respective sequence of forecast errors:
b 𝑇−ℎ
{𝑍b𝑚1 ,𝑡+ℎ |𝑡 }
𝑡=𝑅 and { 𝑍 𝑚1 ,𝑡+ℎ |𝑡 } 𝑡=𝑅 .
𝑇−ℎ b 𝑇−ℎ
Let 𝑑 (12),ℎ,𝑡 := L 𝑚1 ,𝑡+ℎ − L 𝑚2 ,𝑡+ℎ , 𝑡 = 𝑅, . . . ,𝑇 − ℎ be the loss differential between

models 𝑚 1 and 𝑚 2 . The DM statistic is given by
𝑑¯(12),ℎ
𝐷𝑀 = , (4.12)
𝜎 ( 𝑑¯(12),ℎ )
b
where 𝑑¯(12),ℎ = 𝑃1ℎ 𝑇−ℎ 𝜎 ( 𝑑¯(12),ℎ ) is an estimator of the variance of the

Í
𝑡=𝑅 𝑑 (12),ℎ,𝑡 and b
sample average of 𝑑 ℎ,𝑡 . Note that the loss differential will be a sequence of dependent
random variables and this must be taken into account.
McCracken (2020) showed that if 𝑑 (12),ℎ,𝑡 is a covariance-stationary random
variable with autocovariances decaying to zero,
𝑑
𝐷 𝑀 −→ N(0, 1),
as 𝑃 ℎ −→ ∞.5 The above result is valid, for instance, when the forecasts are computed
in a rolling window framework with fixed length.
The DM statistic can be trivially computed by the 𝑡-statistic of a regression of
the loss differential on an intercept. Furthermore, the DM test can be extended
by controlling for additional variables in the regression that may explain the loss
differential, thereby moving from an unconditional to a conditional expected loss
perspective; see Giacomini and White (2006) for a discussion.
Finally, there is also evidence that the use of Bartlett kernel should be avoided
when computing the standard error in the DM statistic and a rectangular kernel should
be used instead. Furthermore, it is advisable to use the small-sample adjustment of
Harvey, Leybourne and Newbold (1997):
√︄
𝑃 ℎ + 1 − 2ℎ + ℎ(ℎ − 1)/𝑃 ℎ
𝑀𝐷𝑀 = × 𝐷 𝑀.
𝑝ℎ
For a deeper discussion on the Diebold-Mariano test, see Clark and McCracken
(2013) or Diebold (2015).
5 In the original paper, Diebold and Mariano (1995) imposed only covariance-stationarity of 𝑑ℎ,𝑡 .
McCracken (2020) gave a counterexample where the asymptotic normality does not hold even in the
case where 𝑑ℎ,𝑡 is covariance-stationary but the autocovariances do not converge to zero.
122 Medeiros
4.3.2 Li-Liao-Quaedvlieg Test
The unconditional version of the DM test is based on the unconditional average

performance of the competing forecasts. Therefore, it ‘integrates out’ potential
heterogeneity across subsample periods. Recently, Li, Liao and Quaedvlieg (2021)
proposed a conditional test for superior predictive ability (SPA) where the null
hypothesis states that the conditional expected loss of a benchmark model is no larger
than those of the competing alternatives, uniformly across all conditioning states.
Such conditioning states are determined by a conditioning variable chosen ex-ante
by the practitioner. As a consequence, the conditional SPA (CSPA) null hypothesis
proposed by the authors asserts that the benchmark method is a uniformly weakly
dominating method among all predictive models under consideration.
Set 𝑑 (1𝑚),ℎ,𝑡 to be the loss differential between model 𝑚 ∈ M ℎ and model 𝑚 1
(benchmark). Given a user-specified conditioning variable 𝐶𝑡 , define:
ℎ 𝑚,ℎ (𝑐) := E(𝑑 (1𝑚),ℎ,𝑡 |𝐶𝑡 = 𝑐).
Note that ℎ 𝑚,ℎ (𝑐) ≥ 0 indicates that the benchmark method is expected to (weakly)
outperform the competitor conditional on 𝐶𝑡 = 𝑐. The null hypothesis of the CSPA
test is written as:
H0 : ℎ 𝑚,ℎ (𝑐) ≥ 0, ∀𝑐 ∈ C, and 𝑚 ∈ M ℎ , (4.13)
where C is the support of 𝑐.

Under H0 , the benchmark (model 𝑚 1 ) outperforms all 𝑚 models uniformly across
all conditioning states. Evidently, by the law of iterated expectations, this also implies
that the unconditional expected loss of the benchmark is smaller than those of the
alternative methods. However, the CSPA null hypothesis is generally much more
stringent than its unconditional counterpart. As such, the (uniform) conditional
dominance criterion may help the researcher differentiate competing forecasting
methods that may appear unconditionally similar.
The practical implementation of the test poses one key difficulty which is the
estimation of the unknown conditional expectation function ℎ 𝑚,ℎ (𝑐). One way of
overcoming such problem is to nonparametrically estimate ℎ 𝑚,ℎ (𝑐) by a sieve method
as described earlier in this chapter. Hence, define 𝑷(𝑐) = [ 𝑝 1 (𝑐), . . . , 𝑝 𝐽 (𝑐)] ′, where
𝑝 𝑖 (𝑐), 𝑖 = 1, . . . , 𝐽 is an approximating basis function. Therefore,
ℎ 𝑚,ℎ (𝑐) = 𝑷(𝑐) ′b

b 𝒃𝑚, (4.14)
where " #
𝑇−ℎ
−1 1 ∑︁
𝒃𝑚 = 𝑸
b b 𝑷(𝑐 𝑡 )𝑑 (1𝑚),ℎ,𝑡
𝑃 ℎ 𝑡=𝑅+1
and
𝑇−ℎ
b= 1
∑︁
𝑸 𝑷(𝑐 𝑡 )𝑷(𝑐 𝑡 ) ′ .
𝑃 ℎ 𝑡=𝑅+1
b𝑚,ℎ,𝑡 = 𝑑 (1𝑚),ℎ,𝑡 − b
Let 𝑍 ℎ 𝑚,ℎ,𝑡 := ℎ 𝑚,ℎ (𝑐 𝑡 ) and
ℎ 𝑚,ℎ,𝑡 , where b
−1 −1
b := 𝑰 𝑀 ⊗ 𝑸
𝛀 b b 𝑺 𝑰𝑀 ⊗ 𝑸
b ,
where 𝑰 𝑀 is a (𝑀 × 𝑀) identity matrix and b 𝑺 is HAC estimator of the long-run

covariance matrix of b𝒁 ℎ,𝑡 ⊗ 𝑷(𝑐 𝑡 ). b
𝒁 ℎ,𝑡 is the vector stacking 𝑍
b𝑚,ℎ,𝑡 for different
models.
The standard error of b
ℎ 𝑚,ℎ,𝑡 is given by
h i 1/2
𝜎𝑚 (𝑐) := 𝑷(𝑐) ′𝛀
b b (𝑚,𝑚) 𝑷(𝑐) ,
b (𝑚,𝑚) is the 𝑀 × 𝑀 block of 𝛀

where 𝛀 b corresponding to model 𝑚.
For a given significance level 𝛼, the rejection decision of the CSPA test is
determined in the following steps.
1. Simulate a 𝑀-dimensional zero-mean Gaussian random vector 𝝃 ∗ with covariance
∗ (𝑐) := 𝑷(𝑐) ′ 𝜉 ∗ /b
matrix 𝛀. 𝑚 𝜎𝑚 (𝑐).
b Set 𝑡 𝑚
2. Repeat step one many times. For some constant 𝑧 > 0, define 𝑞bas the 1− 𝑧/log(𝑃 ℎ )-
quantile of max1≤𝑚≤𝑀 sup𝑐 ∈ C b ∗ in the simulated sample and set
𝑡𝑚
(
h i
Vb := (𝑚, 𝑐) : bℎ 𝑚,ℎ (𝑐) ≤ min inf b ℎ 𝑚,ℎ (𝑐) + 𝑃−1/2 𝑞bb
ℎ 𝜎𝑚 (𝑐)
1≤𝑚≤𝑀 𝑐 ∈ C
) (4.15)
+ 2𝑃−1/2
ℎ 𝜎𝑚 (𝑐)
𝑞bb .
The value of 𝑧 suggested by the authors is 0.1.

3. Set 𝑞b1−𝛼 as the (1 − 𝛼)-quantile of sup (𝑚,𝑐) ∈ V ∗ (𝑐). Reject the null hypothesis
𝑡𝑚
cb
if and only if
h i
𝜂b1−𝛼 := min inf b ℎ 𝑚,ℎ (𝑐) + 𝑃−1/2
ℎ 𝑞
b 𝑚 (𝑐) < 0. (4.16)
1≤𝑚≤𝑀 𝑐 ∈ C
The set V defined in (4.15) defines an adaptive inequality selection such that,
with probability tending to 1, V contains all (𝑚, 𝑐)’s that minimize ℎ 𝑚,ℎ (𝑐). By
inspection of the null hypothesis (4.13), it is clear that whether the null hypothesis
holds or not is uniquely determined by the functions’ values at these extreme points.
The selection step focuses the test on the relevant conditioning region.
124 Medeiros
4.3.3 Model Confidence Sets
The MCS method, proposed by Hansen, Lunde and Nason (2011), consists of a
sequence of statistic tests which yields the construct of a set of superior models,
where the null hypothesis of equal predictive ability (EPA) is not rejected at a certain
confidence level. The EPA statistic tests is calculated for an arbitrary loss function
that satisfies general weak stationarity conditions
The MCS procedure starts from an initial set of models M 0 of dimension 𝑀
encompassing all the model specifications available to the user, and delivers, for
a given confidence level 1 − 𝛼, a smaller set M1−𝛼 ∗ of dimension 𝑀 ∗ ≤ 𝑀. M1−𝛼 ∗
consists as the set of the superior models. The best scenario is when the final set
consists of a single model, i.e., 𝑀 ∗ = 1.
Formally, let 𝑑 𝑚𝑛,𝑡+ℎ denotes the loss differential between models 𝑚 and 𝑛:
𝑑 𝑚𝑛,𝑡+ℎ = L (𝑚, 𝑡 + ℎ) − L (𝑛, 𝑡 + ℎ), 𝑚, 𝑛 = 1, . . . , 𝑀 and 𝑡 = 𝑅, . . . ,𝑇 − ℎ. Let
1 ∑︁
𝑑 𝑚,·,𝑡+ℎ = 𝑑 𝑚,𝑛,𝑡+ℎ , 𝑚 = 1, . . . , 𝑀, (4.17)
𝑀 −1
𝑛∈M
be the simple loss of model 𝑚 relative to any other model 𝑛 at time 𝑡 + ℎ.

The EPA hypothesis for a given set of models M can be formulated in two
alternative ways:
H0, M : 𝑐 𝑚𝑛 = 0, ∀𝑚, 𝑛 = 1, . . . , 𝑀, (4.18)

H 𝐴,M : 𝑐 𝑚𝑛 ≠ 0, for some 𝑚, 𝑛 = 1, . . . , 𝑀, (4.19)
or
H0, M : 𝑐 𝑚· = 0, ∀𝑚 = 1, . . . , 𝑀, (4.20)
H 𝐴,M : 𝑐 𝑚· ≠ 0, for some 𝑚 = 1, . . . , 𝑀, (4.21)
where 𝑐 𝑚𝑛 = E(𝑑 𝑚𝑛,𝑡+ℎ ) and 𝑐 𝑚· = E(𝑑 𝑚·𝑡+ℎ )

In order to test the two hypothesis above, the following two statistics are constructed:
𝑑¯𝑚𝑛 𝑑¯𝑚·
𝑡 𝑚𝑛 = and 𝑡 𝑚· = , (4.22)
𝜎 ( 𝑑¯𝑚𝑛 )
b 𝜎 ( 𝑑¯𝑚· )
b
where
𝑇−ℎ
1 ∑︁ 1 ∑︁ ¯
𝑑¯𝑚𝑛 := 𝑑 𝑚𝑛,𝑡+ℎ , and 𝑑¯𝑚· := 𝑑 𝑚𝑛 ,
𝑃 ℎ 𝑡=𝑅 𝑀 −1
𝑛∈M
𝜎 ( 𝑑¯𝑚𝑛 ) and b
and b 𝜎 ( 𝑑¯𝑚· ) are estimates of the standard deviation of 𝑑¯𝑚𝑛 and 𝑑¯𝑚· ,
respectively. The standard deviations are estimated by block bootstrap.
The null hypotheses of interest map naturally into the following two test statistics:
T𝑅,M = max |𝑡 𝑚𝑛 | (4.23)

𝑚,𝑛∈M
and
Tmax, M = max 𝑡 𝑚· . (4.24)
𝑚∈M
The MCS procedure consists on a sequential testing procedure which eliminates at

each step the worst model, until the EPA hypothesis is not rejected for all the models
belonging to the set. The choice of the worst model to be eliminated has been made
using an elimination rule as follows:

𝑒 𝑅,M = arg max sup 𝑡 𝑚𝑛 and 𝑒 max, M = arg max 𝑡 𝑚· . (4.25)
𝑚 𝑛∈M 𝑚∈M
Therefore, the MCS consists of the following steps:

Algorithm (MCS Procedure) 1. Set M = M 0 ;
2. test for EPA–hypothesis: if EPA is not rejected terminate the algorithm and set
∗
M1−𝛼 = M. Otherwise, use the elimination rules defined in equations (9) to
determine the worst model;
3. remove the worst model, and go to step 2.
4.4 Linear Models
We start by reviewing some ML methods based on the assumption that the target
function 𝑓 ℎ ( 𝑿 𝑡 ) is linear. We focus on methods that have not been previously
discussed in the first chapter of this book.
Under linearity, the pseudo-true model is given as
𝑓 ℎ∗ (𝒙 𝑡 ) := E(𝑌𝑡+ℎ | 𝑿 𝑡 = 𝒙 𝑡 ) = 𝒙 𝑡′ 𝜽.
Therefore, the class of approximating functions is also linear. We consider three dif-
ferent approaches: Factor-based regression, the combination of factors and penalized
regression, and ensemble methods.
4.4.1 Factor Regression
The core idea of factor-based regression is to replace the large dimensional set of
potential predictors, 𝑾 𝑡 ∈ R𝑑 , by a low dimensional set of latent factors 𝑭 𝑡 , which
take values on R 𝑘 , 𝑘 << 𝑑, and that are estimated as rotations of the original dataset.
More specifically, let
𝑾 𝑡 = 𝚲𝑭 𝑡 + 𝑽 𝑡 , 𝑡 = 1, . . . ,𝑇, (4.26)
such that the regressors 𝑿 𝑡 can be re-defined as
′
˜ 𝑡 = 𝑌𝑡 , . . . ,𝑌𝑡−𝑟+1 , 𝑭 𝑡′ , . . . , 𝑭 ′
𝑿 𝑡−𝑠+1 .
126 Medeiros
Note that the dimension of 𝑿 ˜ 𝑡 is 𝑝˜ = 𝑟 + 𝑘 𝑠, such that 𝑝˜ << 𝑝 as 𝑘 << 𝑑.

Set 𝜷 := ( 𝜷 𝐴𝑅 , 𝜷1,𝑊 , . . . , 𝜷 𝑠,𝑊 ) ′ such that:
𝑌𝑡+ℎ = 𝒀 𝑡′ 𝜷 𝐴𝑅 + 𝑾 𝑡′ 𝜷1,𝑊 + . . . + 𝑾 𝑡−𝑠+1

′
𝜷 𝑠,𝑊 + 𝑈𝑡+ℎ ,
where 𝒀 𝑡 = (𝑌𝑡 , . . . ,𝑌𝑡−𝑟+1 ) ′. Hence, the forecasting model becomes
𝑌𝑡+ℎ = 𝒀 𝑡′ 𝜷 𝐴𝑅 + 𝑭 𝑡′ 𝚲′ 𝜷1,𝑊 + . . . + 𝑭 𝑡−𝑠+1

′
𝚲′ 𝜷 𝑠,𝑊
+ 𝑽 𝑡′ 𝜷1,𝑊 + . . . + 𝑽 𝑡−𝑠+1
′
𝜷 𝑠,𝑊 + 𝑈𝑡+ℎ (4.27)
˜ ′
= 𝑿 𝑡 𝜽 + 𝜖 𝑡+ℎ ,
where
′
𝜖 𝑡+ℎ = 𝑽 𝑡′ 𝜷1,𝑊 + . . . + 𝑽 𝑡−𝑠+1 𝜷 𝑠,𝑊 + 𝑈𝑡+ℎ ,
′
and 𝜽 = 𝜷 ′𝐴𝑅 , 𝚲′ 𝜷1,𝑊 , · · · , 𝚲′ 𝜷 𝑠,𝑊 .

In order to estimate 𝚲 and 𝑭 𝑡 we make the following assumption.
Assumption (Factor Model) Assume:
(a) E(𝑭 𝑡 ) = 0, E 𝑭 𝑡 𝑭 𝑡′ = 𝑰 𝑟 and 𝚲′𝚲 is a diagonal matrix;

(b) All eigenvalues of 𝚲′𝚲/𝑛 are bounded away from zero and infinity as 𝑛 → ∞;
(c) ∥𝚺 − 𝚲𝚲′ ∥ = 𝑂 (1); and
(d) ∥𝚲∥ max ≤ 𝐶. □
Under Assumption 3, the factors can be estimated as the 𝑘 most important
principal components of the sample covariance matrix of 𝑾 and the parameter 𝜽 can
be estimated by OLS.
The principal components should be estimated in steps as follows:
1. For each 𝑡 = max(𝑟, 𝑠, ℎ) + 1, . . . ,𝑇 − ℎ, construct the (𝑅 × 𝑑) matrix
©𝑊1,𝑡−𝑅−ℎ+1 · · · 𝑊𝑑,𝑡−𝑅−ℎ+1 ª
.. .. ...
®
𝑾 = . . ®.
®
®
« 𝑊1,𝑡−ℎ · · · 𝑊𝑑,𝑡−𝑅−ℎ+1 ¬
2. Construct the standardized version 𝑾e of 𝑾 such that each column of 𝑾 e has zero
1 e′e
mean and 𝑅 𝑾 𝑾 = 𝑰, where 𝑰 is a (𝑑 × 𝑑) identity matrix.
3. Compute the ordered (in ascending order) eigenvalues and the respective eigen-
vectors of 𝑾.e The 𝑖th factor is given by 𝑭 𝑖 = 𝑾𝜸
e 𝑖 , 𝑖 = 1, . . . , min(𝑑, 𝑅), where
𝜸 𝑖 is the eigenvector associated with the 𝑖th eigenvalue.
4. Select the number of factors by one of the methods described below.
The optimal number of factors can be selected in different ways. For example,
Ahn and Horenstein (2013) suggest selecting the factors as6
6 See also Onatski (2010).

𝜆𝑖
𝑘 = arg
b max ,
𝑘 ∈ {1,...,max(𝑑,𝑅)−1} 𝜆𝑖+1
where 𝜆𝑖 is the 𝑖th larger eigenvalue of 𝑾.

e
Bai and Ng (2002) suggest the following estimator for the number of factors:
𝑘 = arg
b max 𝐼𝐶 (𝑘),
𝑘 ∈ {1,...,min(𝑑,𝑅)−1}
where 𝐼𝐶 is a information criterion and possible options are:

𝑑+𝑅 𝑑𝑅
𝐼𝐶1 (𝑘) = log[𝑆(𝑘)] + 𝑘 log ,
𝑑𝑅 𝑑+𝑅
𝑛 +𝑇 2
𝐼𝐶2 (𝑘) = log[𝑆(𝑘)] + 𝑘 log 𝐶𝑑𝑅 ,
𝑑𝑅
2
log 𝐶𝑑𝑅
𝐼𝐶3 (𝑘) = log[𝑆(𝑘)] + 𝑘 2
, or
𝐶𝑑𝑅
(𝑑 + 𝑅 − 𝑘) log(𝑑𝑅)
𝐼𝐶4 (𝑘) = log[𝑆(𝑘)] + 𝑘 ,
𝑑𝑅
1 e b b 2
√︁
where 𝑆(𝑘) = 𝑑𝑅 ∥ 𝑾 − 𝚲 𝑘 𝑭 𝑘 ∥ 2 and 𝐶𝑑𝑅 := min(𝑑, 𝑅).
The factor-based regression model described here assumes that all regressors are
important but the relevant information can be summarized by a small number of
factors. On the other hand, several models presented in Chapter 1 take an opposite
direction by assuming a sparsity hypothesis, i.e., only a small number of regressors
are in fact relevant.
Although, factor and sparse models are two widely used methods to impose a
low-dimensional structure in high-dimension, they are seemingly mutually exclusive
and there is a large debate in the literature on the relevance of each of them; see, for
example, Giannone, Lenza and Primiceri (2021) and Fava and Lopes (2020).
Recently, Fan, Masini and Medeiros (2021) propose a lifting method that combines
the merits of these two models in a supervised learning methodology that allows to
efficiently explore all the information in high-dimensional datasets. We describe the
method in the next section.
4.4.2 Bridging Sparse and Dense Models
The method proposed by Fan et al. (2021) is based on a flexible model for high-
dimensional panel data, called factor-augmented regression (FarmPredict) model
with both observable or latent common factors, as well as idiosyncratic components.
This model not only includes both principal component (factor) regression and
sparse regression as specific models, but also significantly weakens the cross-
sectional dependence and hence, facilitates model selection and interpretability. The
128 Medeiros
methodology consists of three steps. Although their methodology is more general, it

can be adapted to forecasting problems.
The method can be implemented in the following steps.
1. Run principal component analysis on the vector (𝑌𝑡 , 𝑾 𝑡′ ) ′ and select the number
of factors by an appropriate method.
2. Compute the idiosyncratic errors 𝑽b𝑡 = (𝑌𝑡 , 𝑾 𝑡′ ) ′ − 𝚲
bb𝑭 𝑡 , where 𝚲
b is the matrix of
estimated loadings.
3. Run a LASSO regression of 𝑉 b1𝑡 on 𝑽b−1𝑡 , where 𝑽 b−1𝑡 is the vector with all the
estimated idiosyncratic residuals with the exception of the first one. Call the fitted
value 𝑉
e1𝑡 .
4. Estimate the forecasting model by OLS:
𝑌b𝑡+ℎ |𝑡 =𝑌𝑡 𝛽1, 𝐴𝑅 + . . . +𝑌𝑡−𝑟+1 𝛽𝑟 , 𝐴𝑅

′ ′ (4.28)
+b 𝜷1,𝐹 + . . . + b
𝑭𝑡 b 𝜷 𝑠,𝐹 + 𝑉
𝑭 𝑡−𝑠+1 b e𝑡− 𝑝+1 𝝅 𝑝 .
e𝑡 𝝅 1 + · · · + 𝑉
Another variation of the method is as follows.

1. Define a benchmark model for forecasting 𝑌𝑡+ℎ . For example, an autoregressive
bench . Set 𝑅
model. Call the forecasts 𝑌b𝑡+ℎ b𝑡 = 𝑌𝑡+ℎ − 𝑌bbench .
𝑡+ℎ
′
2. Run principal component analysis for the vector 𝑅 b𝑡 , 𝑾 𝑡′ and compute the
′
idiosyncratic residuals: 𝑽
b𝑡 = 𝑅 b𝑡 , 𝑾 𝑡′ − 𝚲bb𝑭𝑡 .
PCR = 𝑌 ′
3. Set 𝑌b𝑡+ℎ bbench + 𝝀b1 b𝑭 𝑡 . 𝝀b1 is the OLS estimate of the coefficient of the
𝑡+ℎ
regression of 𝑅b𝑡+ℎ on b𝑭 𝑡 . Run a LASSO regression of 𝑉
b1𝑡 on 𝑽
b−1𝑡 . Call the fitted
value 𝑉e1𝑡 .
4. Finally, construct the forecast as:
PCR e
𝑌𝑡+ℎ = 𝑌b𝑡+ℎ + 𝑉1𝑡 . (4.29)
4.4.3 Ensemble Methods
4.4.3.1 Bagging
The term bagging means Bootstrap Aggregating and was proposed by Breiman
(1996) to reduce the variance of unstable predictors7. For example, imagine a linear
predictive regression model, where the variables included in the final model depends
on the result of a standard 𝑡-test. Furthermore, suppose that some of the regressors
are relevant but the ‘true’ parameters, although being non-zero, are quite small as
compared to the size of the sample used to estimate the model. In this case, the
𝑡-statistic can yield the rejection or not of the null hypothesis of a zero coefficient due
7 An unstable predictor has large variance. Intuitively, small changes in the data yield large changes
in the predictive model.
to sample variability and the regressor can be erroneously removed from the model,
generating poorer forecasts. The idea of Baggging is to alleviate this problem by
estimating the model, running the test, and computing forecasts, in several different
samples and taking the average of the forecasts over the samples. Each sample is
constructed by bootstrap, which is a well-known resampling technique.
Bagging was popularized in the time series literature by Inoue and Kilian (2008),
who to construct forecasts from multiple regression models with local-to-zero
regression parameters and errors subject to possible serial correlation or conditional
heteroscedasticity. Bagging is designed for situations in which the number of predictors
is moderately large relative to the sample size.
The bagging algorithm in time-series settings have to take into account the time
dependence dimension when constructing the bootstrap samples.
Algorithm (Bagging for Time-Series Models) The Bagging algorithm is defined as
follows:
1. For each 𝑡, arrange the data as
𝑌𝜏+ℎ , 𝑿 ′𝜏 , 𝜏 = 𝑡 − 𝑅 − ℎ + 1, . . . , 𝑡 − ℎ,

in the form of a matrix 𝑽 of dimension 𝑅 × 𝑝.

2. Construct (block) bootstrap samples with 𝑅 observations of the form
n o
∗
𝑌 (𝑖)1 , 𝑿 ′∗ ∗ ′∗
(𝑖)1 , . . . , 𝑌 (𝑖) 𝑅 , 𝑿 (𝑖) 𝑅 , 𝑖 = 1, . . . , 𝐵,
by drawing blocks of 𝑀 rows of 𝑽 with replacement.

3. Compute the 𝑖th bootstrap forecast as
(
∗ 0 if |𝑡 ∗𝑗 | < 𝑐 ∀ 𝑗,
𝑌b(𝑖)𝑡+ℎ |𝑡 = ∗ ∗ (4.30)
b 𝑿 (𝑖)𝑡 otherwise,
𝝀 (𝑖) e
∗
𝑿 (𝑖)𝑡 := 𝑺∗(𝑖)𝑡 𝒁 ∗(𝑖)𝑡 and 𝑺𝑡 is a diagonal selection matrix with 𝑗th diagonal
where e
element given by (
1 if |𝑡 𝑗 | > 𝑐,
I { |𝑡 𝑗 |>𝑐 } =
0 otherwise,
∗
𝝀 (𝑖) is the OLS estimator at each
𝑐 is a pre-specified critical value of the test. b
bootstrap repetition.
4. Compute the average forecasts over the bootstrap samples:
𝐵
1 ∑︁ b∗
𝑌˜𝑡+ℎ |𝑡 = 𝑌 .
𝐵 𝑖=1 (𝑖)𝑡 |𝑡−1
In algorithm 2, above, one requires that it is possible to estimate and conduct

inference in the linear model. This is certainly infeasible if the number of predictors
is larger than the sample size (𝑝 > 𝑇), which requires the algorithm to be modified.
130 Medeiros
Garcia, Medeiros and Vasconcelos (2017) and Medeiros, Vasconcelos, Veiga and
Zilberman (2021) adopt the following changes of the algorithm:
Algorithm (Bagging for Time-Series Models and Many Regressors) The Bagging
algorithm is defined as follows.
0. Run 𝑝 univariate regressions of 𝑌𝑡+ℎ on each covariate in 𝑿 𝑡 . Compute 𝑡-statistics
and keep only the ones that turn out to be significant at a given pre-specified
ˇ 𝑡.
level. Call this new set of regressors as 𝑿
1–4. Same as before but with 𝑿 𝑡 replaced by 𝑿 ˇ 𝑡.
4.4.3.2 Complete Subset Regression
Complete Subset Regression (CSR) is a method for combining forecasts developed by

Elliott et al. (2013, 2015). The motivation was that selecting the optimal subset of 𝑿 𝑡
to predict 𝑌𝑡+ℎ by testing all possible combinations of regressors is computationally
very demanding and, in most cases, unfeasible. For a given set of potential predictor
variables, the idea is to combine forecasts by averaging8 all possible linear regression
models with fixed number of predictors. For example, with 𝑝 possible predictors,
there are 𝑝 unique univariate models and
𝑝!
𝑝 𝑞,𝑛 =
( 𝑝 − 𝑞)!𝑞!
different 𝑞-variate models for 𝑞 ≤ 𝑄. The set of models for a fixed value of 𝑞 as is
known as the complete subset.
When the set of regressors is large the number of models to be estimated increases
rapidly. Moreover, it is likely that many potential predictors are irrelevant. In these
cases it was suggested that one should include only a small, 𝑞, ˜ fixed set of predictors,
such as five or ten. Nevertheless, the number of models still very large, for example,
with 𝑝 = 30 and 𝑞 = 8, there are 5, 852, 925 regressions to be estimated. An alternative
solution is to follow Garcia et al. (2017) and Medeiros et al. (2021) and adopt a
similar strategy as in the case of Bagging high-dimensional models. The idea is
to start fitting a regression of 𝑌𝑡+ℎ on each of the candidate variables and save the
𝑡-statistics of each variable. The 𝑡-statistics are ranked by absolute value, and we select
the 𝑝˜ variables that are more relevant in the ranking. The CSR forecast is calculated
on these variables for different values of 𝑞. Another possibility is to pre-select the
variables by elastic-net or other selection method; see, Chapter 1 for details.
8 It is possible to combine forecasts using any weighting scheme. However, it is difficult to beat
uniform weighting (Genre, Kenny, Meyler & Timmermann, 2013).
4.5 Nonlinear Models
4.5.1 Feedforward Neural Networks
4.5.1.1 Shallow Neural Networks
Neural Network (NN) is one of the most traditional nonlinear sieve methods. NN can
be classified into shallow (single hidden layer) or deep networks (multiple hidden
layers). We start describing the shallow version. The most common shallow NN is
the feedforward neural network.
Definition (Feedforward NN model) In the single hidden layer feedforward NN
(sieve) model, the approximating function 𝑓 ℎ,𝐷 ( 𝑿 𝑡 ) is defined as
𝐽
∑︁
𝑓 ℎ,𝐷 ( 𝑿 𝑡 ) := 𝑓 ℎ,𝐷 ( 𝑿 𝑡 ; 𝜽) = 𝛽0 + 𝛽 𝑗 𝑆(𝜸 ′𝑗 𝑿 𝑡 + 𝛾0, 𝑗 ),
𝑗=1
(4.31)
𝐽
∑︁
= 𝛽0 + 𝛽 𝑗 𝑆( 𝜸˜ ′𝑗 𝑿
˜ 𝑡 ),
𝑗=1
In (4.31), 𝑿 ˜ 𝑡 = (1, 𝑿 𝑡′ ) ′, 𝑆 𝑗 (·) is a basis function and the parameter vector to

be estimated is given by 𝜽 = (𝛽0 , . . . , 𝛽𝐾 , 𝜸 1′ , . . . , 𝜸 ′𝐽 , 𝛾0,1 , . . . , 𝛾0, 𝐽 ) ′, where 𝜸˜ 𝑗 =
(𝛾0, 𝑗 , 𝜸 ′𝑗 ) ′. □
NN models form a very popular class of nonlinear sieves where the function
˜ 𝑡 ). Such kind of model has been used in many
𝑔 ℎ, 𝑗 ( 𝑿 𝑡 ) in (4.5) is given by 𝑆( 𝜸˜ ′𝑗 𝑿
applications of forecasting for many decades. Usually, the basis functions 𝑆(·) are
called activation functions and the parameters are called weights. The terms in the sum
are called hidden-neurons as an unfortunate analogy to the human brain. Specification
(4.31) is also known as a single hidden layer NN model as is usually represented
in the graphical form as in Figure 4.2. The green circles in the figure represent the
input layer which consists of the regressors of the model (𝑿 𝑡 ). In the figure there are
four input variables. The blue and red circles indicate the hidden and output layers,
respectively. In the example, there are five elements (called neurons in the NN jargon)
in the hidden layer. The arrows from the green to the blue circles represent the linear
combination of inputs: 𝜸 ′𝑗 𝑿 𝑡 + 𝛾0, 𝑗 , 𝑗 = 1, . . . , 5. Finally, the arrows from the blue
to theÍred circles represent the linear combination of outputs from the hidden layer:
𝛽0 + 5𝑗=1 𝛽 𝑗 𝑆(𝜸 ′𝑗 𝑿 𝑡 + 𝛾0, 𝑗 ).
There are several possible choices for the activation functions. In the early days,
𝑆(·) was chosen among the class of squashing functions as per the definition below.
Definition (Squashing (sigmoid) function) A function 𝑆 : R −→ [𝑎, 𝑏], 𝑎 < 𝑏, is a
squashing (sigmoid) function if it is non-decreasing, lim 𝑆(𝑥) = 𝑏 and lim 𝑆(𝑥) =
𝑥−→∞ 𝑥−→−∞
𝑎. □
132 Medeiros
Fig. 4.2: Graphical representation of a single hidden layer neural network
Historically, the most popular choices are the logistic and hyperbolic tangent
functions such that:
1
Logistic: 𝑆(𝑥) =
1 + exp(−𝑥)
exp(𝑥) − exp(−𝑥)
Hyperbolic tangent: 𝑆(𝑥) = .
exp(𝑥) + exp(−𝑥)
The popularity of such functions was partially due to theoretical results on function
approximation. Funahashi (1989) establishes that NN models as in (4.31) with generic
squashing functions are capable of approximating any continuous functions from
one finite dimensional space to another to any desired degree of accuracy, provided
that 𝐽𝑇 is sufficiently large. Cybenko (1989) and Hornik, Stinchombe and White
(1989) simultaneously proved approximation capabilities of NN models to any
Borel measurable function and Hornik et al. (1989) extended the previous results
and showed the NN models are also capable to approximate the derivatives of the
unknown function. Barron (1993) relate previous results to the number of terms in
the model.
Stinchcombe and White (1989) and Park and Sandberg (1991) derived the same
results of Cybenko (1989) and Hornik et al. (1989) but without requiring the activation
function to be sigmoid. While the former considered a very general class of functions,
the later focused on radial-basis functions (RBF) defined as:
Radial Basis: 𝑆(𝑥) = exp(−𝑥 2 ).
More recently, Yarotsky (2017) showed that the rectified linear units (ReLU) as
Rectified Linear Unit: 𝑆(𝑥) = max(0, 𝑥),
are also universal approximators and ReLU activation function is one of the most
popular choices among practitioners due to the following advantages:
1. Estimating NN models with ReLU functions is more efficient computationally as
compared to other typical choices, such as logistic or hyperbolic tangent. One
reason behind such improvement in performance is that the output of the ReLU
function is zero whenever the inputs are negative. Thus, fewer units (neurons)
are activated, leading to network sparsity.
2. In terms of mathematical operations, the ReLU function involves simpler oper-
ations than the hyperbolic tangent and logistic functions, which also improves
computational efficiency.
3. Activation functions like the hyperbolic tangent and the logistic functions may
suffer from the vanishing gradient problem, where gradients shrink drastically
during optimization, such that the estimates are no longer improved. ReLU avoids
this by preserving the gradient since it is an unbounded function.
However, ReLU functions suffer from the dying activation problem: many ReLU
units yield output values of zero, which happens when the ReLU inputs are negative.
While this characteristic gives ReLU its strengths (through network sparsity), it
becomes a problem when most of the inputs to these ReLU units are in the negative
range. The worst-case scenario is when the entire network dies, meaning that it
becomes just a constant function. A solution to this problem is to use some modified
versions of the ReLU function, such as the Leaky ReLU (LeReLU):
Leaky ReLU: 𝑆(𝑥) = max(𝛼𝑥, 𝑥),
where 0 < 𝛼 < 1.

For each estimation window, 𝑡 = max(𝑟, 𝑠, ℎ) + 1, . . . ,𝑇 − ℎ, model (4.31) can be
written in matrix notation. Let 𝚪 = ( 𝜸˜ 1 , . . . , 𝜸˜ 𝐽 ) be a ( 𝑝 + 1) × 𝐽 matrix,
134 Medeiros
©1 𝑋1,𝑡−𝑅−ℎ+1 · · · 𝑋 𝑝,𝑡−𝑅−ℎ+1 ª
®
1 𝑋1,𝑡−𝑅−ℎ+2 · · · 𝑋 𝑝,𝑡−𝑅−ℎ+2 ®
𝑿 = . ®, and
®
.. .. ..
. . ®
®
®
«1 𝑋 1,𝑡−ℎ · · · 𝑋 𝑝,𝑡−ℎ ¬
| {z }
𝑅×( 𝑝+1)
′ ˜ ′ ˜
©1 𝑆( 𝜸˜ 1 𝑿 𝑡−𝑅−ℎ+1 ) · · · 𝑆( 𝜸˜ 𝐽 𝑿 𝑡−𝑅−ℎ+1 ) ª
′ ˜ ′ ˜
®
1 𝑆( 𝜸˜ 1 𝑿 𝑡−𝑅−ℎ+2 ) · · · 𝑆( 𝜸˜ 𝐽 𝑿 𝑡−𝑅−ℎ+2 ) ®
®
O( 𝑿𝚪) = . . .. .. ® .
. .. . ®
. . ®
®
1 𝑆( 𝜸
˜ ′𝑿 ˜ 𝑡−ℎ ) · · · 𝑆( 𝜸˜ ′ 𝑿 ˜ 𝑡−ℎ )
« 1 𝐽 ¬
| {z }
𝑅×( 𝐽+1)
Therefore, by defining 𝜷 = (𝛽0 , 𝛽1 , . . . , 𝛽𝐾 ) ′, the output of a feedforward NN is given

by:
𝒇 𝐷 ( 𝑿, 𝜽) = [ 𝑓 𝐷 ( 𝑿 𝑡−𝑅−ℎ+1 ; 𝜽), . . . , 𝑓 𝐷 ( 𝑿 𝑡−ℎ ; 𝜽)] ′
 𝛽0 + 𝐽𝑗=1 𝛽 𝑗 𝑆(𝜸 ′𝑗 𝑿 𝑡−𝑅−ℎ+1 + 𝛾0, 𝑗 ) 

 Í 
 
..
= (4.32)
 
 . 

′
 Í𝐽 
 𝛽0 + 𝑗=1 𝛽 𝑗 𝑆(𝜸 𝑗 𝑿 𝑡−ℎ + 𝛾0, 𝑗 ) 
 
= O( 𝑿𝚪) 𝜷.
The number of hidden units (neurons), 𝐽, and the choice of activation functions
are known as the architecture of the NN model. Once the architecture is defined, the
dimension of the parameter vector 𝜽 = [𝚪 ′, 𝜷 ′] ′ is 𝐷 = ( 𝑝 + 1) × 𝐽 + (𝐽 + 1) and can
easily get very large such that the unrestricted estimation problem defined as
𝜽 = arg min ∥𝒀 − O ( 𝑿𝚪) 𝜷∥ 22

b
𝜽 ∈R𝐷
is unfeasible. A solution is to use regularization as in the case of linear models and

consider the minimization of the following function:
𝑄(𝜽) = ∥𝒀 − O ( 𝑿𝚪) 𝜷∥ 22 + 𝑝(𝜽), (4.33)
where usually 𝑝(𝜽) = 𝜆𝜽 ′ 𝜽. Traditionally, the most common approach to minimize

(4.33) is to use Bayesian methods as in MacKay (1992a), (1992b), and Foresee and
Hagan (1997). See also Chapter 2 for more examples of regularization with nonlinear
models.
A more modern approach is to use a technique known as Dropout (Srivastava,
Hinton, Krizhevsky, Sutskever & Salakhutdinov, 2014). The key idea is to randomly
drop neurons (along with their connections) from the neural network during estimation.
A NN with 𝐽 neurons in the hidden layer can generate 2 𝐽 possible thinned NN by
just removing some neurons. Dropout samples from this 2 𝐽 different thinned NN
and train the sampled NN. To predict the target variable, we use a single unthinned
network that has weights adjusted by the probability law induced by the random drop.
This procedure significantly reduces overfitting and gives major improvements over
other regularization methods.
We modify equation (4.31) by
𝐽
∑︁
𝑓 𝐷− ( 𝑿 𝑡 ) = 𝛽0 + 𝑠 𝑗 𝛽 𝑗 𝑆(𝜸 ′𝑗 [𝒓 ⊙ 𝑿 𝑡 ] + 𝑣 𝑗 𝛾0, 𝑗 ),
𝑗=1
where 𝑠, 𝑣, and 𝒓 = (𝑟 1 , . . . , 𝑟 𝑝 ) are independent Bernoulli random variables each

with probability 𝑞 of being equal to 1. The NN model is thus estimated by using
𝑓 𝐷− ( 𝑿 𝑡 ) instead of 𝑓 𝐷 ( 𝑿 𝑡 ) where, for each training example, the values of the entries
of 𝑠, 𝑣, and 𝒓 are drawn from the Bernoulli distribution. The final estimates for 𝛽 𝑗 ,
𝜸 𝑗 , and 𝛾0, 𝑗 are multiplied by 𝑞.
4.5.1.2 Deep Neural Networks
A Deep Neural Network model is a straightforward generalization of specification

(4.31), where more hidden layers are included in the model, as represented in Figure
4.3. In the figure, we represent a Deep NN with two hidden layers with the same
number of hidden units in each. However, the number of hidden units (neurons) can
vary across layers.
As pointed out in Mhaska, Liao and Poggio (2017), while the universal approxim-
ation property holds for shallow NNs, deep networks can approximate the class of
compositional functions as well as shallow networks but with exponentially lower
number of training parameters and sample complexity.
Set 𝐽ℓ as the number of hidden units in layer ℓ ∈ {1, . . . , 𝐿}. For each hidden layer
ℓ define 𝚪ℓ = ( 𝜸˜ 1ℓ , . . . , 𝜸˜ 𝑘ℓ ℓ ). Then the output Oℓ of layer ℓ is given recursively by
′ · · · 𝑆( 𝜸˜ ′𝑘ℓ ℓ O1ℓ−1 (·)) ª

©1 𝑆( 𝜸˜ 1ℓ O1ℓ−1 (·))
®
1 𝑆( 𝜸˜ ′ O2ℓ−1 (·)) · · · 𝑆( 𝜸˜ ′𝑘ℓ ℓ O2ℓ−1 (·)) ®®
1ℓ
Oℓ (Oℓ−1 (·)𝚪ℓ ) = .

} .. . .. .. ®
| {z .. . . ®
®
𝑝×( 𝐽ℓ +1) ®
′ ′
«1 𝑆( 𝜸˜ 1ℓ O 𝑛ℓ−1 (·)) · · · 𝑆( 𝜸˜ 𝐽ℓ ℓ O 𝑛ℓ−1 (·)) ¬
where O0 := 𝑿. Therefore, the output of the Deep NN is the composition
𝒉 𝐷 ( 𝑿) = O 𝐿 (· · · O3 (O2 (O1 ( 𝑿𝚪1 )𝚪2 )𝚪3 ) · · · ) 𝜷.

136 Medeiros
Fig. 4.3: Deep neural network architecture
The estimation of the parameters is usually carried out by stochastic gradient

descend methods with dropout to control the complexity of the model.
4.5.2 Long Short Term Memory Networks
Broadly speaking, Recurrent Neural Networks (RNNs) are NNs that allow for feedback
among the hidden layers. RNNs can use their internal state (memory) to process
sequences of inputs. In the framework considered in this chapter, a generic RNN
could be written as
𝑯 𝑡 = 𝒇 (𝑯 𝑡−1 , 𝑿 𝑡 ),
𝑌b𝑡+ℎ |𝑡 = 𝑔(𝑯 𝑡 ),
where 𝑌b𝑡+ℎ |𝑡 is the prediction of 𝑌𝑡+ℎ given observations only up to time 𝑡, 𝒇 and 𝑔
are functions to be defined and 𝑯 𝑡 is what we call the 𝑘-dimensional (hidden) state.
From a time-series perspective, RNNs can be seen as a kind of nonlinear state-space
model.
RNNs can remember the order that the inputs appear through its hidden state
(memory) and they can also model sequences of data so that each sample can be
assumed to be dependent on previous ones, as in time-series models. However, RNNs
are hard to be estimated as they suffer from the vanishing/exploding gradient problem.
For each estimation window, set the cost function to be
𝑡−ℎ
∑︁ 2
Q (𝜽) = 𝑌𝜏+ℎ − 𝑌b𝜏+ℎ | 𝜏 ,
𝜏=𝑡−𝑅−ℎ+1
where 𝜽 is the vector of parameters to be estimated. It is easy to show that the gradient
𝜕Q (𝜽)
𝜕𝜽 can be very small or diverge. Fortunately, there is a solution to the problem
proposed by Hochreiter and Schmidhuber (1997). A variant of RNN which is called
Long-Short-Term Memory (LSTM) network . Figure 4.4 shows the architecture of
a typical LSTM layer. A LSTM network can be composed of several layers. In the
figure, red circles indicate logistic activation functions, while blue circles represent
hyperbolic tangent activation. The symbols ‘X’ and ‘+’ represent, respectively, the
element-wise multiplication and sum operations. The RNN layer is composed of
several blocks: the cell state and the forget, input, and ouput gates. The cell state
introduces a bit of memory to the LSTM so it can ‘remember’ the past. LSTM learns
to keep only relevant information to make predictions, and forget non relevant data.
The forget gate tells which information to throw away from the cell state. The output
gate provides the activation to the final output of the LSTM block at time 𝑡. Usually,
the dimension of the hidden state (𝑯 𝑡 ) is associated with the number of hidden
neurons.
Algorithm 4 describes analytically how the LSTM cell works. 𝒇 𝑡 represents the
output of the forget gate. Note that it is a combination of the previous hidden-state
(𝑯 𝑡−1 ) with the new information (𝑿 𝑡 ). Note that 𝒇 𝑡 ∈ [0, 1] and it attenuates the signal
coming com 𝒄 𝑡−1 . The input and output gates have the same structure. Their function
is to filter the ‘relevant’ information from the previous time period as well as from
the new input. 𝒑 𝑡 scales the combination of inputs and previous information. This
signal is then combined with the output of the input gate (𝒊 𝑡 ). The new hidden state is
an attenuation of the signal coming from the output gate. Finally, the prediction is a
linear combination of hidden states. Figure 4.5 illustrates how the information flows
in a LSTM cell.
Algorithm Mathematically, RNNs can be defined by the following algorithm:
1. Initiate with 𝒄0 = 0 and 𝑯 0 = 0.
2. Given the input 𝑿 𝑡 , for 𝑡 ∈ {1, . . . ,𝑇 }, do:
𝒇𝑡 = Logistic(𝑾 𝑓 𝑿 𝑡 + 𝑼 𝑓 𝑯 𝑡−1 + 𝒃 𝑓 )
𝒊𝑡 = Logistic(𝑾 𝑖 𝑿 𝑡 + 𝑼 𝑖 𝑯 𝑡−1 + 𝒃 𝑖 )
𝒐𝑡 = Logistic(𝑾 𝑜 𝑿 𝑡 + 𝑼 𝑜 𝑯 𝑡−1 + 𝒃 𝑜 )
𝒑𝑡 = Tanh(𝑾 𝑐 𝑿 𝑡 + 𝑼 𝑐 𝑯 𝑡−1 + 𝒃 𝑐 )
𝒄𝑡 = ( 𝒇 𝑡 ⊙ 𝒄 𝑡−1 ) + (𝒊 𝑡 ⊙ 𝒑 𝑡 )
𝑯𝑡 = 𝒐 𝑡 ⊙ Tanh(𝒄 𝑡 )
𝑌b𝑡+ℎ |𝑡 = 𝑾 𝑦 𝑯 𝑡 + 𝑏 𝑦
where 𝑼 𝑓 , 𝑼 𝑖 , 𝑼 𝑜 ,𝑼 𝑐 ,𝑼 𝑓 , 𝑾 𝑓 , 𝑾 𝑖 , 𝑾 𝑜 , 𝑾 𝑐 , 𝑾 𝑦 , 𝒃 𝑓 , 𝒃 𝑖 , 𝒃 𝑜 , 𝒃 𝑐 , and 𝑏 𝑦 are

parameters to be estimated.
138 Medeiros
CELL STATE
X +
X
X
X +
X
X
FORGET GATE
X +
X
X
INPUT GATE
X +
X
X
OUTPUT GATE
Fig. 4.4: Architecture of the Long-Short-Term Memory Cell (LSTM)

Fig. 4.5: Information flow in a LTSM Cell
4.5.3 Convolution Neural Networks
Convolutional Neural Networks (CNNs) are a class of Neural Network models that
have proven to be very successful in areas such as image recognition and classification
and are becoming popular for time series forecasting.
Figure 4.6 illustrates graphically a typical CNN. It is easier to understand the
architecture of a CNN through an image processing application.
Fig. 4.6: Representation of a Convolution Neural Network
As can be seen in Figure 4.6, the CNN consist of two main blocks: the feature
extraction and the prediction blocks. The prediction block is a feedforward deep NN
as previously discussed. The feature extraction block has the following key elements:
140 Medeiros
one or more convolutional layer; a nonlinear transformation of the data; one or more
pooling layers for dimension reduction; and a fully-connected (deep) feed-forward
neural network.
The elements above are organized in a sequence of layers:
convolution + nonlinear transformation → pooling → convolution + nonlinear
transformation → pooling → · · · → convolution + nonlinear transformation →
pooling → Fully-connected (deep) NN
To a computer, an image is matrix of pixels. Each entry of the matrix is the intensity
of the pixel: 0 − 255. The dimension of the matrix is the resolution of the image.
For coloured images, there is a third dimension to represent the colour channels:
red, green, and blue. Therefore, the image is a three dimensional matrix (tensor):
Height × Width × 3. An image kernel is a small matrix used to apply effects (filters),
such as blurring, sharpening, outlining, for example. In CNNs, kernels are used for
feature extraction, a technique for determining the most important portions of an
image. In this context the process is referred to more generally as convolution.
The convolution layer is defined as follows. Let 𝑿 ∈ R 𝑀×𝑁 be the input data and
𝑾 ∈ R𝑄×𝑅 the filter kernel.
For 𝑖 = 1, . . . 𝑀 − 𝑄 + 1, 𝑗 = 1, . . . , 𝑁 − 𝑅 + 1 write:
𝑄 ∑︁
∑︁ 𝑅
𝑂𝑖 𝑗 = [𝑾 ⊙ [ 𝑿] 𝑖:𝑖+𝑄−1, 𝑗: 𝑗+𝑅−1 ] 𝑞,𝑟
𝑞=1 𝑟=1 (4.34)
′
= 𝜾𝑄 𝑾⊙ [ 𝑿] 𝑖:𝑖+𝑄−1, 𝑗: 𝑗+𝑅−1 𝜾𝑅
where ⊙ is the element-by-element multiplication, 𝜾𝑄 ∈ R𝑄 and 𝜾𝑅 𝑅 are vector of ones,
[ 𝑿] 𝑖:𝑖+𝑄−1, 𝑗: 𝑗+𝑅−1 is the block of the matrix 𝑋 running from row 𝑖 to row 𝑖 + 𝑄 − 𝑗
and from column 𝑗 to column 𝑗 + 𝑅 − 1, and [ 𝑿] 𝑖:𝑖+𝑄−1, 𝑗: 𝑗+𝑅−1 ] 𝑞,𝑟 is the element of
[ 𝑿] 𝑖:𝑖+𝑄−1, 𝑗: 𝑗+𝑅−1 in position (𝑞, 𝑟).
𝑂 𝑖 𝑗 is the discrete convolution between 𝑾 and [ 𝑿] 𝑖:𝑖+𝑄−1, 𝑗: 𝑗+𝑅−1 :
𝑂 𝑖 𝑗 = 𝑾 ∗ [ 𝑿] 𝑖:𝑖+𝑄−1, 𝑗: 𝑗+𝑅−1 .
Figure 4.7 shows an example of the transformation implied by the convolution layer.
In this example a (3 × 3) filter is applied to (6 × 6) matrix of inputs. The output is
a (4 × 4) matrix. Each element of the output matrix is the sum of the dot product
between the entries of the input matrix (shaded red area) and the ones from the weight
matrix. Note that the shaded red (3 × 3) matrix is slided to the right and down by one
entry. Sometimes, in order to reduce the dimension of the output, one can apply the
stride technique, i.e., slide over more than one entry of the input matrix. Figure 4.8
shows an example.
Due to border effects, the output of the convolution layer is of smaller dimension
than the input. One solution is to use the technique called padding, i.e., filling a
border with zeroes. Figure 4.9 illustrates the idea. Note that in the case presented in
the figure the input and the ouput matrices have the same dimension.
Fig. 4.7: Example of a convolution layer

142 Medeiros
Fig. 4.8: Example of a convolution layer with stride
Each convolution layer may have more than one convolution filter. Figure 4.10
shows a convolution layer where the input is formed by three (6 × 6) matrices and
where there are two filters. The output of the layer is a set of two (4 × 4) matrices. In
this case stride is equal to one and there is no padding.
Usually the outputs of the convolution layer are sent through a nonlinear activation
function as, for example, the ReLU. See Figure 4.11 for an illustration.
The final step is the application of a dimension reduction technique called pooling.
One common polling approach is the max pooling, where the final output is the
maximum entry in a sub-matrix of the output of the convolution layer. See, for
example, the illustration in Figure 4.12
The process describe above is than repeated as many times as the number of
convolution layers in the network.
Summarizing, the user has to define the following hyperparameters concerning
the architecture of the convolution NN:
1. number of convolution layers (𝐶);
2. number of pooling layers (𝑃);
3. number (𝐾𝑐 ) and dimensions (𝑄 𝑐 height, 𝑅𝑐 width and 𝑆 𝑐 depth) of filters in
each convolution layer 𝑐 = 1, . . . , 𝐶;
4. architecture of the deep neural network.
The parameters to be estimated are
1. Filter weights: 𝑾 𝑖𝑐 ∈ R𝑄𝑐 ×𝑅𝑐 ×𝑆𝑐 , 𝑖 = 1, . . . , 𝐾𝑐 , 𝑐 = 1, . . . , 𝐶;
2. ReLU biases: 𝜸 𝑐 ∈ R𝐾𝑐 , 𝑐 = 1, . . . , 𝐶;
3. All the parameters of the fully connected deep.
0 0 0 0 0 0 0 0
0 3 1 1 2 8 4 0 3
0 1 0 7 3 2 6 0 1 0 -1
0
0
2
1
3
4
5
1
1
2
1
6
3
5
0
0 * 1 0 -1
1 0 -1
=
0 3 2 1 3 7 2 0
0 9 2 6 2 5 1 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 3 1 1 2 8 4 0 3 -4
0 1 0 7 3 2 6 0 1 0 -1
0
0
2
1
3
4
5
1
1
2
1
6
3
5
0
0 * 1 0 -1
1 0 -1
=
0 3 2 1 3 7 2 0
0 9 2 6 2 5 1 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 3 1 1 2 8 4 0 3 -4 -2
0 1 0 7 3 2 6 0 1 0 -1
0
0
2
1
3
4
5
1
1
2
1
6
3
5
0
0 * 1 0 -1
1 0 -1
=
0 3 2 1 3 7 2 0
0 9 2 6 2 5 1 0
0 0 0 0 0 0 0 0
Fig. 4.9: Example of an output of a convolution layer with padding

144 Medeiros
* =
Filter: 3 x 3 x 3
4x4x1
Input data: 6 x 6 x 3
* = Output: 4 x 4 x 2
Filter: 3 x 3 x 3
4x4x1
Input data: 6 x 6 x 3
Number of filters
Fig. 4.10: Example of a convolution layer with two convolution filters
* = ReLU +g01
* = ReLU +g02
Fig. 4.11: Example of a convolution layer with two convolution filters and nonlinear
transformation
3 1 2 4 Max-pooling 3 1 2 4 Max-pooling
1 7 3 6 Stride = 2 7 1 7 3 6 Stride = 2 7 6
2 5 1 3 2 5 1 3
9 6 2 1 9 6 2 1
3 1 2 4 Max-pooling 3 1 2 4 Max-pooling
1 7 3 6 Stride = 2 7 6 1 7 3 6 Stride = 2 7 6
2 5 1 3 9 2 5 1 3 9 3
9 6 2 1 9 6 2 1
Fig. 4.12: Example of a convolution layer with two convolution filters and nonlinear
transformation
4.5.4 Autoenconders: Nonlinear Factor Regression
Autoencoders are the primary model for dimension reduction in the ML literature.
They can be interpreted as a nonlinear equivalents to PCA. An autoenconder is a
special type of deep neural network in which the outputs attempt to approximate
the input variables. The input variables pass through neurons in the hidden layer(s),
creating a compressed representation of the input variables. This compressed input is
decoded (decompressed) into the output layer. The layer of interest is the hidden layer
with the smallest numbers of neurons, since the neurons in this layer represent the
latent non-linear factors that we aim to extract. To illustrate the basic structure of an
autoencoder, Figure 4.13 illustrates an autoencoder consisting of five inputs and three
hidden layers with four, one and four neurons, respectively. The second hidden layer
in the diagram represents the latent single factor we wish to extract, 𝑂 21 . The layer
preceding it is the encoding layer while the layer that follows it is the decoding layer.
As other deep neural networks, autoencoders can be written using the same
recursive formulas as before. The estimated non-linear factors can serve as inputs for
linear and nonlinear forecasting models as the ones described in this Chapter or in
Chapters 1 and 2.
4.5.5 Hybrid Models
Recently, Medeiros and Mendes (2013) proposed the combination of LASSO-based

estimation and NN models. The idea is to construct a feedforward single-hidden layer
NN where the parameters of the nonlinear terms (neurons) are randomly generated
and the linear parameters are estimated by LASSO (or one of its generalizations).
146 Medeiros
Input Hidden 1 Hidden 2 Hidden 3 Output
𝑋1 𝑋ˆ 1
𝑂1(1) 𝑂1(3)
𝑋2 𝑋ˆ 2
𝑂2(1) 𝑂2(3)
𝑋3 𝑂1(2) 𝑋ˆ 3
𝑂3(1) 𝑂3(3)
𝑋4 𝑋ˆ 4
𝑂4(1) 𝑂4(3)
𝑋5 𝑋ˆ 5
Fig. 4.13: Graphical representation of an Autoecoder
Similar ideas were also considered by Kock and Teräsvirta (2014) and Kock and
Teräsvirta (2015).
Trapletti, Leisch and Hornik (2000) and Medeiros, Teräsvirta and Rech (2006)
proposed to augment a feedforward shallow NN by a linear term. The motivation is
that the nonlinear component should capture only the nonlinear dependence, making
the model more interpretable. This is in the same spirit of the semi-parametric models
considered in Chen (2007).
Inspired by the above ideas, Medeiros et al. (2021) proposed combining random
forests with adaLASSO and OLS. The authors considered two specifications. In the
first one, called RF/OLS, the idea is to use the variables selected by a Random Forest
in a OLS regression. The second approach, named adaLASSO/RF, works in the
opposite direction. First select the variables by adaLASSO and than use them in a
Random Forest model. The goal is to disentangle the relative importance of variable
selection and nonlinearity to forecast inflation.
In this chapter we review the most recent advances in using Machine Learning
models/methods to forecast time-series data in a high-dimensional setup, where the
number of variables used as potential predictors is much larger than the available
sample to estimate the forecasting models.
We start the chapter by discussing how to construct and compare forecasts from
different models. More specifically, we discuss the Diebold-Mariano test of equal
predictive ability and the Li-Liao-Quaedvlieg test of conditional superior predictive
ability. Finally, we illustrate how to construct model confidence sets.
References 147
In terms of linear ML models, we complement the techniques described in Chapter

1 by focusing of factor-based regression, the combination of factors and penalized
regressions and ensemble methods.
After presenting the linear models, we review neural network methods. We discuss
both shallow and deep networks, as well as long shot term memory and convolution
neural networks.
We end the chapter by discussing some hybrid methods and new proposals in the
forecasting literature.
Acknowledgements The author wishes to acknowledge Marcelo Fernandes and Eduardo Mendes
as well as the editors, Felix Chan and László Mátyás, for insightful comments and guidance.
References
Ahn, S. & Horenstein, A. (2013). Eigenvalue ratio test for the number of factors.
Econometrica, 81, 1203–1227.
Bai, J. & Ng, S. (2002). Determining the number of factors in approximate factor
models. Econometrica, 70, 191–221.
Barron, A. (1993). Universal approximation bounds for superpositions of a sigmoidal
function. IEEE Transactions on Information Theory, 39, 930–945.
Breiman, L. (1996). Bagging predictors. Machine Learning, 24, 123–140.
Chen, X. (2007). Large sample sieve estimation of semi-nonparametric models. In
J. Heckman & E. Leamer (Eds.), Handbook of econometrics. Elsevier.
Clark, T. & McCracken, M. (2013). Advances in forecast evaluation. In G. Elliott &
A. Timmermann (Eds.), Handbook of economic forecasting (Vol. 2, p. 1107-
1201). Elsevier.
Cybenko, G. (1989). Approximation by superposition of sigmoidal functions.
Mathematics of Control, Signals, and Systems, 2, 303–314.
Diebold, F. (2015). Comparing predictive accuracy, twenty years later: A personal
perspective on the use and abuse of Diebold-Mariano tests. Journal of Business
and Economic Statistics, 33, 1–9.
Diebold, F. & Mariano, R. (1995). Comparing predictive accuracy. Journal of
Business and Economic Statistics, 13, 253–263.
Elliott, G., Gargano, A. & Timmermann, A. (2013). Complete subset regressions.
Journal of Econometrics, 177(2), 357–373.
Elliott, G., Gargano, A. & Timmermann, A. (2015). Complete subset regressions
with large-dimensional sets of predictors. Journal of Economic Dynamics and
Control, 54, 86–110.
Fan, J., Masini, R. & Medeiros, M. (2021). Bridging factor and sparse models (Tech.
Rep. No. 2102.11341). arxiv.
Fava, B. & Lopes, H. (2020). The illusion of the illusion of sparsity. Brazilian
Journal of Probability and Statistics. (forthcoming)
148 Medeiros
Foresee, F. D. & Hagan, M. . T. (1997). Gauss-newton approximation to Bayesian

regularization. In IEEE international conference on neural networks (vol. 3)
(pp. 1930–1935). New York: IEEE.
Funahashi, K. (1989). On the approximate realization of continuous mappings by
neural networks. Neural Networks, 2, 183–192.
Garcia, M., Medeiros, M. & Vasconcelos, G. (2017). Real-time inflation forecasting
with high-dimensional models: The case of brazil. International Journal of
Forecasting, 33(3), 679–693.
Genre, V., Kenny, G., Meyler, A. & Timmermann, A. (2013). Combining expert
forecasts: Can anything beat the simple average? International Journal of
Forecasting, 29, 108–121.
Giacomini, R. & White, H. (2006). Tests of conditional predictive ability. Economet-
rica, 74, 1545–1578.
Giannone, D., Lenza, M. & Primiceri, G. (2021). Economic predictions with big
data: The illusion of sparsity. Econometrica, 89, 2409–2437.
Grenander, U. (1981). Abstract inference. New York, USA: Wiley.
Hansen, P., Lunde, A. & Nason, J. (2011). The model confidence set. Econometrica,
79, 453–497.
Harvey, D., Leybourne, S. & Newbold, P. (1997). Testing the equality of prediction
mean squared errors. International Journal of Forecasting, 13, 281–291.
Hochreiter, S. & Schmidhuber, J. (1997). Long short-term memory. Neural
Computation, 9, 1735–1780.
Hornik, K., Stinchombe, M. & White, H. (1989). Multi-layer Feedforward networks
are universal approximators. Neural Networks, 2, 359–366.
Inoue, A. & Kilian, L. (2008). How useful is bagging in forecasting economic time
series? a case study of U.S. consumer price inflation. Journal of the American
Statistical Association, 103, 511-522.
Kock, A. & Teräsvirta, T. (2014). Forecasting performance of three automated
modelling techniques during the economic crisis 2007-2009. International
Journal of Forecasting, 30, 616–631.
Kock, A. & Teräsvirta, T. (2015). Forecasting macroeconomic variables using neural
network models and three automated model selection techniques. Econometric
Reviews, 35, 1753–1779.
Li, J., Liao, Z. & Quaedvlieg, R. (2021). Conditional superior predictive ability.
Review of Economic Studies. (forthcoming)
MacKay, D. J. C. (1992a). Bayesian interpolation. Neural Computation, 4, 415–447.
MacKay, D. J. C. (1992b). A practical Bayesian framework for backpropagation
networks. Neural Computation, 4, 448–472.
McAleer, M. & Medeiros, M. (2008). A multiple regime smooth transition hetero-
geneous autoregressive model for long memory and asymmetries. Journal of
Econometrics, 147, 104–119.
McCracken, M. (2020). Diverging tests of equal predictive ability. Econometrica,
88, 1753–1754.
Medeiros, M. & Mendes, E. (2013). Penalized estimation of semi-parametric additive
time-series models. In N. Haldrup, M. Meitz & P. Saikkonen (Eds.), Essays in
References 149
nonlinear time series econometrics. Oxford University Press.

Medeiros, M., Teräsvirta, T. & Rech, G. (2006). Building neural network models for
time series: A statistical approach. Journal of Forecasting, 25, 49–75.
Medeiros, M., Vasconcelos, G., Veiga, A. & Zilberman, E. (2021). Forecasting
inflation in a data-rich environment: The benefits of machine learning methods.
Journal of Business and Economic Statistics, 39, 98–119.
Mhaska, H., Liao, Q. & Poggio, T. (2017). When and why are deep networks
better than shallow ones? In Proceedings of the thirty-first aaai conference on
artificial intelligence (aaai-17) (pp. 2343–2349).
Onatski, A. (2010). Determining the number of factors from empirical distribution
of eigenvalues. Review of Economics and Statistics, 92, 1004–1016.
Park, J. & Sandberg, I. (1991). Universal approximation using radial-basis-function
networks. Neural Computation, 3, 246–257.
Patton, A. (2015). Comment. Journal of Business & Economic Statistics, 33, 22-24.
Samuel, A. (1959). Some studies in machine learning using the game of checkers.
IBM Journal of Research and Development, 3.3, 210–229.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. (2014).
Simple way to prevent neural networks from overfitting. Journal of Machine
Learning Research, 15, 1929–1958.
Stinchcombe, M. & White, S. (1989). Universal approximation using feedforward
neural networks with non-sigmoid hidden layer activation functions. In
Proceedings of the international joint conference on neural networks (pp.
613–617). Washington: IEEE Press, New York, NY.
Stock, J. & Watson, M. (2002a). Forecasting using principal components from a
large number of predictors. Journal of the American Statistical Association,
97, 1167–1179.
Stock, J. & Watson, M. (2002b). Macroeconomic forecasting using diffusion indexes.
Journal of Business & Economic Statistics, 20, 147–162.
Suarez-Fariñas, Pedreira, C. & Medeiros, M. (2004). Local-global neural networks:
A new approach for nonlinear time series modelling. Journal of the American
Statistical Association, 99, 1092–1107.
Teräsvirta, T. (1994). Specification, estimation, and evaluation of smooth transition
autoregressive models. Journal of the American Statistical Association, 89,
208–218.
Teräsvirta, T. (2006). Forecasting economic variables with nonlinear models. In
G. Elliott, C. Granger & A. Timmermann (Eds.), (Vol. 1, p. 413-457). Elsevier.
Teräsvirta, T., Tjøstheim, D. & Granger, C. (2010). Modelling nonlinear economic
time series. Oxford, UK: Oxford University Press.
Trapletti, A., Leisch, F. & Hornik, K. (2000). Stationary and integrated autoregressive
neural network processes. Neural Computation, 12, 2427–2450.
West, K. (2006). Forecast evaluation. In G. Elliott, C. Granger & A. Timmermann
(Eds.), (Vol. 1, pp. 99–134). Elsevier.
Yarotsky, D. (2017). Error bounds for approximations with deep ReLU networks.
Neural Networks, 94, 103–114.
Chapter 5
Causal Estimation of Treatment Effects From
Observational Health Care Data Using Machine
Learning Methods
William Crown
Abstract The econometrics literature has generally approached problems of causal

inference from the perspective of obtaining an unbiased estimate of a parameter in
a structural equation model. This requires strong assumptions about the functional
form of the model and data distributions. As described in Chapter 3, there is a
rapidly growing literature that has used machine learning to estimate causal effects.
Machine learning models generally require far fewer assumptions. Traditionally, the
identification of causal effects in econometric models rests on theoretically justified
controls for observed and unobserved confounders. The high dimensionality of
many datasets offers the potential for using machine learning to uncover potential
instruments and expand the set of observable controls. Health care is an example of
high dimensional data where there are many causal inference problems of interest.
Epidemiologists have generally approached such problems using propensity score
matching or inverse probability treatment weighting within a potential outcomes
framework. This approach still focuses on the estimation of a parameter in a structural
model. A more recent method, known as doubly robust estimation, uses mean
differences in predictions versus their counterfactual that have been updated by
exposure probabilities. Targeted maximum likelihood estimators (TMLE) optimize
these methods. TMLE methods are not, inherently, machine learning methods.
However, because the treatment effect estimator is based on mean differences in
individual predictions of outcomes for those treated versus the counterfactual, super
learning machine learning approaches have superior performance relative to traditional
methods. In this chapter, we begin with the same assumption of selection of observable
variables within a potential outcomes framework. We briefly review the estimation
of treatment effects using inverse probability treatment weights and doubly robust
estimators. These sections provide the building blocks for the discussion of TMLE
methods and their estimation using super learner methods. Finally, we consider the
extension of the TMLE estimator to include instrumental variables in order to control
for bias from unobserved variables correlated with both treatment and outcomes.
William Crown B
Brandeis University, Waltham, Massachusetts, USA, e-mail: wcrown@brandeis.edu
152 Crown
5.1 Introduction
Several aspects of the changing healthcare data landscape including the rapid growth in
the volume of healthcare data, the fact that much of it is unstructured, the ability to link
different types of data together (claims, EHR, sociodemographics, genomics), and the
speed with which the data are being refreshed create serious challenges for traditional
statistical methods from epidemiology and econometrics while, simultaneously,
creating opportunities for the use of machine learning methods.
Methods such as logistic regression have long been used to predict whether a
patient is at risk of developing a disease or having a health event such as a heart
attack—potentially enabling intervention before adverse outcomes occur. Such models
are rapidly being updated with machine learning methods such as lasso, random
forest, support vector machines, and neural network models to predict such outcomes
as hospitalization (Hong, Haimovich & Taylor, 2018; Futoma, Morris & Lucas, 2015;
Shickel, Tighe, Bihorac & Rashidi, 2018; Rajkomar et al., 2018) or the onset of
disease (Yu et al., 2010). These algorithms that offer the potential to improve the
sensitivity and specificity of predictions used in health care operations and to guide
clinical care (Obermeyer & Emanuel, 2016). In some areas of medicine, such as
radiology, machine learning methods show great promise for improved diagnostic
accuracy (Obermeyer & Emanuel, 2016; Ting et al., 2017). However, the applications
of machine learning in health care have been almost exclusively about prediction;
rarely, are machine learning methods used for causal inference.
5.2 Naïve Estimation of Causal Effects in Outcomes Models with

Binary Treatment Variables
Everything else equal, randomized trials are the strongest design for estimating
unbiased treatment effects because when subjects are randomized to treatment, the
alternative treatment groups are asymptotically ensured to balance on the basis of both
observed and unobserved covariates. In the absence of randomization, economists
often use quasi-experimental designs that attempt to estimate the average treatment
effect (ATE) that one would have attained from a randomized trial on the same patient
group to answer the same question. Randomized controlled trials highlight the fact
that, from a conceptual standpoint, it is important to consider the estimation problem
in the context of both observable and unobservable variables.
Suppose we wish to estimate:
𝑌 = 𝐵0 + 𝐵1𝑇 + 𝐵2 𝑋 + 𝐵3𝑈 + 𝑒,
where 𝑌 is a health outcome of interest, 𝑇 is a treatment variable, 𝑋 is a matrix of

observed covariates, 𝑈 is a matrix of unobserved variables, 𝑒 is vector of residuals, and
𝐵0 , 𝐵1 , 𝐵2 , and 𝐵3 are parameters, or parameter vectors to be estimated. Theoretically,
randomization eliminates the correlation between any covariate and treatment. This is
5 Causal Estimation of Treatment Effects from Observational Health Care Data 153
important because any unobserved variable that is correlated both with treatment and
with outcomes will introduce a correlation between treatment and the residuals. This,
by definition, introduces bias (Wooldridge, 2002). In health economic evaluations,
the primary goal is to obtain unbiased and efficient estimates of the treatment effect,
𝐵1 . However, the statistical properties of 𝐵1 may be influenced by a variety of factors
that may introduce correlation between the treatment variable and the residuals.
Consider the matrix of unobserved variables 𝑈. If 𝑈 has no correlation with 𝑇,
then its omission from the equation will have no effect on the bias of the estimate of
𝐵1 . However, if 𝑐𝑜𝑣(𝑇,𝑈) is not equal to zero, then
𝐸 (𝐵1 ) = 𝐵1 + 𝐵3 𝑐𝑜𝑣(𝑇,𝑈),
where 𝐸 (𝐵1 ) is the expected value of 𝐵1 . In other words, the estimator for treatment
effect will be biased by the amount of 𝐵3 𝑐𝑜𝑣(𝑇,𝑈). In brief, a necessary condition for
obtaining unbiased estimates in any health economic evaluation using observational
data is the inclusion of strong measures on all important variables hypothesized
to be correlated with both the outcome and the treatment (i.e., measurement of all
important confounders noted earlier).
Leaving aside the issue of unobserved variables for the moment, one difficulty
with the standard regression approach is that there is nothing in the estimation method
that assures that the groups are, in fact, comparable. In particular, there may be
subpopulations of patients in the group receiving the intervention for whom there is
no overlap with the covariate distributions of patients in the comparison group and
vice versa. This concept is known as positivity or lack of common support. Positivity
requires that, for each value of X in the treated group, the probability of observing 𝑋
in the comparison group is positive. In the absence of positivity or common support,
the absence of bias in the treatment effect estimate is possible only if the estimated
functional relationship between the outcome 𝑌 and the covariate matrix 𝑋 holds for
values outside of the range of common support (Jones & Rice, 2009). Crump, Hotz,
Imbens and Mitnik (2009) point out that, even with some evidence of positivity, bias
and variance of treatment effect estimates can be very sensitive to the functional form
of the regression model.
One approach to this problem is to match patients in the intervention group with
patients in the comparison group who have the exact same pattern of observable
covariates (e.g., age, gender, race, medical comorbidities). However, in practice this
exact match approach requires enormous sample sizes—particularly when the match
is conducted on numerous strata. Instead, Rosenbaum and Rubin (1983) proposed
matching on the propensity score. Propensity score methods attempt to control for
distributional differences in observed variables in the treatment cohorts so that the
groups being compared are at least comparable on observed covariates.
154 Crown
5.3 Is Machine Learning Compatible with Causal Inference?
A number of papers have reviewed the role of machine learning for estimating
treatment effects (e.g., Athey & Imbens, 2019; Knaus, Lechner & Strittmatter, 2021).
Some machine learning approaches use regression-based methods for prediction. For
example, Lasso, Ridge, and elastic net methods utilize correction factors to reduce the
risk of over fitting (Hastie, Tibshirani & Friedman, 2009; Tibshirani, 1996). However,
as noted by Mullainathan and Spiess (2017) and discussed in Chapter 3, causal
interpretation should not be given to the parameters of such models. Unfortunately,
there is nothing magical about machine learning that protects against the usual
challenges encountered in drawing causal inferences in observational data analysis. In
particular, just because machine learning methods are operating on high-dimensional
data does not protect against bias. Increasing sample size—for example, assembling
more and more and more medical claims data for an outcomes study–does not correct
the problem of bias if the dataset is lacking in key clinical severity measures such as
cancer stage in a model of breast cancer outcomes (Crown, 2015). Moreover, even in
samples with very large numbers of observations such as medical claims or electronic
medical record datasets, models with very large numbers of variables become sparse
in high dimensional space. This can lead to a breakdown in assumptions of positivity
for example. Perhaps most importantly, machine learning can provide a statistical
method for estimating causal effects but only in the context of an appropriate causal
framework. The use of machine learning without a causal framework is fraught.
On the other hand, economists, epidemiologists, and health services researchers
have been trained that they must have a theory that they test through model estimation.
The major limitation of this approach is that it makes it very difficult to escape
from the confines of what we already know (or think we know). Machine learning
methods such as Lasso can provide a manner for systematically selecting variables
to be included in the model, as well as exploring alternative functional forms for
the outcome equation. The features thus identified then can become candidates for
inclusion in a causal modeling framework that is estimated using standard econometric
methods—preferably using a different set of observations than those used to identify
the features. However, as discussed in Chapter 3, parameter estimates from machine
learning models cannot be assumed to be unbiased estimates of causal parameters in
a structural equation.
Machine learning methods may also help in the estimation of traditional eco-
nometric or epidemiologic models using propensity score or inverse probability
treatment weights. Machine learning can aid in the estimation of causal models in
other ways as well. An early application of machine learning for causal inference in
health economics and outcomes research (HEOR) was to fit propensity score (PS)
models to create matched samples and estimate average treatment effects (ATEs) by
comparing these samples (Westreich, Lessler & Jonsson Funk, 2010; Rosenbaum &
Rubin, 1983). While traditionally, logistic regression was used to fit PS models, it
has been shown that ‘off the shelf’ ML methods for prediction and classification (e.g.,
random forests, classification and regression trees, Lasso) are sometimes more flexible
and can lead to lower bias in treatment effect estimates (Setoguchi, Schneeweiss,
Brookhart, Glynn & Cook, 2008). However, these approaches in themselves are
imperfect as they are tailored to minimize root mean square error (RMSE) as opposed
to targeting the causal parameter. Some extensions of these methods have addressed
specific challenges of using the PS for confounding adjustment, by customizing the
loss function of the ML algorithms (e.g., instead of minimizing classification error, to
maximize balance in the matched samples) (Rosenbaum & Rubin, 1983). However,
the issue remains that giving equal importance to many covariates when creating
balance may not actually minimize bias (e.g., if many of the covariates are only weak
confounders). It is recommended that balance on variables that are thought to be
the most prognostic to the outcome should be prioritized; however, this ultimately
requires subjective judgement (Ramsahai, Grieve & Sekhon, 2011).
Finally, machine learning methods can be used to estimate causal treatment effects
directly. As discussed in Chapter 3, Athey and Imbens (2016), Wager and Athey (2018),
and Athey, Tibshirani and Wager (2019) propose the use of tree-based approaches
to provide a step-function approximation to the outcome equation. These methods
place less emphasis on parametric estimation of treatment effects. Chapter 3 also
discusses the estimation of double debiased estimators (Belloni, Chen, Chernozhukov
& Hansen, 2012; Belloni, Chernozhukov & Hansen, 2013; Belloni, Chernozhukov &
Hansen, 2014a; Belloni, Chernozhukov & Hansen, 2014b; Belloni, Chernozhukov,
Fernndez-Val & Hansen, 2017; Chernozhukov et al., 2017; and Chernozhukov et al.,
2018).
In this chapter, we discuss the estimation of Targeted Maximum Likelihood
Estimation (TMLE) models (Schuler & Rose, 2017; van der Laan & Rose, 2011).
TMLE is not, itself, a machine learning method. However, the fact that TMLE uses
the potential outcomes framework and bases estimates of ATE upon predictions of
outcomes and exposures makes it a natural approach to implement using machine
learning techniques—particularly, super learner methods. This has many potential
advantages including reducing the necessity of correctly chosing the specification of
the outcome model and reducing the need to make strong distributional assumptions
about the data.
5.4 The Potential Outcomes Model
Imbens (2020) provides a comprehensive review and comparison of two major causal
inference frameworks with application for health economics—(1) directed acyclic
graphs (DAGs) (Pearl, 2009) and (2) potential outcomes (Rubin, 2006). Imbens
(2020) notes that the potential outcomes framework has been more widely used
in economics but that DAGs can be helpful in clarifying the assumptions made in
the analysis such as the role of ‘back door’ and ‘front door’ criteria in identifying
treatment effects. Richardson (2013) introduce Single World Intervention Graphs
(SWIGs) as a framework for unifying the DAG and potential outcomes approaches to
causality. Dahabreh, Robertson, Tchetgen and Stuart (2019) use SWIGs to examine
156 Crown
the conditions under which it is possible to generalize the results of a randomized

trial to a target population of trial-eligible individuals.
The potential outcomes framework has mainly focused upon the estimation of
average effects of binary treatments. It has made considerable progress not only on
questions related to the identification of treatment effects but also problems of study
design, estimation, and inference (Imbens, 2020). For these reasons, the remainder
of this chapter will focus on the estimation of causal effects using the potential
outcomes framework. In particular, we will focus on the building blocks for TMLE
which, as mentioned above, is a statistical technique that can be implemented within
the potential outcomes framework. Because TMLE estimates ATE as the average
difference in predicted outcomes for individuals exposed to an intervention relative to
their counterfactual outcome, it lends itself to implementation using machine learning
methods.
Potential outcomes. In observational studies we observe outcomes only for the
treatment that individuals receive. There is a potential outcome that could be observed
if that individual was exposed to an alternative treatment but this potential outcome is
not available. A straightforward estimate of treatment effect is the expected difference
between the outcome that an individual had for the treatment received versus the
outcome that they would have had if exposed to an alternative treatment 𝐸 [𝑌1 −𝑌0 ]
The most straightforward estimate of this ATE is the parameter estimate for a
binary treatment variable in a regression model. It is also possible to estimate the
conditional average treatment effect (CATE) given a vector of patient attributes 𝑋.
We assume that the observed data are n independent and identically distributed copies
of 𝑂 = (𝑍,𝑇,𝑌 , 𝑋) ∼ 𝑃0 , where Y is a vector of observed outcomes, 𝑇 indicates
the treatment group, 𝑋 is a matrix of observed control variables, 𝑍 is a matrix of
instruments for unobserved variables correlated with both outcomes and treatment
assignment, and 𝑃0 is the true underlying distribution from which the data are drawn.
If there are no unobserved confounders to generate endogeneity bias 𝑍 is not needed.
Assumptions needed for causal inference with observational data. Drawing causal
inference in observational studies requires several assumptions (Robins, 1986; van der
Laan & Rose, 2011; van der Laan & Rubin, 2006; Rubin, 1974). The first of
these—the Stable Unit Value Assumption (SUTVA)—is actually a combination of
several assumptions. It states that (1) an individual’s potential outcome under his or
her observed exposure history is the outcome that will actually be observed for that
person (also known as consistency) (Cole & Frangakis, 2009), (2) the exposure of any
given individual does not affect the potential outcomes of any other individuals (also
known as non-interference) and (3) the exposure level is the same for all exposed
individuals (Rubin, 1980; Rubin, 1986; Cole & Hernán, 2008).
In addition, causal inference with observational data requires an assumption that
there are no unmeasured confounders. That is, all common causes of both the exposure
and the outcome have been measured (Greenland & Robins, 1986) and the exposure
mechanism and potential outcomes are independent after conditioning on the set
of covariates. Unmeasured confounders are a common source of violation of the
assumption of exchangeability of treatments (Hernán, 2011). When results from
RCTs and observational studies have been found to differ it is often assumed that this
is due to the failure of observational studies to adequately control for unmeasured
confounders.
Finally, there is the assumption of positivity which states that, within a given
strata of X, every individual has a nonzero probability of receiving either exposure
condition; this is formalized as 0 < 𝑃( 𝐴 = 1|𝑋) < 1 for a binary exposure (Westreich
& Cole, 2010; Petersen, Porter, Gruber, Wang & Laan, 2012). If the positivity
assumption is violated, causal effects will not be identifiable (Petersen et al., 2012).
5.5 Modeling the Treatment Exposure Mechanism–Propensity

Score Matching and Inverse Probability Treatment Weights
One mechanism for testing the positivity assumption is to model treatment exposure
as a function of baseline covariates for the treated and comparison groups in
the potential outcomes model. Propensity score methods are widely used for this
purpose in empirical research. Operationally, propensity score methods begin with
the estimation of a model to generate the fitted probability, or propensity, to receive
the intervention versus comparison treatment. (The term ‘treatment’ is very broad
and can be anything from a pharmaceutical intervention to alternative models of
benefit design or organizing patient care such as accountable care organizations
[ACOs]). Observations that have a similar estimated propensity to be in either
the study group or comparison group will tend to have similar observed covariate
distributions (Rosenbaum & Rubin, 1983). Once the propensity scores have been
estimated, it is possible to pair patients receiving the treatment with patients in the
comparison group on the basis of having similar propensity scores. This is known as
propensity score matching. Thus, propensity score methods can be thought of as a
cohort balancing method undertaken prior to the application of traditional multivariate
methods (Brookhart et al., 2006; Johnson et al., 2006). More formally, Rosenbaum and
Rubin (1983) define the propensity score for subject 𝑖 as the conditional probability
of assignment to a treatment (𝑇 = 1) versus comparison (𝑇 = 0) given covariates, 𝑋:
𝑃𝑟 (𝑇 = 1|𝑋).
The validity of this approach assumes that there are no variables that influence
treatment selection for which we lack measures. Note that the propensity score model
of treatment assignment and the model of treatment outcomes share the common
assumption of strong ignorability. That is, both methods assume that any missing
variables are uncorrelated with both treatment and outcomes and can be safely
ignored.
One criticism of propensity score matching is that it is sometimes not possible to
identify matches for all the patients in the intervention group. This leads to lose of
sample size. Inverse probability treatment weighting (IPTW) retains the full sample
by using the propensity score to develop weights that, for each subject, are the inverse
158 Crown
of the predicted probability of the treatment that the subject received. This approach
gives more weight to subjects who have a lower probability of receiving a particular
treatment and less weight to those with a high probability of receiving the treatment.
Hirano, Imbens and Ridder (2003) propose using inverse probability treatment
weights to estimate average treatment effects:
𝑛
1 ∑︁ 𝑌𝑖 𝑇 𝑌𝑖 (1 − 𝑇)
𝐴𝑇 𝐸 = − .
𝑁 𝑖=0 𝑃𝑆 1 − 𝑃𝑆
IPTW is also useful for estimating more complex causal models such as marginal
structural models (Joffe, Have, Feldman & Kimmel, 2004). When the propensity
score model is correctly specified, IPTW can lead to efficient estimates of ATE under
a variety of data-generating processes. However, IPTW can generate biased estimates
of ATE if the propensity score model is mis-specified. This approach is similar to
weighting methodologies long used in survey research.
Intuitively, propensity score matching is very appealing because it forces an
assessment of the amount of overlap (common support) in the populations in the
study and comparison groups. There is a large, and rapidly growing, literature using
propensity score methods to estimate treatment effects with respect to safety and
health economic outcomes (Johnson, Crown, Martin, Dormuth & Siebert, 2009;
Mitra & Indurkhya, 2005). In application, there are a large number of methodologies
that can be used to define the criteria for what constitutes a ‘match’ (Baser, 2006;
Sekhon & Grieve, 2012).
5.6 Modeling Outcomes and Exposures: Doubly Robust Methods
Most of the medical outcomes literature using observational data to estimate ATE
focuses on the modeling of the causal effect of an intervention on outcomes by
balancing the comparison groups with propensity score matching or IPTW. Doubly
robust estimation is a combination of propensity score matching and covariate
adjustment using regression methods (Bang & Robins, 2005; Scharfstein, Rotnitzky
& Robins, 1999). Doubly robust estimators help to protect against bias in treatment
effects because the method is consistent when either the propensity score model or
the regression model is mis-specified (Robins, Rotnitzky & Zhao, 1994). See Robins
et al. (1994), Imbens and Wooldridge (2009) and Abadie and Cattaneo (2018) for
surveys.
After estimating the propensity score (PS) in the usual fashion the general
expressions for the doubly robust estimates in response to the presence or absence of
exposure (𝐷 𝑅1 ) and (𝐷 𝑅0 ), respectively are given by Funk et al. (2011):
𝑌𝑥=1 𝑋 𝑌ˆ1 (𝑋 − 𝑃𝑆)

𝐷 𝑅1 = −
𝑃𝑆 𝑃𝑆
𝑌𝑥=0 (1 − 𝑋) 𝑌ˆ0 (𝑋 − 𝑃𝑆)

𝐷 𝑅0 = − .
1 − 𝑃𝑆 1 − 𝑃𝑆
When 𝑋 = 1, 𝐷 𝑅1 and 𝐷 𝑅0 simplify to
𝑌𝑥=1 𝑌ˆ1 (1 − 𝑃𝑆)

𝐷 𝑅1 = −
𝑃𝑆 𝑃𝑆
and
𝐷 𝑅0 = 𝑌ˆ0 .
Similarly, when 𝑋 = 0, 𝐷 𝑅1 and 𝐷 𝑅0 simplify to
𝐷 𝑅1 = 𝑌ˆ1
and
𝑌𝑥=0 𝑌ˆ0 𝑃𝑆
𝐷 𝑅0 = + .
1 − 𝑃𝑆 1 − 𝑃𝑆
Note that for exposed individuals (where 𝑋 = 1), 𝐷 𝑅 is a function of their
observed outcomes under exposure and predicted outcomes under exposure given
covariates, weighted by a function of the 𝑃𝑆. The estimated value for 𝐷 𝑅0 is simply
the individuals’ predicted response, had they been unexposed based on the parameter
estimates from the outcome regression among the unexposed and the exposed
individuals’ covariate values (𝑍).
Similarly, for the unexposed (𝑌 = 0) 𝐷 𝑅0 is calculated as a function of the observed
response combined with the predicted response weighted by a function of the 𝑃𝑆,
while 𝐷 𝑅1 is simply the predicted response in the presence of exposure conditional
on covariates. Finally, the ATE is estimated as the difference between the means of
𝐷 𝑅1 and 𝐷 𝑅0 calculated across the entire study population.
With some algebraic manipulation the doubly robust estimator can be shown to
be the mean difference in response if everyone was either exposed or unexposed to
the intervention plus the product of two bias terms—one from the propensity score
model and one from the outcome model. If bias is zero from either the propensity
score model or the outcome model it will “zero out” any bias from the other model.
Recently the doubly robust literature has focused on the case with a relatively
large number of pretreatment variables (Chernozhukov et al., 2017; Athey, Imbens &
Wager, 2018; van der Laan & Rose, 2011; Shi, Blei & Veitch, 2019). Overlap issues in
covariate distributions, absent in all the DAG discussions, become prominent among
practical problems (Crump et al., 2009; D’Amour, Ding, Feller, Lei & Sekhon, 2021).
Basu, Polsky and Manning (2011) provide Monte Carlo evidence on the finite
sample performance of OLS, propensity score estimates, IPTW, and doubly robust
estimates. They find that no single estimator can be considered best for estimating
treatment effects under all data-generating processes for healthcare costs. IPTW
estimators are least likely to be biased across a range of data-generating processes
but can be biased if the propensity score model is mis-specified. IPTW estimators are
generally less efficient than regression estimators when the latter are unbiased. Doubly
robust estimators can be biased as a result of mis-specification of both the propensity
160 Crown
score model and the outcome model. The intent is that they offer the opportunity to
offset bias from mis-specifying either the propensity score model or the regression
model by getting the specification right for at least one; however, this comes at an
efficiency cost. As a result, the efficiency of doubly robust estimators tends to fall
between that of IPTW and regression estimators. These findings are consistent with
several simulation studies that have compared doubly robust estimation to multiple
imputation for addressing missing data problems (Kang & Schafer, 2007; Carpenter,
Kenward & Vansteelandt, 2006). These studies have found that IPTW by itself, or
doubly robust estimation, can be sensitive to the specification of the imputation
model—especially when some observations have a small estimated probability of
being observed. Using propensity score matching prior to estimating treatment effects
(as with doubly robust methods) has obvious appeal as it appears to simulate a
randomized trial with observational data. However, it is still possible that unobserved
variables may be correlated with both outcomes and the treatment variable, resulting
in biased estimates of ATEs. As a result, researchers should routinely test for residual
confounding or endogeneity even after propensity score matching. This can be done
in a straightforward way by including the residuals from the propensity model as an
additional variable in the outcome model (Terza, Basu & Rathouz, 2008; Hausman,
1983; Hausman, 1978). If the coefficient for the residuals variable is statistically
significant, this indicates that residual confounding or endogeneity remain.
5.7 Targeted Maximum Likelihood Estimation (TMLE) for

Causal Inference
In epidemiologic and econometrics studies, estimation of causal effects using ob-

servational data is necessary to evaluate medical treatment and policy interventions.
Numerous estimators can be used for estimation of causal effects. In the epidemiologic
literature propensity score methods or G-computation have been widely used within
the potential outcomes framework. In this chapter, we briefly discuss these other
methods as the building blocks for targeted maximum likelihood estimation (TMLE).
TMLE is a well-established alternative method with desirable statistical properties but
it is not as widely utilized as these other methods such as g-estimation and propensity
score techniques. In addition, implementation of TMLE benefits from the use of
machine learning super learner methods.
TMLE is related to G-computation and propensity score methods in that TMLE
involves estimation of both 𝐸 (𝑌 |𝑇, 𝑋) and 𝑃(𝑇 = 1|𝑋). G-computation is used to
estimate the outcome model in TMLE. G-computation (Robins & Hernan, 2009)
is an estimation approach that is especially useful for incorporating time-varying
covariates. Despite its potential usefulness, however, it is not widely used due to lack of
understanding of its theoretical underpinnings and empirical implementation (Naimi,
Cole & Kennedy, 2017). TMLE is a doubly robust, maximum-likelihood–based
estimation method that includes a secondary “targeting” step that optimizes the
bias-variance tradeoff for the parameter of interest. Although TMLE is not specifically
a causal modeling method, it has features that make it a particularly attractive

method for causal effect estimation in observational data. First, because it is a doubly
robust method it will yield unbiased estimates of the parameter of interest if either
𝐸 (𝑌 |𝑇, 𝑋) or 𝑃(𝑇 = 1|𝑋) is consistently estimated. Even if the outcome regression
is not consistently estimated, the final ATE estimate will be unbiased as long as the
exposure mechanism is also not consistently estimated. Conversely, if the outcome is
consistently estimated, the targeting step will preserve this unbiasedness and may
remove finite sample bias (van der Laan & Rose, 2011). Additionally, TMLE is an
asymptotically efficient estimator when both the outcome and exposure mechanisms
are consistently estimated (van der Laan and Rose, 2011). Furthermore, TMLE is
a substitution estimator; these estimators are more robust to outliers and sparsity
than are nonsubstitution estimators (van der Laan & Rose, 2011). Finally, when
estimated using machine learning TMLE has the flexibility to incorporate a variety
of algorithms for estimation of the outcome and exposure mechanisms. This can help
minimize bias in comparison with use of misspecified regressions for outcomes and
exposure mechanisms.
The estimation of ATE using TMLE models is comprised of several steps (Schuler
and Rose, 2017). We assume that the observed data are n independent and identically
distributed copies of 𝑂 = (𝑇,𝑌 , 𝑋) ∼ 𝑃0 , where 𝑃0 is the true underlying distribution
from which the data are drawn.
Step 1. The first step is to generate an initial estimate of 𝐸 (𝑌 |𝑇, 𝑋) using g-
estimation. 𝐸 (𝑌 |𝑇, 𝑋) is the conditional expectation of the outcome, given the
exposure and the covariates. As noted earlier, g-estimation addresses many common
problems in causal modeling such as time-varying covariates while minimizing
assumptions about the functional form of the outcome equation and data distribution.
We could use any regression model to estimate 𝐸 (𝑌 |𝑇, 𝑋) but the use of super
learning allows us to avoid choosing a specific functional form for the model. This
model is then used to generate the set of potential outcomes corresponding to T = 1
and T = 0, respectively. That is, the estimated outcome equation is used to generate
predicted outcomes for the entire sample assuming that everyone is exposed to the
intervention and that no one is exposed to the intervention, respectively. The mean
difference in the two sets of predicted outcomes is the g-estimate of the ATE:
1 ∑︁
𝐴𝑇 𝐸 𝐺−𝑐𝑜𝑚 𝑝 = 𝜙𝐺−𝑐𝑜𝑚 𝑝 = 𝐸 [𝑌 |𝑇 = 1, 𝑋] − 𝐸 [𝑌 |𝑇 = 0, 𝑋].
𝑛
However, because we used machine learning, the expected outcome estimates
have the optimal bias-variance tradeoff for estimating the outcome, not the ATE. As
a result, the ATE estimate may be biased. Nor can we compute the standard error of
the ATE. (We could bootstrap the standard errors but they would only be correct if
the estimand was asymptotically normally distributed.)
Step 2. The second step is to estimate the exposure mechanism 𝑃(𝑇 = 1|𝑋). As
with the outcome equation, we use super learner methods to estimate P (T = 1|X)
using a variety of machine learning algorithms. For each individual, the predicted
probability of exposure to the intervention is given by the propensity score 𝑃1 . The
individual’s predicted probability of exposure to the comparison treatment is 𝑃ˆ0 ,
162 Crown
where 𝑃ˆ0 = 1 − 𝑃ˆ1 . In Step 1 we estimated the expected outcome, conditional on

treatment and confounders. As noted earlier these machine learning estimates have an
optimal bias-variance trade-off for estimating the outcome (conditional on treatment
and confounders), rather than the ATE. In Step 3, the estimates of the exposure
mechanism are used to optimize the bias-variance trade-off for the ATE so we can
make valid inferences based upon the results.
Step 3. Updating the initial estimate of 𝐸 (𝑌 |𝑇, 𝑋) for each individual. This is
done by first calculating 𝐻𝑡 (𝑇 = 𝑡, 𝑋) = 𝐼 (𝑇=1)𝑃ˆ 1
− 𝐼 (𝑇=0)
𝑃ˆ 0
based upon the previously
ˆ ˆ
calculated values for 𝑃1 and 𝑃0 and each patient’s actual exposure status. This step
is very similar to IPTW but is based upon the canonical gradient (van der Laan &
Rose, 2011). We need to use H(X) to estimate the Efficient Influence Function (EIF)
of the ATE. Although we do not discuss the details here, in semi-parametric theory
an Influence Function is a function that indicates how much an estimate will change
if the input changes. If an Efficient Influence Function exists for an estimand (in this
case, the ATE), it means the estimand can be estimated efficiently. The existence
of the EIF for the ATE is what enables TMLE to use the asymptotic properties of
semi-parametric estimators to support reliable statistical inference based upon the
estimated ATE.
To estimate the EIF, the outcome variable 𝑌 is regressed on 𝐻𝑡 specifying a
fixed intercept to estimate 𝑙𝑜𝑔𝑖𝑡 (𝐸 ∗ (𝑌 |𝑇, 𝑋)) = 𝑙𝑜𝑔𝑖𝑡 (𝑌ˆ𝑡 ) + 𝛿𝐻𝑡 . We also calculate
𝐻1 (𝑇 = 1, 𝑋) = 1/𝑃ˆ1 and 𝐻0 (𝑇 = 0, 𝑋) = −1/𝑃ˆ0 . 𝐻1 is interpreted as the inverse
probability of exposure to the intervention; 𝐻0 is interpreted as the negative inverse
probability of exposure. Finally, we generate updated (“targeted”) estimates of the
set of potential outcomes using information from the exposure mechanism to reduce
bias. Note that in the 𝑙𝑜𝑔𝑖𝑡 (𝑌ˆ1∗ ) equation, 𝑙𝑜𝑔𝑖𝑡 (𝑌ˆ1 ) is not a constant value 𝐵0 . Rather,
it is a vector of values. This means that it is a fixed intercept rather than a constant
intercept. In the TMLE literature, 𝛿 is called the fluctuation parameter, because it
provides information about how much to change, or fluctuate, the initial outcome
estimates. Similarly, 𝐻 (𝑋) is referred to as the clever covariate because it ‘cleverly’
helps us solve for the EIF and then update the estimates.
Step 4. In the final step, the fluctuation parameter and clever covariate are used to
update the initial estimates of the expected outcome, conditional on confounders and
treatment These estimates have the same interpretation as the original estimates of
potential outcomes in step 1 but their values have been updated for potential exposure
bias. To update the estimate from step 1 we first need to transform the predicted
outcomes to the logit scale. Then, it is a simple matter of adjusting these estimates by
𝛿𝐻 (𝑋):
𝑙𝑜𝑔𝑖𝑡 (𝑌ˆ1∗ ) = 𝑙𝑜𝑔𝑖𝑡 (𝑌ˆ1 ) + 𝛿𝐻1 and 𝑙𝑜𝑔𝑖𝑡 (𝑌ˆ0∗ ) = 𝑙𝑜𝑔𝑖𝑡 (𝑌ˆ0 ) + 𝛿𝐻0 .
After retransforming the updated, estimated values 𝑙𝑜𝑔𝑖𝑡 (𝑌ˆ1∗ ) we calculate ATE for
the target parameter of interest as the mean difference in predicted values of outcomes
for individuals receiving the treatment relative to their predicted counterfactual
outcome if they had not received treatment.
𝑛
1 ∑︁ ˆ ∗ ˆ ∗
𝐴𝑇 𝐸 =
𝑌 − 𝑌0 .
𝑛 𝑖=1 1
The ATE is interpreted as the causal difference in outcomes that would be apparent
if all individuals in the population of interest participated in the intervention group
versus not participating in the intervention.
To obtain the standard errors for the estimated ATE we need to compute the
influence curve (IC):
𝐼𝐶 = (𝑌 − 𝐸 ∗ [𝑌 |𝑇, 𝑋]) + 𝐸 ∗ [𝑌 |𝑇 = 1, 𝑋] − 𝐸 ∗ [𝑌 |𝑇 = 0, 𝑋] − 𝐴𝑇 𝐸 .
Once we have the IC, its standard error is simply

1 √︁
𝑆𝐸 𝐼𝐶 = 𝑉 𝑎𝑟 (𝐼𝐶).
𝑛
5.8 Empirical Applications of TMLE in Health Outcomes Studies
There is a growing literature of empirical implementations of TMLE, along with

simulations comparing TMLE to alternative methods. Kreif et al. (2017) use TMLE
to estimate the causal effect of nutritional interventions on clinical outcomes among
critically ill children, comparing TMLE causal effect estimates to those from models
using g-computation and inverse probability treatment weights. After adjusting for
time-dependent confounding they find that three methods generate similar results.
Pang et al. (2016) compare the performance of TMLE and IPTW on the marginal
causal effect of statin use on the one-year risk of death among patients who had
previously suffered a myocardial infarction. Using simulation methods they show that
TMLE methods perform better with richer specifications of the outcome model and
is less likely to be biased because of its doubly-robust property. On the other hand,
IPTW had a better mean square error in a high dimensional setting. And violations of
the positivity assumption, which are common in high dimensional data, are an issue
for both methods. (Schuler & Rose, 2017) found that TMLE outperformed traditional
methods such as g-estimation and inverse probability weighting in simulation.
5.8.1 Use of Machine Learning to Estimate TMLE Models
As just described, TMLE uses doubly robust maximum likelihood estimation to

update the initial outcome model using estimated probabilities of exposure (Funk
et al., 2011). The average treatment effect (ATE) is then estimated as the average
difference in the predicted outcome for treated patients versus their outcome if they
had not been treated (their counterfactual). This approach does not require machine
learning to accomplish and, as a result, TMLE is not inherently a machine learning
164 Crown
estimation technique. However due to complexity of specifying the exposure and

outcome mechanisms machine learning methods have a number of advantages for
estimating TMLE models. In particular, the Super Learner is an ensembling ML
approach that is recommended to be used with TMLE to help overcome bias due
to model misspecification (van der Laan & Rose, 2018; Funk et al., 2011; van der
Laan & Rubin, 2006). Super Learning can draw upon the full repertoire of ML (even
non-parametric neural network models) and traditional econometric/epidemiological
methods, and produce estimates that are asymptotically as good as the best performing
model—eliminating the need to make strong assumptions about functional form and
estimation method up front (van der Laan & Rubin, 2006). Estimation using machine
learning methods is facilitated by the fact that the ATE estimated with TMLE is based
on the predicted values of outcomes for the treated and counterfactual groups, rather
than an estimated parameter value.
5.9 Extending TMLE to Incorporate Instrumental Variables
It is reasonable to expect that health care datasets such as medical claims and
electronic medical record data will not contain all variables needed to estimate
treatment exposure and treatment outcomes. As a result, TMLE models that utilize
only observed variables will be biased. Toth and van der Laan (2016) show that the
extension of TMLE to incorporate the effects of unobserved variables correlated with
both treatment and outcomes is conceptually straightforward and basically involves
the incorporation of an instrument for unobserved variables into the outcome and
exposure equations. As above we assume that the observed data are n independent and
identically distributed copies of 𝑂 = (𝑍,𝑇,𝑌 , 𝑋) ∼ 𝑃0 , where 𝑃0 is the true underlying
distribution from which the data are drawn. However, now 𝑂 also includes 𝑍 which
are instruments for unobserved variables correlated with both treatment exposure
and outcomes. Under the IV model
𝐸 [𝑌 |𝑍, 𝑋] = 𝑤 0 (𝑋) + 𝑚 0 (𝑋)𝜋0 (𝑍, 𝑋),
where 𝑚 0 (𝑋) is the model of treatment on outcomes, 𝜋0 (𝑍, 𝑋) is the model of

treatment exposure, and 𝑤 0 (𝑋) = 𝐸 [𝑌 − 𝑇 𝑚 0 (𝑋)|𝑋]. In other words, 𝑤 0 (𝑋) returns
the expected value of 𝑌 conditional on whether 𝑇 = 1 or 𝑇 = 0. Estimation of the TMLE
for the IV model begins by obtaining initial estimates of the outcome model 𝑚(𝑍, 𝑋),
the exposure model 𝜋(𝑍, 𝑋), and the instrument propensity score 𝑔(𝑋). From these,
an initial estimate of the potential outcomes for the treated and comparison groups is
generated.
The second step of estimating the TMLE for a parameter requires specifying a
loss function L(P) where the expectation of the loss function is minimized at the true
probability distribution. It is common to use the squared error loss function. The
efficient influence function (EIF) can be written as
𝐷 ∗ (𝑚, 𝑔, 𝑄 𝑥 )(𝑂) =
= 𝐻 (𝑋){𝜋0 (𝑍, 𝑋) − 𝐸 0 (𝜋0 (𝑍, 𝑋)|𝑋 }{𝑌 − 𝜋0 (𝑍, 𝑋)𝑚 0 (𝑋) − 𝑤 0 (𝑋)}−
− 𝐻 (𝑋){𝜋0 (𝑍, 𝑋) − 𝐸 0 (𝜋0 (𝑍, 𝑋)|𝑋 }𝑚 0 (𝑋)}(𝑇 − 𝜋0 (𝑍, 𝑋)) + 𝐷 𝑋 (𝑄 𝑋 ),
where 𝐻 (𝑋) is the “clever covariate” that is a function of the inverse probability
treatment weights of the exposure variable, along with 𝜁 −2 (𝑋), a term that measures
instrument strength. 𝐻 (𝑋), 𝜁 −2 (𝑋), and 𝐷 𝑋 (𝑄 𝑋 ) are defined as
𝐸 [𝑉 2 ] − 𝐸 [𝑉]𝑉  −2
 
−1
𝐻 (𝑋) = 𝑉 𝑎𝑟 (𝑉) 

 𝜁 (𝑋),

 𝑉 − 𝐸 [𝑉] 
 
𝜁 −2 (𝑋) = 𝑉 𝑎𝑟 𝑍 |𝑋 (𝜋(𝑍, 𝑋)|𝑋),
𝐷 𝑋 (𝑄 𝑋 ) = 𝑐{𝑚 0 (𝑋) − 𝑚 𝜓 (𝑉)}.

Here 𝑉 is a variable in 𝑋 for which we wish to estimate the treatment effect. In
the targeting step, a linear model for 𝑚 0 (𝑋) is fitted using only the clever covariate
𝐻 (𝑋). This model is used to generate potential outcomes for [𝑌 |𝑇 = 1] and [𝑌 |𝑇 = 0].
Finally, the average treatment effect is estimated as the mean difference in predicted
values of outcomes for individuals receiving the treatment relative to their predicted
counterfactual outcome if they had not received treatment.
𝑛
1 ∑︁ ˆ ∗ ˆ ∗
𝐴𝑇 𝐸 =
𝑌 − 𝑌0 .
𝑛 𝑖=1 1
5.10 Some Practical Considerations on the Use of IVs
Although estimation of TMLE models with IVs is theoretically straightforward,

estimates can be very sensitive to weak instruments. Weak instruments are also likely
to be associated with residual correlation of the instrument with the residuals in the
outcome equation creating opportunities to introduce bias in the attempt to correct
for it. There is an extensive literature on the practical implications of implementing
IVs which has direct relevance for their use in TMLE models.
Effective implementation of instrumental variables methods requires finding
variables that are correlated with treatment selection but uncorrelated with the
outcome variable. This turns out to be extremely difficult to do. The difficulty
of finding instrumental variables that are correlated with treatment selection but
uncorrelated with outcomes often leads to variables that have weak correlations with
treatment. It is important to recognize that an extensive literature has now shown that
the use of variables that are only weakly correlated with treatment selection and/or
that are even weakly correlated with the residuals in the outcome equation can lead
to larger bias than ignoring the endogeneity problem altogether (Bound, Jaeger &
166 Crown
Baker, 1995; Staiger & Stock, 1997; Hahn & Hausman, 2002; Kleibergen & Zivot,
2003; Crown, Henk & Vanness, 2011). Excellent introductions and summaries of
the instrumental variable literature are provided in Basu, Navarro and Urzua (2007),
Brookhart, Rassen and Schneeweiss (2010), and Murray (2007).
Bound et al. (1995) show that the incremental bias of instrumental variable versus
ordinary least squares (OLS) is inversely proportional to strength of the instrument
and number of variables that are correlated with treatment but not the outcome
variable. Crown et al. (2011) conducted a Monte Carlo simulation analysis of the
Bound et al. (1995) results to provide empirical estimates of the magnitude of bias in
instrumental variables under alternative assumptions related to the strength of the
correlation between the instrumental variable and the variable that it is intended to
replace. They also examine how bias in the instrumental variable estimator is related
to the strength of the correlation between the instrumental variable and the observed
residuals (the contamination of the instrument). Finally, they examine how bias
changes in relation to sample size for a range of study sizes likely to be encountered
in practice. The results were sobering. For the size of samples used in most studies,
the probability that instrumental variable is outperformed by OLS is substantial, even
when the asymptotic results indicate bias to be lower for instrumental variable, when
the endogeneity problem is serious, and when the instrumental variable has a strong
correlation with the treatment variable. This suggests that methods focusing upon
observed data, such as propensity score matching or IPTW, will generally be more
efficient than those that attempt to control for unobservables, although it is very
important to test for whether any residual confounding or endogeneity remains. These
results have implications for attempts to include IVs in the estimation of TMLE as
well. In particular, more research is needed on the effects of residual correlation of
instruments on the bias and efficiency of TMLE.
5.11 Alternative Definitions of Treatment Effects
This chapter has focused upon the use of TMLE for the estimation of ATEs. There
are multiple potential definitions for treatment effects, however, and it is important
to distinguish among them. The most basic distinctions are between the average
treatment effect (ATE), the average treatment effect of the treated (ATT), and the
marginal treatment effect (MTE) (Jones & Rice, 2009; Basu, 2011; Basu et al., 2007).
These alternative treatment effect estimators are defined as differences in expected
values of an outcome variable of interest (𝑌 ) conditional on covariates (𝑋) as follows
(Heckman & Navarro, 2003):
𝐴𝑇 𝐸 : 𝐸 (𝑌1 −𝑌0 |𝑋),

𝐴𝑇𝑇 : 𝐸 (𝑌1 −𝑌0 |𝑋,𝑇 = 1),
𝑀𝑇 𝐸 : 𝐸 (𝑌1 −𝑌0 |𝑋, 𝑍,𝑉 = 0),
where 𝑇 refers to the treatment, 𝑍 is an instrumental variable (or variables) that,

conditional on 𝑋, is correlated with treatment selection but not outcomes, and V
measures the net utility arising from treatment. Basu et al. (2007) show that the
MTE is the most general of the treatment effects, since both the ATE and ATT
can be derived from the MTE once it has been estimated. ATEs are defined as the
expected difference in outcomes between two groups, conditional upon their observed
covariates; some patients may not, in fact, have received the treatment at all. ATEs are
very common in clinical trials testing the efficacy or safety of a treatment. An example
from observational data might be the parameter estimate for a dummy variable
comparing diabetes patients enrolled in a disease management program with those
not enrolled in the program. Similarly, ATTs are defined as the expected difference in
outcomes between one treated group versus another treated group. Researchers often
attempt to estimate ATTs for therapeutic areas where multiple treatments exist and
make head-to-head comparisons among treatments (e.g., depressed patients treated
with selective serotonin reuptake inhibitors vs. tricyclic antidepressants). As with
all statistical parameters, ATEs and ATTs can be defined for both populations and
samples. For most of the estimators discussed in this chapter, the distinction between
ATEs and ATTs has little implication for choice of statistical estimator. Although
researchers generally refer to the estimation of ATEs in medical outcomes studies,
most such studies are actually estimates of ATT. MTEs, on the other hand, are
relatively new to the empirical literature and have important implications for choice
of the statistical estimator. In particular, the estimation of MTEs highlights two key
issues: (i) the existence of common support among the treatment groups; and (ii) the
presence of unobserved essential heterogeneity in treatment selection and outcomes.
The first of these characteristics links the estimation of MTEs to propensity score
methods while the second links the estimation of MTEs to instrumental variables.
Notably, the TMLE estimator described in this chapter, when implemented using IV,
is an estimator of MTEs.
By referring to the econometric literature on IV estimation, it is clear that the
modeling of IVs in TMLE can be given a utility maximization interpretation. When
heterogeneity in treatment response exists and patients select into treatment based
upon their expectation of the utility that they will receive from treatment, it becomes
necessary to model treatment selection in order to interpret the instrumental variable
estimates (Basu et al., 2007). The probability of a patient selecting treatment T can
be modeled as a function of the utility, 𝑇 ∗ , that a person expects to receive from the
treatment. Let 𝑍 be an instrumental variable correlated with treatment selection but
uncorrelated with 𝑌 ,
𝑃𝑟 (𝑇 = 1|𝑋, 𝑍).
Note that the probability of treatment includes the instrumental variable Z. This leads
to the sample selection model that has the following form:
𝑇 = 1 if 𝑃𝑟 (𝑇 ∗ > 0); 𝑇 = 0 otherwise.
That is, if the expected utility 𝑇 ∗ associated with treatment 𝑇 is greater than 0
(standard normal scale), the individual will choose treatment 𝑇 over the alternative.
168 Crown
𝑇 ∗ = 𝐶0 + 𝐶1 𝑋1 + 𝐶𝑐 𝑍 + 𝑒𝑇 ,
𝑌 = 𝐵0 + 𝐵1𝑖𝑣 𝑇 ∗ +𝐵2 𝑋 + 𝑒𝑌 ,
where 𝐶0 , 𝐶1 and 𝐶2 are parameters to be estimated, 𝐵1𝑖𝑣 is the instrumental variable

estimate of treatment effectiveness and the remaining variables and parameters are as
previously defined. There are many extensions of the basic sample selection model to
account for different functional forms, multiple outcome equations, etc. (Cameron &
Trivedi, 2013; Maddala, 1983).
Vytlacil (2002) points out that the semiparametric sample selection model is
equivalent to the method of local instrumental variables (LIV). LIV estimation
enables the identification of MTEs, which are defined as the average utility gain to
patients who are indifferent to the treatment alternatives given 𝑋 and 𝑍 (Basu et al.,
2007; Heckman & Navarro, 2003; Evans & Basu, 2011; Basu, 2011). A particularly
attractive feature of MTEs is that all mean treatment effect estimates can be derived
from MTEs. For instance, the ATT is derived as a weighted average of the MTEs
over the support of the propensity score (conditional on 𝑋). Evans and Basu (2011)
provide a very clear description of LIV methods, MTEs, and the relationship of
MTEs to other mean treatment effect estimates.
5.12 A Final Word on the Importance of Study Design in

Mitigating Bias
Most studies comparing average treatment effects (ATEs) from observational studies
with randomized controlled trials (RCTs) for the same disease states have found a
high degree of agreement (Anglemyer, Horvath & Bero, 2014; Concato, Shah &
Horwitz, 2000; Benson & Hartz, 2000). However, other studies have documented
considerable disagreement in such results introduced by the heterogeneity of datasets
and other factors (Madigan et al., 2013). In some cases, apparent disagreements have
been shown to be due to avoidable errors in observational study design which, upon
correction, found similar results from the observational studies and RCTs (Dickerman,
García-Albéniz, Logan, Denaxas & Hernán, 2019; Hernán et al., 2008).
For any question involving causal inference, it is theoretically possible to design a
randomized trial to answer that question. This is known as designing the target trial
(Hernán, 2021). When a study is designed to emulate a target trial using observational
data some features of the target trial design may be impossible to emulate. Emulating
treatment assignment requires data on all features associated with the implementation
of the treatment intervention. This is the basis for the extensive use of propensity
score matching and inverse probability weighting in the health outcomes literature.
It has been estimated that a relatively small percentage of clinical trials can be
emulated using observational data (Bartlett, Dhruva, Shah, Ryan & Ross, 2019).
However, observational studies can still be designed with a theoretical target trial
References 169
in mind–specifying a randomized trial to answer the question of interest and then

examining where the available data may limit the ability to emulate this trial (Berger
& Crown, 2021). Aside from lack of comparability in defining treatment groups, there
are a number of other problems frequently encountered in the design of observational
health outcomes studies including immortal time bias, adjustment for intermediate
variables, and reverse causation. The target trial approach is one method for avoiding
such issues.
Numerous observational studies have designed target trials designed to emulate
existing RCTs in order to compare the results from RCTs to those of the emulations
using observational data (Seeger et al., 2015; Hernán et al., 2008; Franklin et al.,
2020; Dickerman et al., 2019). In general, such studies demonstrate higher levels
of agreement than comparisons of ATE estimates from observational and RCTS
within a disease area that do not attempt to emulate study design characteristics
such as inclusion/exclusion criteria, follow-up periods, etc. For example, a paper
comparing RCT emulation results for 10 cardiovascular trials found that the hazard
ratio estimate from the observational emulations was within the 95% CI from the
corresponding RCT in 8 of 10 studies. In 9 of 10, the results had the same sign and
statistical significance.
To date, all of the trial emulations have used traditional propensity score or IPTW
approaches. None have used doubly-robust methods such as TMLE implemented with
Super Learner machine learning methods. In addition to simulation studies, it would
be useful to examine the ability of methods like TMLE to estimate similar treatment
effects as randomized trials—particularly in cases where traditional methods have
failed to do so.
References
Abadie, A. & Cattaneo, D., Matias. (2018). Econometric methods for program
evaluation. Annual Review of Economics, 10, 465–503.
Anglemyer, A., Horvath, H. & Bero, L. (2014). Healthcare outcomes assessed with
observational study designs compared with those assessed in randomized trials.
The Cochrane Database of Systematic Reviews, 4.
effects. Proceedings of the National Academy of Sciences, 113, 7353–7360.
Athey, S. & Imbens, G. (2019). Machine learning methods that economists should
know about. Annual Review of Economics, 11, 685–725.
Athey, S., Imbens, G. & Wager, S. (2018). Approximate residual balancing: Debiased
inference of average treatment effects in high dimensions. Journal of the Royal
Statistical Society, Series B (Methodological), 80, 597–623.
Athey, S., Tibshirani, J. & Wager, S. (2019). Generalized random forests. Annals of
Statistics, 47, 399–424.
Bang, H. & Robins, J. (2005). Doubly robust estimation in missing data and causal
inference models. Biometrics, 61, 692–972.
170 Crown
Bartlett, V., Dhruva, S., Shah, N., Ryan, P. & Ross, J. (2019). Feasibility of using
real-world data to replicate clinical trial evidence. JAMA Network Open, 2,
e1912869.
Baser, O. (2006). Too much ado about propensity score models? comparing methods
of propensity score matching. Value in Health: The Journal of the International
Society for Pharmacoeconomics and Outcomes Research, 9, 377–385.
Basu, A. (2011). Economics of individualization in comparative effectiveness
research and a basis for a patient-centered health care. Journal of Health
Economics, 30, 549-59.
Basu, A., Navarro, S. & Urzua, S. (2007). Use of instrumental variables in the
presence of heterogeneity and self-selection: An application to treatments of
breast cancer patients. Health Economics, 16, 1133–1157.
Basu, A., Polsky, D. & Manning, W. (2011). Estimating treatment effects on
healthcare costs under exogeneity: Is there a ’magic bullet’? Health Services &
Outcomes Research Methodology, 11, 1-26.
Belloni, A., Chen, D., Chernozhukov, V. & Hansen, C. (2012). Sparse models and
methods for optimal instruments with an application to eminent domain. SSRN
Electronic Journal, 80, 2369–2429.
Belloni, A., Chernozhukov, V., Fernndez-Val, I. & Hansen, C. (2017). Program
evaluation and causal inference with high-dimensional data. Econometrica, 85,
233–298.
Belloni, A., Chernozhukov, V. & Hansen, C. (2013). Inference for high-dimensional
sparse econometric models. Advances in Economics and Econometrics: Tenth
World Congress Volume 3, Econometrics, 245–295.
Belloni, A., Chernozhukov, V. & Hansen, C. (2014a). High-dimensional methods
and inference on structural and treatment effects. The Journal of Economic
Perspectives, 28, 29–50.
Studies, 81, 29–50.
Benson, K. & Hartz, A. (2000). A comparison of observational studies and
randomized, controlled trials. The New England Journal of Medicine, 342,
1878–1886.
Berger, M. & Crown, W. (2021). How can we make more rapid progress in the
leveraging of real-world evidence by regulatory decision makers? Value in
Health, 25, 167–170.
Bound, J., Jaeger, D. & Baker, R. (1995). Problems with instrumental variables
estimation when the correlation between the instruments and the endogenous
explanatory variable is weak. Journal of the American Statistical Association,
90, 443–450.
Brookhart, M., Rassen, J. & Schneeweiss, S. (2010). Instrumental variable methods
for comparative effectiveness research. Pharmacoepidemiology and Drug
safety, 19, 537-554.
Brookhart, M., Schneeweiss, S., Rothman, K., Glynn, R., Avorn, J. & Sturmer, T.
(2006). Variable selection for propensity score models. American Journal of
References 171
Epidemiology, 163, 1149-1156.

Cameron, A. & Trivedi, P. (2013). Regression analysis of count data (2nd ed.).
Carpenter, J., Kenward, M. & Vansteelandt, S. (2006). A comparison of multiple
imputation and doubly robust estimation for analyses with missing data. Journal
of the Royal Statistical Society Series A, 169, 571–584.
Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C. & Newey,
W. (2017). Double/debiased/neyman machine learning of treatment effects.
American Economic Review, 107, 261–265.
& Robins, J. (2018). Double/debiased machine learning for treatment and
structural parameters. The Econometrics Journal, 21, 1–C68.
Cole, S. & Frangakis, C. (2009). The consistency statement in causal inference: a
definition or an assumption? Epidemiology, 20, 3–5.
Cole, S. & Hernán, M. (2008, 10). Constructing inverse probability weights for
marginal structural models. American Journal of Epidemiology, 168, 656–64.
Concato, J., Shah, N. & Horwitz, R. (2000). Randomized, controlled trials, observa-
tional studies, and the hierarchy of research designs. New England Journal of
Medicine, 342, 1887-1892.
Crown, W. (2015). Potential application of machine learning in health outcomes
research and some statistical cautions. Value in Health, 18, 137–140.
Crown, W., Henk, H. & Vanness, D. (2011). Some cautions on the use of instrumental
variables estimators in outcomes research: How bias in instrumental variables
estimators is affected by instrument strength, instrument contamination, and
sample size. Value in Health, 14, 1078–1084.
Crump, R., Hotz, V., Imbens, G. & Mitnik, O. (2009). Dealing with limited overlap
in estimation of average treatment effects. Biometrika, 96, 187–199.
Dahabreh, I., Robertson, S., Tchetgen, E. & Stuart, E. (2019). Generalizing causal
inferences from randomized trials: Counterfactual and graphical identification.
Biometrics, 75, 685–694.
D’Amour, A., Ding, P., Feller, A., Lei, L. & Sekhon, J. (2021). Overlap in observational
studies with high-dimensional covariates. Journal of Econometrics, 221, 644–
654.
Dickerman, B., García-Albéniz, X., Logan, R., Denaxas, S. & Hernán, M. (2019, 10).
Avoidable flaws in observational analyses: an application to statins and cancer.
Nature Medicine, 25, 1601–1606.
Evans, H. & Basu, A. (2011). Exploring comparative effect heterogeneity with
instrumental variables: prehospital intubation and mortality (Health, Econo-
metrics and Data Group (HEDG) Working Papers). HEDG, c/o Department of
Economics, University of York.
Franklin, J., Patorno, E., Desai, R., Glynn, R., Martin, D., Quinto, K., . . . Schneeweiss,
S. (2020). Emulating randomized clinical trials with nonrandomized real-
world evidence studies: First results from the RCT DUPLICATE initiative.
Circulation, 143, 1002–1013.
Funk, J., Westreich, D., Wiesen, C., Stürmer, T., Brookhart, M. & Davidian, M.
172 Crown
(2011, 03). Doubly robust estimation of causal effects. American Journal of

Epidemiology, 173, 761–767.
Futoma, J., Morris, M. & Lucas, J. (2015). A comparison of models for predicting
early hospital readmissions. Journal of Biomedical Informatics, 56, 229–238.
Greenland, S. & Robins, J. (1986). Identifiability, exchangeability, and epidemiolo-
gical confounding. International Journal of Epidemiology, 15, 413–419.
Hahn, J. & Hausman, J. (2002). A new specification test for the validity of instrumental
variables. Econometrica, 70, 163–189.
Hastie, T., Tibshirani, R. & Friedman, J. (2009). The elements of statistical learning:
Data mining, inference and prediction (2nd ed.). Springer Verlag, New York.
Hausman, J. (1978). Specification tests in econometrics. Econometrica, 46, 1251–
1271.
Hausman, J. (1983). Specification and estimation of simultaneous equation models.
In Handbook of econometrics (pp. 391–448). Elsevier.
Heckman, J. & Navarro, S. (2003). Using matching, instrumental variables and
control functions to estimate economic choice models. Review of Economics
and Statistics, 86.
Hernán, M. (2011). Beyond exchangeability: The other conditions for causal inference
in medical research. Statistical Methods in Medical Research, 21, 3–5.
Hernán, M. (2021). Methods of public health research–strengthening causal inference
from observational data. The New England Journal of Medicine, 385, 1345–
1348.
Hernán, M., Alonso, A., Logan, R., Grodstein, F., Michels, K., Willett, W., . . . Robins,
J. (2008). Observational studies analyzed like randomized experiments an
application to postmenopausal hormone therapy and coronary heart disease.
Hirano, K., Imbens, G. & Ridder, G. (2003). Efficient estimation of average treatment
effects using the estimated propensity score. Econometrica, 71, 1161–1189.
Hong, W., Haimovich, A. & Taylor, R. (2018). Predicting hospital admission
at emergency department triage using machine learning. PLOS ONE, 13,
e0201016.
Imbens, G. (2020). Potential outcome and directed acyclic graph approaches to
causality: Relevance for empirical practice in economics. Journal of Economic
Literature, 58, 1129–1179.
Imbens, G. & Wooldridge, J. (2009). Recent developments in the econometrics of
program evaluation. Journal of Economic Literature, 47, 5–86.
Joffe, M., Have, T., Feldman, H. & Kimmel, S. (2004). Model selection, confounder
control, and marginal structural models: Review and new applications. The
American Statistician, 58, 272–279.
Johnson, M., Bush, R., Collins, T., Lin, P., Canter, D., Henderson, W., . . . Petersen,
L. (2006). Propensity score analysis in observational studies: outcomes
after abdominal aortic aneurysm repair. American Journal of Surgery, 192,
336–343.
Johnson, M., Crown, W., Martin, B., Dormuth, C. & Siebert, U. (2009). Good
research practices for comparative effectiveness research: analytic methods
References 173
to improve causal inference from nonrandomized studies of treatment effects

using secondary data sources: the ispor good research practices for retrospective
database analysis task force report–part iii. Value Health, 12, 1062–1073.
Jones, A. & Rice, N. (2009). Econometric evaluation of health policies. In The
Oxford Handbook of Health Economics.
Kang, J. & Schafer, J. (2007). Demystifying double robustness: A comparison of
alternative strategies for estimating a population mean from incomplete data.
Statistical Science, 22, 523–539.
Kleibergen, F. & Zivot, E. (2003). Bayesian and classical approaches to instrumental
variable regression. Journal of Econometrics, 29-72.
Knaus, M., Lechner, M. & Strittmatter, A. (2021). Machine learning estimation of het-
erogeneous causal effects: Empirical monte carlo evidence. The Econometrics
Journal, 24.
Kreif, N., Tran, L., Grieve, R., Stavola, B., Tasker, R. & Petersen, M. (2017). Estimating
the comparative effectiveness of feeding interventions in the pediatric intensive
care unit: A demonstration of longitudinal targeted maximum likelihood
estimation. American Journal of Epidemiology, 186, 1370–1379.
Maddala, G. S. (1983). Limited-dependent and qualitative variables in econometrics.
Madigan, D., Ryan, P., Schuemie, M., Stang, P., Overhage, J. M., Hartzema, A.,
. . . Berlin, J. (2013). Evaluating the impact of database heterogeneity on
observational study results. American Journal of Epidemiology, 178, 645–651.
Mitra, N. & Indurkhya, A. (2005). A propensity score approach to estimating
the cost-effectiveness of medical therapies from observational data. Health
Economics, 14, 805—815.
Mullainathan, S. & Spiess, J. (2017). Machine learning: An applied econometric
approach. Journal of Economic Perspectives, 31, 87–106.
Murray, M. (2007). Avoiding invalid instruments and coping with weak instruments.
Journal of Economic Perspectives, 20, 111–132.
Naimi, A., Cole, S. & Kennedy, E. (2017). An introduction to g methods. International
Journal of Epidemiology, 46, 756–762.
Obermeyer, Z. & Emanuel, E. (2016). Predicting the future — big data, machine
learning, and clinical medicine. The New England Journal of Medicine, 375,
1216–1219.
Pang, M., Schuster, T., Filion, K., Schnitzer, M., Eberg, M. & Platt, R. (2016).
Effect estimation in point-exposure studies with binary outcomes and high-
dimensional covariate data - a comparison of targeted maximum likelihood
estimation and inverse probability of treatment weighting. The international
Journal of Biostatistics, 12.
Pearl, J. (2009). Causality (2nd ed.). Cambridge University Press.
Petersen, M., Porter, K., Gruber, S., Wang, Y. & Laan, M. (2012). Diagnosing and
responding to violations in the positivity assumption. Statistical Methods in
Medical Research, 21, 31–54.
Rajkomar, A., Oren, E., Chen, K., Dai, A., Hajaj, N., Liu, P., . . . Dean, J. (2018).
Scalable and accurate deep learning for electronic health records. npj Digital
174 Crown
Medicine, 18.
Ramsahai, R., Grieve, R. & Sekhon, J. (2011, 12). Extending iterative matching
methods: An approach to improving covariate balance that allows prioritisation.
Health Services and Outcomes Research Methodology, 11, 95–114.
Richardson, T. (2013, April). Single world intervention graphs (swigs): A unification
of the counterfactual and graphical approaches to causality (Tech. Rep. No.
Working Paper Number 128). Center for Statistics and the Social Sciences.
University of Washington.
Robins, J. (1986). A new approach to causal inference in mortality studies with
sustained exposure periods - application to control of the healthy worker
survivor effect. Computers & Mathematics With Applications, 14, 923–945.
Robins, J. & Hernan, M. (2009). Estimation of the causal effects of time varying
exposures. In In: Fitzmaurice g, davidian m, verbeke g, and molenberghs g
(eds.) advances in longitudinal data analysis (pp. 553–599). Boca Raton, FL:
Chapman & Hall.
Robins, J., Rotnitzky, A. G. & Zhao, L. (1994). Estimation of regression coefficients
when some regressors are not always observed. Journal of The American
Statistical Association, 89, 846–866.
Rosenbaum, P. & Rubin, D. (1983). The central role of the propensity score in
observational studies for causal effects. Biometrika, 70, 41–55.
Rubin, D. B. (1974). Estimating causal effects if treatment in randomized and
nonrandomized studies. Journal of Educational Psychology, 66, 688–701.
Rubin, D. B. (1980). Randomization analysis of experimental data: The fisher
randomization test. Journal of the American Statistical Association, 75(371),
575–582.
Rubin, D. B. (1986). Statistics and causal inference: Comment: Which ifs have causal
answers. Journal of the American Statistical Association, 81, 961–962.
Rubin, D. B. (2006). Matched sampling for causal effects. Cambridge University
Press, Cambridge UK.
Scharfstein, D., Rotnitzky, A. G. & Robins, J. (1999). Adjusting for nonignorable
drop-out using semiparametric nonresponse models. JASA. Journal of the
American Statistical Association, 94, 1096–1120. (Rejoinder, 1135–1146).
Schuler, M. & Rose, S. (2017). Targeted maximum likelihood estimation for causal
inference in observational studies. American Journal of Epidemiology, 185,
65–73.
Seeger, J., Bykov, K., Bartels, D., Huybrechts, K., Zint, K. & Schneeweiss, S. (2015,
10). Safety and effectiveness of dabigatran and warfarin in routine care of
patients with atrial fibrillation. Thrombosis and Haemostasis, 114, 1277–1289.
Sekhon, J. & Grieve, R. (2012). A matching method for improving covariate balance
in cost-effectiveness analyses. Health Economics, 21, 695–714.
Setoguchi, S., Schneeweiss, S., Brookhart, M., Glynn, R. & Cook, E. (2008).
Evaluating uses of data mining techniques in propensity score estimation: A
simulation study. Pharmacoepidemiology and Drug Safety, 17, 546–555.
Shi, C., Blei, D. & Veitch, V. (2019). Adapting neural networks for the estimation of
treatment effects..
References 175
Shickel, B., Tighe, P., Bihorac, A. & Rashidi, P. (2018). Deep ehr: A survey of
recent advances on deep learning techniques for electronic health record (ehr)
analysis. Journal of Biomedical and Health Informatics., 22, 1589–1604.
Staiger, D. & Stock, J. H. (1997). Instrumental variables regression with weak
instruments. Econometrica, 65, 557–586.
Terza, J., Basu, A. & Rathouz, P. (2008). Two-stage residual inclusion estimation:
Addressing endogeneity in health econometric modeling. Journal of Health
Economics, 27, 531–543.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of
the Royal Statistical Society: Series B (Methodological), 58, 267–288.
Ting, D., Cheung, C., Lim, G., Tan, G., Nguyen, D. Q., Gan, A., . . . Wong, T.-Y.
(2017). Development and validation of a deep learning system for diabetic
retinopathy and related eye diseases using retinal images from multiethnic
populations with diabetes. JAMA, 318, 2211-2223.
Toth, B. & van der Laan, M. J. (2016, June). TMLE for marginal structural models
based on an instrument (Tech. Rep. No. Working Paper 350). U.C. Berkeley
Division of Biostatistics Working Paper Series.
van der Laan, M. & Rose, S. (2011). Targeted learning: Causal inference for
observational and experimental data. Springer.
van der Laan, M. & Rose, S. (2018). Targeted learning in data science: Causal
inference for complex longitudinal studies.
van der Laan, M. & Rubin, D. (2006). Targeted maximum likelihood learning.
International Journal of Biostatistics, 2, 1043–1043.
Vytlacil, E. (2002). Independence, monotonicity, and latent index models: An
equivalence result. Econometrica, 70, 331–341.
113(523), 1228-1242.
Westreich, D. & Cole, S. (2010, 02). Invited commentary: Positivity in practice.
American Journal of Epidemiology, 171, 674–677; discussion 678–681.
Westreich, D., Lessler, J. & Jonsson Funk, M. (2010). Propensity score estim-
ation: Neural networks, support vector machines, decision trees (cart), and
meta-classifiers as alternatives to logistic regression. Journal of Clinical
Wooldridge, J. (2002). Econometric analysis of cross-section and panel data. MIT
Press.
Chapter 6
Econometrics of Networks with Machine
Learning
Oliver Kiss and Gyorgy Ruzicska
Abstract Graph structured data, called networks, can represent many economic
activities and phenomena. Such representations are not only powerful for developing
economic theory but are also helpful in examining their applications in empirical
analyses. This has been particularly the case recently as data associated with net-
works are often readily available. While researchers may have access to real-world
network structured data, in many cases, their volume and complexities make analysis
using traditional econometric methodology prohibitive. One plausible solution is
to embed recent advancements in computer science, especially machine learning
algorithms, into the existing econometric methodology that incorporates large net-
works. This chapter aims to cover a range of examples where existing algorithms in
the computer science literature, machine learning tools, and econometric practices
can complement each other. The first part of the chapter provides an overview of
the challenges associated with high-dimensional, complex network data. It discusses
ways to overcome them by using algorithms developed in computer science and
econometrics. The second part of this chapter shows the usefulness of some ma-
chine learning algorithms in complementing traditional econometric techniques by
providing empirical applications in spatial econometrics.
6.1 Introduction
Networks are fundamental components of a multitude of economic interactions.

Social relationships, for example, might affect how people form new connections,
Oliver Kiss B
Central European University, Budapest, Hungary and Vienna, Austria e-mail: Kiss_Oliver@phd.ceu
.edu
Gyorgy Ruzicska
Central European University, Budapest, Hungary and Vienna, Austria, e-mail: Ruzicska_Gyorgy@
phd.ceu.edu
178 Kiss and Ruzicska
while ownership and managerial networks could affect how companies interact in a
competitive environment. Likewise, geographic networks can influence where nations
export to and import from in international trade. Researchers should incorporate the
observable network dependencies in their analyses whenever social, geographical, or
other types of linkages influence economic outcomes. Such data is also increasingly
available due to the rise of digitization and online interactions. In the literature,
economic studies with networks have analyzed, among other topics, peer effects
(Sacerdote, 2001), social segregation (Lazarsfeld & Merton, 1954), production
networks (Acemoglu, Carvalho, Ozdaglar & Tahbaz-Salehi, 2012), and migration
networks (Ortega & Peri, 2013).
There is a growing literature on incorporating network structured data into the
econometric estimation framework. Besides, there is an increasing number of papers
in machine learning that extract information from large-scale network data and
perform predictions based on such data sets. On the other hand, there are relatively
few network-related topics studied both by econometricians and machine learning
experts. This chapter aims to provide an overview of the most used econometric
models, machine learning methods, and algorithmic components in their interaction.
We further discuss how different approaches can augment or complement one another
when incorporated into social and economic analyses.
The chapter proceeds as follows. The following section introduces the terminology
used throughout the chapter whenever we refer to a network or its components. This
section is not exhaustive and only provides definitions necessary for understanding
our discussion. In Section 6.3, we highlight the most significant difficulties that arise
when network data is used for econometric estimation. Section 6.4 discusses graph
representation learning, a way to reduce graph dimensionality and extract valuable
information from usually sparse matrices describing a graph. In Section 6.5, we
discuss the problem of sampling networks. Due to the complex interrelations and
the often important properties encoded in neighborhoods, random sampling almost
always destroys salient information. We discuss methods proposed in the literature
aiming to extract better representations of the population. While the techniques
mentioned above have received significant attention in the computer science literature,
they have – to our knowledge – not been applied in any well-known economic work
yet. Therefore, in Section 6.6, we turn our attention to a range of canonical network
models that have been used to analyze spatial interactions and discuss how the
spatial weight matrix can be estimated using machine learning techniques. Then, we
introduce gravity models, which have been the main building blocks of trade models,
and provide a rationale for using machine learning techniques instead of standard
econometric methods for forecasting. The chapter closes with the geographically
weighted regression model and shows an example where econometric and machine
learning techniques effectively augment each other.
6 Econometrics of Networks with Machine Learning 179
6.2 Structure, Representation, and Characteristics of Networks
Networks have been studied in various contexts and fields ranging from sociology
and economics through traditional graph theory to computer science. Due to this
widespread interest, notations and terminology also differ across fields of study.
Throughout this chapter, we rely on a unified notation and terminology introduced in
the paragraphs below.
A network (or graph) 𝐺 is given by a pair (V, E) consisting of a set of nodes
or vertices V = {1, 2, ..., 𝑛} and a set of edges E ⊆ {(𝑖, 𝑗)|𝑖, 𝑗 ∈ V} between them.
An edge (𝑖, 𝑗) is incident to nodes 𝑖 and 𝑗. Networks can be represented with an 𝑛
dimensional positive adjacency matrix A ∈ R+0 | V |×| V | , where each column and row
corresponds to a node in the network. The 𝑎 𝑖 𝑗 element of this matrix contains the
weight of the directed edge originating at node 𝑖 targeting node 𝑗. Throughout this
chapter, the terms network and graph refer to the same object described above.
The values in the adjacency matrix can be understood as the strength of interactions
between the two corresponding nodes. There are various types of interactions that can
be quantified with such links. For example, in spatial econometrics, an edge weight
can denote the distance (or its inverse) between two separate locations represented by
the nodes. In social network analysis, these edge weights may indicate the number of
times two individuals interact with each other in a given period of time.
The diagonal elements (𝑎 𝑖𝑖 ) of the adjacency matrix have a special interpretation.
They indicate if there exists an edge originating at a node pointing to itself. Such
representation is mainly useful in dynamic networks where the actions of an agent
can have effects on its future self. In most static applications, however, the diagonal
elements of the adjacency matrix are zero.
Directed and undirected networks. A network may be undirected if all of its edges
are bidirectional (with identical weights in both directions) or directed if some are
one directional (or if the weights differ). In both types of networks, the lack of a
link between nodes 𝑖 and 𝑗 is represented by the 𝑎 𝑖 𝑗 element of the adjacency matrix
being zero. In a directed network having a one directional link from node 𝑖 to node 𝑗
means that 𝑎 𝑖 𝑗 > 0 and 𝑎 𝑗𝑖 = 0. If the network is undirected, then its adjacency matrix
is symmetric, i.e., 𝑎 𝑖 𝑗 = 𝑎 𝑗𝑖 ∀𝑖, 𝑗 ∈ V.
Different economic and social relationships can be represented by different types
of networks. In trade networks, edges are usually directed as they describe the flow
of goods from one country to another. On the other hand, friendship relationships are
generally considered reciprocal and are, therefore, characterized by undirected edges.
Weighted and unweighted networks. A network may be unweighted if all its ties
have the same strength. In unweighted networks, the elements of the adjacency matrix
are usually binary (𝑎 𝑖 𝑗 ∈ {0, 1}∀𝑖, 𝑗 ∈ V), indicating the existence of a link between
two nodes. In a weighted network setting, edges can be assigned different weights.
These weights usually represent a quantifiable measure of the strength of connection,
and they are incorporated into the elements of the adjacency matrix.
In some applications, such as spatial econometrics, networks are usually weighted,
as edge weights can denote the spatial distance between locations or the number of
times two agents interact in a given time period. In contrast, some settings make it
difficult or impossible to quantify the strength of connections (consider, for example,
the strength of friendship relations) and are, therefore, more likely to be represented
by unweighted networks.
5 4
2 3
1 3
1 4
2
Fig. 6.1: Example for an undirected network (left) and a directed network (right)
Network structured data. Throughout this section, network data or network struc-
tured data refers to a data set containing nodes, potentially with their characteristics,
and relationships between these nodes described by edges, and possibly edge charac-
teristics. How this data is stored usually depends on the size and type of the network.
Small networks can easily be represented by their adjacency matrices which are able
to capture weights and edge directions at the same time. In this case, node and edge
characteristics are usually given in separate cross-sectional data files. The size of
the adjacency matrix (which is usually a sparse matrix) scales quadratically in the
number of nodes. Consider, for example, the networks shown in Figure 6.1. Let us
denote the adjacency matrix of the undirected network by A and that of the directed
network by B. Then corresponding adjacency matrices are
 
  0 1 1 0 1
0 1 0 1 






 0 0 0 0 1
1 0 1 1  
A =  and B = 0
 

 0 0 0 1 .
0 1 0 0 






 0 1 0 0 0
1 1 0 0 



  1 0 0 0 0
 
In practice, large networks are therefore rather described by edge lists instead of the
adjacency matrix. In this representation, we have a data set containing the source
and target node identifiers and additional columns for edge characteristics (such as
weight). Node characteristics are usually stored in a separate file.
Characteristics of networks. Due to the unique structure of network data, core
statistics characterizing a network are also specific and have been designed to capture
certain aspects of the underlying relationships. There are two main types of statistics
regarding graph structured data. One set (usually related to nodes) aims to describe
local relationships by summarizing the neighborhoods of nodes, while others aim to
characterize the network as a whole. Although there is a multitude of such measures,
this chapter relies predominantly on the following:
• Two nodes are neighbors if there is a common edge incident to both of them.
• The set of neighbors of node 𝑖 is N (𝑖) = { 𝑗 ∈ V |(𝑖, 𝑗) ∈ E or ( 𝑗, 𝑖) ∈ E}.
• The degree of node 𝑖 is the number of its neighbors 𝑑 (𝑖) = |N (𝑖)|.
• The degree distribution of a graph 𝐺 (V, E) is the distribution over 𝑑 (𝑖)|𝑖 ∈ V.
• A path between the nodes 𝑖 and 𝑗 is a set of 𝑛 edges in E
{(𝑠1 , 𝑡 1 ), (𝑠2 , 𝑡2 ), . . . , (𝑠 𝑛−1 , 𝑡 𝑛−1 ), (𝑠 𝑛 , 𝑡 𝑛 )},
such that 𝑠1 = 𝑖, 𝑡 𝑛 = 𝑗 and 𝑠 𝑘 = 𝑡 𝑘−1 ∀𝑘 ∈ 2, . . . , 𝑛.

• Two nodes belong to the same connected component if there exists a path between
the two nodes.
• A random walk of length 𝑛 from node 𝑖 is a randomly generated path from node 𝑖.
The next edge is always chosen uniformly from the set of edges originating in the
last visited node.
• The centrality of a node is a measure describing its relative importance in the
network. There are several ways to measure this. For example, degree centrality
uses the degree of each node, while closeness centrality uses the average length of
the shortest path between the node and all other nodes in the graph. More complex
centrality measures usually apply a different aggregation of degrees or shortest
paths. The Katz centrality, for example, uses the number of all nodes that can be
connected through a path, while the contributions of distant nodes are penalized.
These measures can often be used for efficient stratified sampling of graphs. For
example, the PageRank of a node (another centrality measure) is used in PageRank
node sampling, a method presented in Section 6.5.
• The local clustering coefficient of a node measures how sparse or dense the
immediate neighborhood of a node is. Given 𝑑 (𝑖) – the degree of node 𝑖 – it is
straightforward that in a directed network there can be at most 𝑑 (𝑖) (𝑑 (𝑖) − 1) edges
connecting the neighbors of node 𝑖. The local clustering coefficient measures what
fraction of this theoretical maximum of edges is present in the network; thus, in a
directed network it is given by
|{𝑒 𝑗 𝑘 s.t. 𝑗, 𝑘 ∈ N (𝑖) and 𝑒 𝑗 𝑘 ∈ E}|

𝐶 (𝑖) = .
𝑑 (𝑖)(𝑑 (𝑖) − 1)
In an undirected setting, the number of edges present must be divided by 𝑑 (𝑖) (𝑑 (𝑖) −
1)/2.
• The degree correlation – or degree assortativity – of a network measures whether
similar nodes (in terms of their degree) are more likely to be connected in the
network. The phenomenon that high-degree nodes are likely to be connected to
other high-degree nodes is called assortative mixing. On the contrary, if high-
degree nodes tend to have low-degree neighbors, we call it disassortative mixing.

Details on the calculation of this measure are discussed by Newman (2002).
While the list of network characteristics above is far from being exhaustive, it is
sufficient to understand our discussion in the upcoming sections.
Distinction between network structured data and neural networks. This chapter
discusses econometric and machine learning methods that utilize network structured
data. In Section 6.6, we introduce neural network-based machine learning methods,
including deep neural networks, convolutional neural networks, and recurrent neural
networks. Importantly, these neural network architectures are not directly related to
the network structured data. While networks are often used as inputs to these neural
network models, the name "network" in these machine learning models refers to their
architectural design. As defined in this section, networks are graphs that represent
economic activities, while neural networks are techniques used to identify hidden
patterns in the data. Chapter 4 discusses all the neural network architectures described
in this chapter.
6.3 The Challenges of Working with Network Data
Network data differs from traditional data structures in many aspects. These differences
result in unique challenges requiring specific solutions or the refinement of existing
econometric practices. For example, the size of the adjacency matrix can result in
computational challenges, while establishing causal relationships is complicated by
complex interrelations between agents (represented by nodes) in the networks. This
section highlights the most important issues arising from using network structured data
in an econometric analysis. These specific aspects must be considered in theoretical
modeling and empirical studies that use networks.
Curse of dimensionality. The analysis of large-scale real-world network data (like
web graphs or social networks) is often difficult due to the size of the data set. The
number of data points in an adjacency matrix increases quadratically with the number
of agents in a network. This is in contrast with other structured data sets, where new
observations increase the size of the data linearly. Due to its high dimensionality
and computational difficulties, the adjacency matrix cannot be directly included in
an econometric specification in many cases. The traditional solution in econometric
applications has been incorporating network aggregates instead of the whole network
into the estimation. Another common alternative is to multiply the adjacency matrix
(or a transformation of it) by regressors from the right to control for each node’s
neighbors’ aggregated characteristics when modeling their individual outcomes.
For example, peer effects in networks are modeled by Ballester, Calvo-Armengol
and Zenou (2006) using such aggregates. In their model, each agent 𝑖 chooses the
intensity of action 𝑦 𝑖 to maximize:
1 ∑︁
𝑢 𝑖 (𝑦 1 , ...𝑦 𝑛 ) = 𝛼𝑖 𝑦 𝑖 − 𝛽𝑖 𝑦 2𝑖 + 𝛾 𝑎𝑖 𝑗 𝑦𝑖 𝑦 𝑗 ,
2 𝑗≠𝑖
where the adjacency matrix elements 𝑎 𝑖 𝑗 represent the strength of connection between
agents 𝑖 and 𝑗. The main coefficient of interest in such a setting is 𝛾 – often called the
peer effect coefficient – measuring the direct marginal effect of an agent’s choice on
the outcome of a connected peer with a unit connection strength. In this specification,
utility is directly influenced by an agent’s own action along with all its neighbors’
actions. An agent’s optimal action is also indirectly affected by the actions of all
agents belonging to the same connected component in the network. In fact, the optimal
action of nodes in such a setting is directly related to their Katz-Bonancich centrality.
The canonical characterization of outcomes being determined by neighbors’
actions and characteristics is attributed to Manski (1993):
𝑁
∑︁ 𝑁
∑︁
𝑦𝑖 = 𝛼 + 𝛽 𝑎 𝑖 𝑗 𝑦 𝑗 + 𝜇𝑥 𝑖 + 𝛾 𝑎𝑖 𝑗 𝑥 𝑗 + 𝜖𝑖 .
𝑗=1 𝑗=1
where 𝑎 𝑖 𝑗 is defined as in the previous example and 𝑥𝑖 , 𝑥 𝑗 are node-level characteristics

of agent 𝑖 and 𝑗, respectively. This specification is discussed further in Section 6.6.1.
While such approaches are easy to interpret, recent advances in computer sci-
ence provide more efficient machine learning algorithms, which can capture more
complex characteristics in a reduced information space. These methods decrease
the dimensionality of a network by embedding it into a lower-dimensional space.
This lower-dimensional representation can be used to control for node or edge level
information present in the network in downstream tasks, such as regressions. Some
of the most widely used dimensionality reduction techniques are discussed in Section
6.4.
Sampling. Sampling is crucial when it is impossible to observe the whole population
(all nodes and edges) or when the size of the population results in computational
challenges. In a network setting, however, random sampling destroys information that
might be relevant to a researcher since local network patterns carry useful information.
Chandrasekhar and Lewis (2016) provide an early study on the econometrics of
sampled networks. They show for two specific random node sampling approaches1 that
sampling leads to a non-classical measurement error, which results in biased regression
coefficient estimates. We discuss how applying different sampling algorithms might
help preserve important information and discuss common node and edge sampling
approaches in Section 6.5. Most of these approaches have been designed to preserve a
specific network attribute assuming a particular network type. While their theoretical
properties are known, their applicability to real-life data has only been studied in a
handful of papers (Rozemberczki, Kiss & Sarkar, 2020).
1 Both techniques (random node sampling with edge induction and random node-neighbor sampling)
are discussed in detail in Section 6.5.
Identification, reverse causality, and omitted variables. Inference based on network

data is complicated by the network structure being related to agents’ characteristics
and actions (observed and unobserved), making it endogenous in most specifications.
That is, networks determine outcomes that, in turn, affect the network structure.
Therefore, when econometricians try to identify chains of causality, they are often
faced with inherently complex problems that may also incorporate a circle of causality.
To avoid this issue, researchers must control for all the characteristics that could affect
behavior and drive the observed network structure. When some covariates cannot be
controlled for – either because of unobservability or unavailable data – the estimation
could be biased due to the omitted variables problem.
To illustrate these problems, assume that we would like to document peer effects
in an educational setting empirically. In particular, educational outcome (GPA) is
regressed on friends’ GPA and characteristics:
𝑌𝑖 = 𝛼𝑋𝑖 + 𝛽 𝑋¯ −𝑖 + 𝛾𝑌¯−𝑖 + 𝜖 𝑖 ,
where 𝑋𝑖 are observed individual characteristics, and 𝜖𝑖 incorporates the unobserved

individual characteristics and a random error term. 𝑋¯ −𝑖 and 𝑌¯−𝑖 measure the average
of neighbors’ characteristics and GPA, respectively.
This model specifies two types of peer effects. 𝛽 measures the ‘contextual effect’
reflecting how peer characteristics affect individual action. On the other hand, 𝛾 is
the ‘endogenous effect’ of neighbors’ actions. In education, these effects manifest in
higher GPA if the agent’s peers work hard (endogenous effect) or because peers are
intelligent (contextual effect).
We may encounter three problems when identifying coefficients in this regression
model. First, there are presumably omitted variables that we cannot control for. For
example, if peers in a group are exposed to the same shock, 𝜖 𝑖 is correlated with 𝜖−𝑖 and
hence 𝑌¯−𝑖 . Second, if agents select similar peers, the estimation suffers from selection.
In such a case, 𝜖 𝑖 is correlated with both 𝑋¯ −𝑖 and 𝑌¯−𝑖 . Third, this specification is an
example of the reflection problem documented by Manski (1993). The reflection
problem occurs when agents’ actions are determined jointly in equilibrium. Hence 𝜖𝑖
is correlated with 𝑌¯−𝑖 .
To avoid these problems with identification and disentangle causal effects, eco-
nometricians have analyzed experimental and quasi-experimental setups where
network connections could be controlled for. For example, Sacerdote (2001) used
randomization in college roommate allocation to identify peer effects in educational
outcomes. Alternatively, researchers may use instrumental variables to overcome the
endogeneity of the social structure in their estimations. Jackson (2010) discusses
endogeneity, identification, and instrumental variables in further detail and highlights
other problems that arise with identification in social network analysis, including
nonlinearities in social interactions and the issue of timing.
In the machine learning literature, spatiotemporal signal processing has been
applied to model the co-evolution of networks and outcomes. A spatiotemporal
deep learning model combines graph representation learning (extracting structurally
representative information from networks) with temporal deep learning techniques.
These models rely on a temporal graph sequence (a graph of agents observed through
different time periods). They utilize a graph neural network block to perform message
passing at each temporal unit. Then, a temporal deep learning block incorporates
the new information into the model. This combination of techniques can be used to
model both temporal and spatial autocorrelation across the spatial units and agents
(Rozemberczki, Scherer, He et al., 2021).
In general, machine learning methods aimed at establishing causal relationships
in networks are rare. The literature primarily focuses on using models for forecasting
and prediction. While these methods benefit data-driven decision-making and policy
analysis, algorithms focusing on causality remain an important future research
domain.
6.4 Graph Dimensionality Reduction
With the rise of digitalization, data became available on an unprecedented scale in

many research domains. This abundance of data results in new research problems.
One such problem is the availability of too many potentially valuable right-hand side
variables. Regularization techniques such as LASSO2 have a proven track record
of separating valuable variables (i.e., those with high explanatory power) from
less valuable ones. Another traditional technique in econometrics is dimensionality
reduction. This process aims to represent the data in a lower-dimensional space while
retaining useful information. Some of these techniques (mainly principal component
analysis, which uses a linear mapping to maximize variance in the lower-dimensional
representation) have an established history of being used in applied works. However,
there is also a growing number of methodologies primarily applied in machine
learning to achieve the same goal. The analysis of network data is a prime example
of where these algorithms gained importance in the last decade.
Algorithms applied in this domain are collectively called embedding techniques.
These techniques can be used to represent the original high dimensional data in a
significantly lower-dimensional space where the distance of the embedded objects is
traditionally associated with a measure of similarity. Algorithms differ in terms of
what is embedded in a lower-dimensional space (nodes, edges, or the whole graph),
what graph property is used to measure similarity, and how this similarity is used to
encode information in the lower-dimensional space. The representations obtained
through these algorithms can then be used to control for the structural information
present in the network. This section provides an overview of methodologies prevalent
in the machine learning literature that can be useful for applied econometricians
working with graph data.
2 See, e.g., Chapters 1. and 2.

6.4.1 Types of Embeddings
Node embedding. Probably the most common application of embedding techniques

is node embedding. In this case, given a graph 𝐺 (V, E) and the dimensionality
of the embedding 𝑑 << |V | the goal is to assign a 𝑑-dimensional vector to each
node in the graph. The pairwise distances between nodes in this lower-dimensional
– usually Euclidean – space correspond to a similarity measure of choice. That is,
nodes similar to each other in some sense are to be closer in this lower-dimensional
space. There is widespread research on how this similarity measure should be defined.
There are three major approaches in the literature in this aspect. Neighborhood
preserving node embeddings (for example, Diff2Vec (Rozemberczki & Sarkar, 2018)
or Node2Vec (Grover & Leskovec, 2016)) aim to preserve the distance between nodes
in their lower-dimensional representations meaning that neighbors will be closer to
each other. This is usually achieved by the decomposition of proximity matrices, a
methodology described in Section 6.4.2. Another major set of techniques, called
structural embeddings, aims to preserve the structural role of nodes in the graph.
That is, nodes with similar structural properties (e.g., degree, clustering coefficient,
centrality) are to be closer in the embedding space. The third set of techniques
exploits additional information encoded in node attributes. In this case, the pairwise
similarities of nodes are constructed by considering a set of additional node-level
characteristics. In a social network setting, this could, for example, mean a set of
variables describing the socioeconomic status of each node. These methods are
collectively called attributed node embeddings.
Edge embedding. While relationships between edges might seem somewhat less
complex than those between nodes, they may contain just as much useful information.
In some applications focusing on connections between agents (e.g., friendships or
trade), edges are of primary interest. Predicting which of their friends people trust
the most or understanding how information spreads in a network are examples where
edges have a crucial role. Structural information on edges can be encoded in an
adjacency matrix by transforming the original graph into a so-called line graph. Given
a graph 𝐺 (V, E), its line graph is another graph in which each edge is represented
by a node, and each node is represented by an edge. The nodes in the line graph are
connected if the corresponding edges in the original graph 𝐺 (V, E) have a node that
is incident to both of them. Notice that a line graph can be described by an |E | × |E |
adjacency matrix. The problem of high dimensionality thus persists when one aims
to focus on edges without losing potentially valuable network information. In such
cases, edge embeddings might be useful. The core idea is identical to that of the
node embeddings. Given a graph 𝐺 (V, E) and the dimensionality of the embedding
𝑑 << |E | the goal is to assign a 𝑑-dimensional vector to each edge in the graph.
Whole graph embedding. In some cases, the unit of observation is a graph. For
example, one might study ego-nets, where each agent has a corresponding network
that consists of the agent and its neighbors. Another example is analyzing discussions
in a forum where each graph represents a thread. The users participating in the
discussion are the nodes, and the replies to each other are the edges. In such cases,
the goal is to embed whole graphs, that is, represent each graph in a set of graphs
{𝐺 1 , 𝐺 2 , . . . , 𝐺 𝑛 } with a 𝑑-dimensional vector.
6.4.2 Algorithmic Foundations of Embeddings
Matrix factorization. A major set of embedding algorithms relies on the factorization

of matrices encoding graph properties. For example, one might consider a matrix
containing pairwise edge or node similarity measures. These matrices are then
factorized either directly or – more frequently – using graph Laplacian eigenmaps. Let
us denote a pairwise similarity matrix by W. In the case of pairwise node similarities
W ∈ R | V |×|V | , and W𝑖 𝑗 measures the similarity between nodes 𝑖 and 𝑗. Likewise in
edge similarities W ∈ R | E |× | E | , and W𝑢𝑣 measures the similarity between edges 𝑢 and
𝑣. Given a dimensionality (𝑑), the goal is to find an embedding Y ∈ R |V |×𝑑 for nodes
or Y ∈ R | E |×𝑑 for edges that minimizes a pre-defined loss function. A traditional loss
function used by a variety of algorithms is a classical quadratic loss utilizing the
Euclidean (ℓ 2 ) norm as given in Equation (6.1):
∑︁
Y∗ = arg min W𝑖 𝑗 · ||Y𝑖 − Y 𝑗 || 22 . (6.1)
𝑖≠ 𝑗
Y𝑖 in Equation (6.1) is the 𝑑 dimensional representation of node or edge 𝑖 in the

embedding space. Let us define a diagonal matrix D, where
∑︁
D𝑖𝑖 = W𝑖 𝑗 .
𝑖, 𝑗 ∈V
Then the Laplacian of the similarity matrix is given by
L = D − W.
Notice that using the graph Laplacian Equation (6.1) reduces to
Y∗ = arg min 𝑡𝑟 (Y′LY). (6.2)
Belkin and Niyogi (2002) show that one needs to introduce an additional orthogonality
constraint in order to obtain a unique solution. This is, in most cases, solved by
enforcing
Y′DY = I.
The problem is then given by Equation (6.3):
Y∗ = arg min 𝑡𝑟 (Y′LY). (6.3)

Y′ DY=I
By using the Lagrange method it can be seen that the optimal solution in this case
consists of eigenvectors that solve the generalized eigenvalue problem given in
Equation (6.4) (Torres, Chan & Eliassi-Rad, 2020; Belkin & Niyogi, 2002):
LY𝑖∗ = 𝜆𝑖 DY∗𝑖 . (6.4)
To be more precise, the eigenvectors corresponding to the 𝑑 lowest eigenvalues of the

normalized Laplacian D−1 L are the columns of Y∗ ; thus, the embedding vector of
node 𝑗 consists of the 𝑗 th elements of these 𝑑 eigenvectors. The eigenvalues in such a
setting also turned out to be useful. The whole graph embedding method SF (de Lara
& Edouard, 2018) applies the same approach, representing each whole graph in a set
of graphs with the 𝑑 lowest eigenvalues of the graph in question.
A core question is how one defines pairwise similarities. Some neighborhood
preserving methods, such as the seminal Laplacian Eigenmap (Belkin & Niyogi,
2002) or GLEE (Torres et al., 2020), use the adjacency matrix directly. Others rely
on different similarity metrics. For example, Isomap (Balasubramanian & Schwartz,
2002) uses the sum of edge weights along the shortest path between node pairs.
Cai, Zheng and Chang (2018) provide a comprehensive survey on Laplacian-based
embeddings with different similarity measures and alternative objective functions.
Sequence-based methods. A wide set of algorithms relies on observing the neigh-
borhood of each node (either through direct sampling or through multiple random
walks originating at each node) and then learning a representation by maximizing
the likelihood of predicting the correct neighborhood for each node using its lower-
dimensional representation. Given a dimensionality 𝑑 the goal is to find an embedding
Y ∈ R | V |×𝑑 for the nodes in a graph 𝐺 (V, E). A prominent method in this category
is node2vec (Grover & Leskovec, 2016) presented below. Other algorithms such as
DeepWalk (Perozzi, Al-Rfou & Skiena, 2014), LINE (Tang et al., 2015), or diff2vec
(Rozemberczki & Sarkar, 2018) use similar approaches with modifications in terms
of the sampling strategy or the exact likelihood function.
With node2vec, for every node 𝑖 ∈ V, we define a set of neighboring nodes
N (𝑖) ⊂ V. The goal is to maximize a log-likelihood function given in Equation (6.5)
∑︁
max log 𝑃𝑟 (N (𝑖)|Y𝑖 ), (6.5)
Y
𝑖 ∈V
where Y𝑖 is the lower-dimensional representation of node 𝑖. Grover and Leskovec

(2016) make two important assumptions to make the above problem tractable. By
assuming that the probability of observing one node in the neighborhood of node
𝑖 is independent of observing any other node in the neighborhood given node 𝑖’s
representation, they factorize the joint probability of observing a neighborhood to
observing individual nodes. This results in the property given in Equation (6.6).
Ö
𝑃𝑟 (N (𝑖)|Y𝑖 ) = 𝑃𝑟 ( 𝑗 |Y𝑖 ). (6.6)
𝑗 ∈N (𝑖)
An important question is how to model the conditional probability in Equation (6.6).

Grover and Leskovec (2016) propose a simple logit structure using the scalar products
of the embedding vectors given in Equation (6.7)
𝑒𝑥 𝑝(Y𝑖 Y′𝑗 )
𝑃𝑟 ( 𝑗 |Y𝑖 ) = Í . (6.7)
𝑒𝑥 𝑝(Y𝑖 Y′𝑘 )
𝑘 ∈V
With the assumptions above, Equation (6.5) reduces to
∑︁ © ∑︁
max − log 𝑍 𝑘 + Y𝑖 Y′𝑗 ® ,
ª
Y
𝑘 ∈V « 𝑗 ∈N (𝑖) ¬
where ∑︁
log 𝑍 𝑘 = 𝑒𝑥 𝑝(Y𝑖 Y′𝑘 ).
𝑘 ∈V
It is worth noting that 𝑍 𝑘 is computationally expensive to calculate throughout the

optimization process. Therefore, practical implementations usually approximate it
with the so-called negative sampling method (Mikolov, Sutskever, Chen, Corrado &
Dean, 2013).
6.5 Sampling Networks
Real-world graphs are often too large to be analyzed directly. Consider, for example,
social networks with millions of users and billions of edges. In these networks, even
deriving basic descriptive statistics might be computationally challenging, if not
impossible, with resources available to most researchers. Sampling might seem to be
a straightforward solution to this problem. However, applying the proper sampling
technique is a major challenge in analyzing large, web-scale graphs. Choosing
the appropriate sampling technique, in general, has two components. First, large
networks often cause computational issues (Kang, Tsourakakis & Faloutsos, 2009;
Gonzalez, Low, Gu, Bickson & Guestrin, 2012) rendering some traditional sampling
algorithms unusable. Second, networks describe complex underlying relations between
agents, and therefore choosing an inappropriate sampling technique might destroy
valuable information. Sampling severely affects a multitude of core descriptive
network characteristics such as degree distribution, diameter, clustering coefficient,
or transitivity (Easley & Kleinberg, 2010). The choice of the sampling technique is,
therefore, less obvious. It has to be reflective of the downstream task. That is, one has
to choose the sampling technique that retains the valuable information needed for
the task at hand. In general, there are two practical points where graph sampling can
happen during the analysis of graph structured data:
1. At data collection. Consider, for example, field experiments collecting data on
the social network of individuals (see e.g., Crépon, Devoto, Duflo & Parienté,
2015 or Ferrali, Grossman, Platas & Rodden, 2020). In such cases, collecting
the whole network would practically imply collecting data on all 7.75 billion
individuals in the world. By construction, these experiments rely on samples (for
example, the network of people living in the same village or students in the same
school). Analysis relying on such samples should consider the consequences of

such sampling approaches on network features. Another example where sampling
might be necessary at the point of collecting the data is web graphs. Web graphs
are large-scale graph structured data sets from the web. For example, the graph
of users on social network websites (Twitter, Facebook, LinkedIn, etc.) and their
connections. Another web graph is the world wide web itself, where websites
are the nodes and hyperlinks pointing to each other are the links. In these cases
collecting the whole graph might be impossible or impractical.
2. At data analysis. Consider having access to a whole web graph. The analysis of
these graphs is impossible on commercial hardware. In business environments,
hardware constraints might be less of an issue. It is, however, still unnecessarily
costly in many cases to analyze the whole network where performing the analysis
on a proper sample is possible.
There is a wide range of algorithms proposed (Hu & Lau, 2013) and implemented
(Rozemberczki et al., 2020) for different objectives. This section overviews the most
important approaches and the potential applications.
6.5.1 Node Sampling Approaches
One approach is to sample the network by selecting a subset of vertices N𝑠 ⊆ N .

There are four widely discussed methods in the literature for sampling nodes.
Random node sampling. This sampling technique has a single parameter 𝑝 corres-
ponding to the probability of a node being included in the sample. That is, each node
is picked randomly with probability 𝑝 to be included in the sampled network. An edge
will only be part of the sampled network if the nodes connected by it are both included
in the sampled network. While this sampling technique is computationally simple and
works fast for large networks, it heavily distorts the degree distribution of the sampled
graph unless the original graph has a positive or negative binomial degree distribution
(the average degree is different even in these cases) (Stumpf, Wiuf & May, 2005).
It is worth noting that Chandrasekhar and Lewis (2016) offer a method to correct
for this kind of bias. The results of Rozemberczki et al. (2020) show that a potential
application domain of this sampling method is the estimation of network assortativity
through the calculation of degree correlation. Degree correlations estimated from
samples utilizing this method are closer to the ground truth than those using other
types of node sampling for multiple real-life social networks.
Random node sampling with neighborhood. We start by selecting a random subset
of nodes 𝑉𝑠 ⊆ 𝑉 using random node sampling. Then the algorithm induces the
neighboring nodes as the node set and the edges between all of the nodes. This
practically results in observing a set of nodes and their immediate neighborhood.
This technique is useful when we are interested in the occurrence of small, local
patterns such as three-node cliques.
Degree-based node sampling. In many real-life networks, the number of connections

of a node encodes essential information. For example, nodes with a higher degree
in social networks represent people with more friends, connections, or followers.
Highly connected nodes often have important roles in network stability, information
transmission, etc. It is, therefore, a meaningful idea to select nodes in the sample with
probabilities weighted according to their degree if we want to include potentially
more influential agents with a higher probability (Adamic, Lukose, Puniyani &
Huberman, 2001). This approach is, by construction, biased towards the inclusion
of high-degree nodes, and therefore graphs sampled using this technique have very
different distributional properties than their originals (Leskovec & Faloutsos, 2006).
PageRank node sampling. Similarly to the degree-based node sampling, this ap-
proach relies on selecting nodes into the sample using a pre-defined distribution over
the set of nodes. The core idea is to calculate the PageRank (Page, Brin, Motwani &
Winograd, 1999) of each node in the node set. PageRank is a widely used algorithm
to rank nodes in a network. The PageRank of a node 𝑖 ∈ N is given by
1−𝑑 ∑︁ 𝑃𝑅( 𝑗)
𝑃𝑅(𝑖) = +𝑑 ,
|N | 𝐷 +𝑗
𝑗 ∈M𝑖
where 𝑑 is the so called damping factor, M𝑖 is the set of nodes that have an edge
pointing to node 𝑖 , 𝐷 +𝑗 is the out-degree of 𝑗, and |N | is the total number of nodes.
The PageRank of a node shows how likely a random walk is to go through the
given node. Roughly speaking, ‘easily reachable’ nodes in the network have a higher
PageRank.
6.5.2 Edge Sampling Approaches
An alternative to node sampling is edge sampling. These techniques are predominantly

helpful when the underlying research question focuses on the connections rather than
the agents connected. In these cases, we are analyzing edge attributes. For example,
one might be interested in the differences in friendship characteristics (number of
interactions, length of friendship, etc.) between agents from the same and different
cities. In this case, we need to focus on the differences between two distinct sets of
edges. Next, we outline the most prevalent methodologies for edge sampling.
Random edge sampling. Just like with nodes, it is possible to select edges into the
sample with a uniform probability (𝑝). Nodes connected by the edges chosen will
also be part of the sample. Unless the goal is to approximate edge-level average
characteristics, this sampling method has limited use. The resulting sampled graph is
usually (depending on the value of 𝑝) very sparse, but a multitude of salient network
features are destroyed.
Random node-edge sampling. This method slightly differs from the one mentioned
above. The traditional random edge sampling method is biased towards selecting high
degree nodes into the sample by construction. Random node-edge sampling is used
to solve this issue. First, one selects nodes with a uniform probability, 𝑝, and then
for each selected node, we pick a single edge incident to the node into the sample at
random.
Induced edge sampling methods. Random edge sampling with edge set induction
first randomly samples edges with a fixed probability. Nodes incident to these edges
will be included in the sample. Finally, edges between nodes already in the sample
are retained with an induction step. This method is computationally expensive as it
requires iterating through the edge set exactly twice. A similar approach was proposed
by Ahmed, Neville and Kompella (2013), performing only a partial induction step.
Their algorithm (PIES) is a stream sampling algorithm that produces a sample from
an edge stream. This has the advantage of low computational costs as – alongside
the already sampled part of the graph – only a single edge is kept in the memory at
the time. The algorithm iterates through a set of edges and selects each edge into the
sample (along with the nodes incident to it) with a fixed probability. Edges connecting
two nodes already in the sample are also added to the sample, resulting in a partial
induction (edges are only induced if both the nodes they connect have been added to
the sample previously). This method (along with the full induction version) is known
to have a bias towards high degree nodes due to the random edge selection step. High
degree nodes have more incident edges and thus are more likely to be selected into
the sample.
Hybrid Approaches and the Importance of the Problem

Many edge and node selection algorithms can be combined relatively easily to fit
the underlying research question better. For example, Krishnamurthy et al. (2005)
present an approach where at each sampling step, with probability 𝑤, they select an
edge into the sample using random edge sampling. With probability 1 − 𝑤, they use
random node-edge sampling. They claim that these methods are biased in opposite
directions for multiple metrics in their study. This, however, does not hold for every
case: Leskovec and Faloutsos (2006) and Rozemberczki et al. (2020) evaluate a
variety of sampling algorithms using different networks and different network metrics.
They find that metrics calculated from samples generated using a simple random
edge sampling algorithm are usually closer to the ground truth metric than those
calculated from the hybrid samples in their applications. This is yet another case
showing that the choice of the sampling method should always correspond to the
downstream estimation task.
6.5.3 Traversal-Based Sampling Approaches
Traversal or exploration-based sampling methods extract samples from a graph by

probing the neighborhood on one or more seed nodes. The majority of modern
sampling algorithms belong to this category. Next, we outline the main types of these
algorithms and the best-known versions in each group.
6.5.3.1 Search Based Techniques
Search-based techniques sample a network by starting from seed node (s) and looking
for specific neighbors until well-defined stopping criteria are reached.
Breadth first search. With breadth first search (Doerr & Blenn, 2013) we need to
define a starting node 𝑛0 first and add it to a queue denoted by 𝑄. We also define the –
so far – empty set of sampled nodes 𝑉𝑠 . The algorithm proceeds by processing the
elements of the queue. It always removes the first element of the queue, adds it to the
set of sampled nodes, and then adds all those neighbors of the node to the queue that
are neither sampled nor already in the queue. The search process ends once a budget
is exhausted or if there are no more elements in the queue. The easiest example of
such a budget is sampling a fixed number of nodes. In this case, the initial budget is
the number of nodes we want to sample, and at each sampled node, it decreases by
one. The sampling stops once the budget reaches zero. It is also possible to make
certain sampled nodes more expensive depending on their characteristics. The final
sample is the graph induced by the selected nodes.
Depth first search. This algorithm is identical to breadth first search except for how
it selects the next node from the queue into the sample. With depth first search (Doerr
& Blenn, 2013), we always choose the last node in the queue into the sample. The
names of these search methods are reasonably intuitive. Breadth first search will
result in a broad tree of sampled nodes as it iterates through all neighbors of a few
nodes, while depth first search prefers newly added neighbors and, thus, results in a
deep, narrow tree.
Random first search. As its name suggests, this method randomly selects the next
node into the sample from the queue.
Snowball sampling. Snowball sampling (Goodman, 1961) is a restricted version
of the breadth first search approach, where we only add a limited number (𝑘) of
neighbors to the queue for each sampled node.
Forest fire. Forest fire can be thought of as a stochastic version of snowball sampling,
where the restriction on the number of neighbors visited only holds in expectation
(Leskovec, Kleinberg & Faloutsos, 2005). Instead of selecting 𝑘 neighbors at each
iteration, forest fire selects 𝐾 ∼ 𝐺𝑒𝑜𝑚𝑒𝑡𝑟𝑖𝑐( 𝑝) of them at each step. Notice that if
𝑝 = 𝑘1 then it holds that 𝐸 [𝐾] = 𝑘.
Pseudo Code for Search-Based Sampling Algorithms.

Algorithm 1 presents the pseudo code for the basic search-based algorithms described
in this section.
Algorithm 1: Search based sampling algorithms

Data: Graph 𝐺, initial node 𝑠, budget 𝐵 and cost 𝑏
Result: 𝐺𝑠 : Sampled graph
1 𝑉𝑠 ← ∅
2 𝑄 ← 𝐿𝑖𝑠𝑡 (𝑠)
3 while |𝑄 | > 0 and 𝐵 > 0 do
4 n = Q.RemoveFirst (BFS) / n = Q.RemoveLast (DFS) / n = Q.RemoveRandom (RFS) 𝑉𝑠
← 𝑉𝑠 ∪ 𝑛 𝐵 ← 𝐵 − 𝑏 for 𝑣 in neighbors(𝑛) do
5 if 𝑣 ∉ 𝑄 and 𝑣 ∉ 𝑉𝑠 then
6 Q.Append(𝑣);
7 end
8 end
9 end
10 𝐺𝑠 ← G.Induce(𝑉𝑠 )
Community structure expansion. Starting from a randomly selected node, this

algorithm adds new nodes to the sample based on their expansion factor (Maiya &
Berger-Wolf, 2010). Let us denote the set of sampled nodes by 𝑉𝑠 . At each iteration,
we calculate |𝑁 ({𝑣}) − (𝑁 (𝑉𝑠 ) ∪𝑉𝑠 |, the expansion factor for each node 𝑣 ∈ 𝑁 (𝑉𝑠 ),
where 𝑁 (𝑣) is the neighbor set of node 𝑣 and 𝑁 (𝑉𝑠 ) is the union of the neighbors
of all the nodes already in the sample. Then, we select the node with the largest
expansion factor into the sample. The process goes on until the desired number of
sampled nodes is reached. The main intuition behind the algorithm is that at each
iteration the sample is extended by one node. The algorithm selects the node which
reaches the largest number of nodes that are not in the immediate neighborhood of
the sampled nodes yet. This method is known to provide samples better representing
different communities in the underlying graph (Maiya & Berger-Wolf, 2010). This
is because nodes acting as bridges between different communities will have larger
expansion factors and are, thus, more likely to be selected into the sample than
members of communities that already have sampled members.
Shortest path sampling. In shortest path sampling (Rezvanian & Meybodi, 2015),
one chooses non-incident node pairs and adds a randomly selected shortest path
between these nodes to the sample at each iteration. This continues until the desired
sample size is reached. Finally, edges between the sampled nodes are induced.
6.5.3.2 Random Walk-Based Techniques
Random walk-based techniques start from a seed node and traverse the graph
inducing a sample between visited nodes. The basic random walk approach has many
shortcomings addressed by the extended algorithms presented next.
Random walk sampler. The most straightforward random walk sampling technique
(Gjoka, Kurant, Butts & Markopoulou, 2010) starts from a single seed node 𝑠 and
adds nodes to the sample by walking through the graph’s edges randomly until a
pre-defined sample size is reached.
Rejection-constrained Metropolis-Hastings random walk. By construction, higher-
degree nodes are more likely to be included in the sample obtained by a basic random
walk. This can cause problems in many settings (especially social network analysis)
that require a representative sample degree distribution. The Metropolis-Hastings
random walk algorithm (Hübler, Kriegel, Borgwardt & Ghahramani, 2008; Stutzbach,
Rejaie, Duffield, Sen & Willinger, 2008; R.-H. Li, Yu, Qin, Mao & Jin, 2015)
addresses this problem by making the walker more likely to select lower-degree nodes
into the sample. The degree to which lower-degree nodes are preferred can be set
by choosing a single rejection constraint parameter 𝛼. A random neighbor (𝑤) of
the most recently added node (𝑣) is selected each iteration. We generate a random
at
𝛼
uniform number 𝛾 ∼ 𝑈 [0, 1]. If 𝛾 < ||𝑁𝑁 (𝑤)
(𝑣) |
| then node 𝑤 is added to the sample,
otherwise, it is rejected and the iteration is repeated. The process continues until a
pre-determined number of nodes is added to the sample. Notice that a lower degree
makes it more likely that a node is accepted into the sample. Increasing 𝛼 increases
the probability that high degree nodes are rejected.
Non-backtracking random walk. A further known issue of the traditional random
walk sampler is that it is likely to get stuck in densely connected, small communities.
There are multiple solutions to overcome this problem. Non-backtracking random
walks restrict the traditional random walk approach so that the walker cannot go to
the node it came from. More formally, a walker currently at node 𝑗, immediately
before at node 𝑖, will move to the next node 𝑘 ∈ 𝑁 ( 𝑗) \ {𝑖} if 𝑑 ( 𝑗) ≥ 2 randomly. If
𝑖 is the only neighbor of 𝑗, the walker is allowed to backtrack. The walk continues
until the desired sample size is reached. It has been shown that this approach greatly
reduces the bias in the estimated degree distribution compared to traditional random
walks (C.-H. Lee, Xu & Eun, 2012).
Random walk with jumps. An alternative to the non-backtracking approach is random
walk with jumps, where – with a given probability – the next node might be selected
randomly from the whole node set instead of traversing to one of the neighboring
nodes. How this probability is determined can be specific to the application. Ribeiro,
Wang, Murai and Towsley (2012) suggest a degree-proportional probability 𝑤+𝑑𝑤(𝑣) ,
where 𝑤 is a parameter of our choice and 𝑑 (𝑣) is the degree of the last node added to
the sample. Ribeiro et al. (2012) also propose a variety of asymptotically unbiased
estimators for samples obtained using this sampling method.
Frontier of random walkers. With frontier sampling, one starts by selecting 𝑚
random walk seed nodes. Then, at each iteration, a traditional random walk step is
implemented in one of the 𝑚 random walks chosen randomly (Ribeiro & Towsley,
2010). The process is repeated until the desired sample size is reached. This algorithm
has been shown to outperform a wide selection of other traversal-based methods in
estimating the degree correlation on multiple real-life networks (Rozemberczki et al.,
2020).
6.6 Applications of Machine Learning in the Econometrics of

Networks
This section presents three applications in which machine learning methods comple-
ment traditional econometric estimation techniques with network data. Importantly,
we assume that the network data is readily available in these applications and focus
on estimating economic/econometric models.
First, we discuss spatial models and how machine learning can help researchers
learn the spatial weight matrix from the data. Second, we show how specific machine
learning methods can achieve higher prediction accuracy in flow prediction, which
is a commonly studied topic in spatial econometrics. Third, we present an example
where the econometric model of geographically weighted regression has been utilized
to improve the performance of machine learning models. These cases show that
econometrics and machine learning methods do not replace but complement each
other.
6.6.1 Applications of Machine Learning in Spatial Models
Spatial models have been commonly studied in the literature. In international trade,
modeling spatial dependencies is utterly important as geographic proximity has a
significant impact on economic outcomes (e.g., Dell, 2015, Donaldson, 2018, and
Faber, 2014). In epidemiology, spatial connectedness is a major predictor of the
spread of an infectious disease in a population (e.g., Rozemberczki, Scherer, Kiss,
Sarkar & Ferenci, 2021). Even in social interactions, geographic structure largely
determines whom people interact with and how frequently they do so (e.g., Breza
& Chandrasekhar, 2019, Feigenberg, Field & Pande, 2013, and Jackson, Rodriguez-
Barraquer & Tan, 2012). The following discussion highlights some of the workhorse
models of spatial econometrics and presents an application where machine learning
can augment econometric estimations methods. Specifically, we discuss how spatial
dependencies can be learned using two distinct machine learning techniques.
Representing spatial autocorrelation in econometrics. Understanding spatial auto-
correlation is essential for applications that use spatial networks. Spatial autocorrela-
tion is present whenever an economic variable at a given location is correlated with
the values of the same variable at other places. Then, there is a spatial interaction
between the outcomes at different locations. When spatial autocorrelation is present,
standard econometric techniques often fail, and econometricians must account for
such dependencies (Anselin, 2003).
Econometricians commonly specify a particular functional form that generates
the spatial stochastic process to model spatial dependencies, which relates the values
of random variables at different locations. This, in turn, directly determines the
spatial covariance structure. Most papers in the literature apply a non-stochastic
and exogenous spatial weight matrix, which needs to be specified by the researcher.
This is practically an adjacency matrix of a weighted network used in the spatial

domain. Generally, the weight matrix represents the geographic arrangement and
distances in spatial econometrics. For example, the adjacency matrix may correspond
to the inverse of the distance between locations. Alternatively, the elements may
take the value of one when two areas share a common boundary and zero otherwise.
Other specifications for the entries of the spatial weight matrix include economic
distance, road connections, relative GDP, and trade volume between the two locations.
However, researchers should be aware that the weight matrix elements may be
outcomes themselves when using these alternative measures. This can be correlated
with the final outcome causing endogeneity issues in the estimation.
Spatial dependencies can also be determined by directly modeling the covariance
structure using a few parameters or estimating the covariances non-parametrically.
As these methods are less frequently used, the discussion of such techniques is out of
the scope of this chapter. For further details, see Anselin (2003).
The benchmark model in spatial econometrics. The benchmark spatial autocor-
relation model was first presented by Manski (1993). This model does not directly
control for any network characteristics but assumes that outcomes are affected by
own and neighboring agents’ characteristics and actions. It is specified as follows:
𝑦 = 𝜌A𝑦 + X𝛽 + AX𝜃 + 𝑢
(6.8)
𝑢 = 𝜆A𝑢 + 𝜖,
where 𝑦 is the vector of the outcome variable, A is the spatial weight matrix, X is the
matrix of exogenous explanatory variables, 𝑢 is a spatially autocorrelated error term,
and 𝜖 is a random error. The model incorporates three types of interactions:
• an endogenous interaction, where the economic agent’s action depends on its
neighbors’ actions;
• an exogenous interaction, where the agent’s action depends on its neighbors’
observable characteristics;
• a spatial autocorrelation term, driven by correlated unobservable characteristics.
This model is not identifiable, as shown by Manski (1993). However, if we
constrain one of the parameters 𝜌, 𝜃, or 𝜆 to be zero, as proposed by Manski (1993),
the remaining ones can be estimated. When 𝜌 = 0, the model is called the spatial
Durbin error model; when 𝜃 = 0, it is a spatial autoregressive confused model,
and when both 𝜌 = 0 and 𝜃 = 0 are assumed, it is referenced as the spatial error
model. These are rarely used in the literature. Instead, most researchers assume that
𝜆 = 0, which results in the spatial Durbin model, or 𝜃 = 𝜆 = 0, which is the spatial
autoregressive model. As the spatial autoregressive model is the most frequently used
in the literature, our discussion follows this specification.
While this section focuses on geographical connectedness, it is important to note
that the benchmark model is also applicable to problems other than spatial interactions.
The adjacency matrices do not necessarily have to reflect geographical relations. For
example, the model used for peer effects – described in Section 6.3 – takes the form
of the spatial Durbin model.
The spatial autoregressive model (SAR). The SAR model can be specified as follows:
𝑦 = 𝜌A𝑦 + X𝛽 + 𝜖, (6.9)
where 𝜌 is the spatial autoregressive parameter and 𝑦, A, X, and 𝜖 are as described in

Equation (6.8). TheÍweights of the spatial weight matrix are typically row standardized
such that for all 𝑖, 𝑤 𝑖 𝑗 = 1. To simplify the reduced form, let us define
𝑗
S ≡ IN − 𝜌A,
then Equation (6.9) can be expressed as
𝑦 = S−1 X𝛽 + S−1 𝜖 .
Notice that the right-hand side only contains the exogenous node characteristics,
the adjacency matrix, and the error term. The estimation of this model has been
widely studied in the literature, including the maximum likelihood estimation (Ord,
1975), the instrumental variable method (Anselin, 1980), and the generalized method
of moments (L.-F. Lee, 2007, X. Lin & Lee, 2010 and Liu, Lee & Bollinger, 2010).
In some applications, a significant drawback of estimating the SAR model with
the standard econometric techniques is that they assume a non-stochastic adjacency
matrix, A, determined by the researcher. When spatial weights involve endogenous
socioeconomic variables, the elements of the adjacency matrix are likely to be
correlated with the outcome variable. To illustrate this phenomenon, consider the
regression equation where the outcome variable is GDP, and the elements of the
adjacency matrix are trade weights as a share of total trade. Then, the unobservables
that affect the outcome may also be correlated with the weights. As implied by the
gravity model, trade flows can be affected by the unobservable multilateral resistance,
which can be correlated with unobservables in the regression equation (Qu, fei Lee &
Yang, 2021). In such cases, the adjacency matrix is not exogenous, and estimators
that assume the opposite lead to biased estimates (Qu et al., 2021). As Pinkse and
Slade (2010) point out, the endogeneity of spatial weights is a challenging problem,
and there is research to be done in this direction of spatial econometrics.
There have been some efforts in the econometrics literature to make these
dependencies endogenous to the estimation. Using a control function approach, Qu
and fei Lee (2015) developed an estimation method for the case when the entries of
the adjacency matrix are functions of unilateral economic variables, 𝑎 𝑖 𝑗 = ℎ(𝑧𝑖 , 𝑧 𝑗 ).
In their model, the source of endogeneity is the correlation between the error term in
the regression equation for entries of the spatial weight matrix and the error term
in the SAR model. Qu et al. (2021) extend this model with 𝑎 𝑖 𝑗 being determined by
bilateral variables, such as trade flows between countries.
Even when the assumption of the adjacency matrix being non-stochastic is
valid, and we can define simple connections for the spatial weight matrix, in some
applications, it is more difficult to measure the strength of connectivity accurately. For
example, spatial connectedness does not necessarily imply that neighboring locations
interact when using geographic distance in the adjacency matrix. Furthermore, the
interaction strength may not be commensurate with the distance between two zones.
A related issue is when there are multiple types of interactions, such as geographic
distance or number of road connections, and the choice of the specification for
the spatial weight matrix is not apparent. Furthermore, defining the strength of
connections becomes very difficult when the size of the data is large. Determining the
spatial weight structure of dozens of observations is challenging, if not impossible.
In the literature, Bhattacharjee and Jensen-Butler (2013) propose an econometric
approach to identify the spatial weight matrix from the data. The authors show that the
spatial weight matrix is fully identified under the structural constraint of symmetric
spatial weights in the spatial error model. The authors propose a method to estimate the
elements of the spatial weight matrix under symmetry and extend their approach to the
SAR model. Ahrens and Bhattacharjee (2015) propose a two-step LASSO estimator
for the spatial weight matrix in the SAR model, which relies on the identifying
assumptions that the weight matrix is sparse. Lam and Souza (2020) estimate the
optimal spatial weight matrix by obtaining the best linear combination of different
linkages and a sparse adjustment matrix, incorporating errors of misspecification.
The authors use the adaptive LASSO selection method to select which specified
spatial weight matrices to include in the linear combination. When no spatial weight
matrix is specified, the method reduces to estimating a sparse spatial weight matrix.
Learning spatial dependencies with machine learning. The machine learning
literature has focused on developing models for spatiotemporal forecasting, which
can learn spatial dependencies directly from the data. As argued by Rozemberczki,
Scherer, He et al. (2021), neural network-based models have been able to best capture
spatial interactions. There have been various neural network infrastructures proposed
that can capture spatial autocorrelation. While these methods do not provide a
straightforward solution to the endogeneity issue, they can certainly be used for
estimating spatial dependencies. In this chapter, we discuss two of them.
To use the SAR model for spatiotemporal forecasting, machine learning researchers
have incorporated a temporal lag in Equation (6.9). Furthermore, a regularization
term must be added to the estimation equation so that the spatial autoregressive
parameter, 𝜌, and the spatial weight matrix, A, can be identified simultaneously. This
yields the following model:
yt+1 = 𝜌Ayt + Xt+1 𝛽 + 𝛾|A| + 𝑢, (6.10)
where |A| is a 𝑙1 regularizing term and 𝛾 is a tuning parameter set by the researcher.
In this model, controlling for |A| makes the spatial weight matrix more sparse and
also helps identify 𝜌 separately from A. To the best of the authors’ knowledge, there
are no econometric papers in the literature that discuss this specification.
Learning the spatial weight matrix with recurrent neural networks. Ziat, Delasalles,
Denoyer and Gallinari (2017) formalize a recurrent neural network (RNN) architecture
for forecasting time series of spatial processes. The model denoted spatiotemporal
neural network (STNN) learns spatial dependencies through a structured latent
dynamical component. Then, a decoder predicts the actual values from the latent
representations. The main idea behind recurrent neural networks is discussed further
in Chapter 4.
A brief overview of the model is as follows: Assume there are 𝑛 temporal series
with length 𝑇, stacked in X ∈ R𝑇×𝑛 . First, we assume that a spatial weight matrix,
A ∈ R𝑛×𝑛 , is provided. Later, this assumption is relaxed. The model predicts the
series 𝜏 time-steps ahead based on the input variables and the adjacency matrix. The
first component of the model captures the process’s dynamic and is expressed in a
latent space. Assume that each series has a latent space representation in each time
period. Then, the matrix of the latent factors can be denoted by Zt ∈ R𝑛×𝑁 , where
𝑁 is the dimension of the latent space. The latent representation at time 𝑡 + 1, Zt+1 ,
depends on its own latent representation at time 𝑡 (intra-dependency), and on the
latent representation of the neighboring time series at time 𝑡 (inter-dependency). The
dynamical component is expressed as
Zt+1 = ℎ(Zt 𝚯 (0) + AZt 𝚯 (1) ), (6.11)
where 𝚯 (0) ∈ R 𝑁 ×𝑁 and 𝚯 (1) ∈ R 𝑁 ×𝑁 are the parameter matrices to be estimated

and ℎ(.) is a non-linear function.
This specification is different from standard RNN models, where the hidden state
Zt is not only a function of the preceding hidden state Zt−1 but also of the ground
truth values Xt−1 . With this approach, the dynamic of the series is captured entirely
in the latent space. Therefore, spatial dependencies can be modeled explicitly in the
latent factors.
The second component decodes the latent states into a prediction of the series
and is written as 𝑋˜𝑡 = 𝑑 (Zt ) at time t, where 𝑋˜𝑡 is the prediction computed at time
t. Importantly, the latent representations and the parameters of both the dynamic
transition function ℎ(.) and the decoder function 𝑑 (.) can be learned from the data
assuming they are differentiable parametric functions. In the paper, the authors use
ℎ(.) = 𝑡𝑎𝑛ℎ(.) and 𝑑 (.) is a linear function but more complex functions can also be
used.
Then, the learning problem, which captures the dynamic latent space component
and the decoder component can be expressed as
1 ∑︁
𝑑 ∗ , Z∗ , 𝚯 (0)∗ , 𝚯 (1)∗ = arg min Δ(𝑑 (Zt ), 𝑋𝑡 )+
𝑑,Z,𝚯 (0) ,𝚯 (1) 𝑇 𝑡
𝑇−1 (6.12)
1 ∑︁
𝜆 ||Zt+1 − ℎ(Zt 𝚯 (0) + AZt 𝚯 (1) )|| 2 ,
𝑇 𝑡=1
where Δ is a loss function and 𝜆 is a hyperparameter set by cross-validation.

The first term measures the proximity of the predictions 𝑑 (Zt ) and the observed
values 𝑋𝑡 , while the second term captures the latent space dynamics of the series.
The latter term takes its minimum when Zt+1 and ℎ(Zt ) are as close as possible. The
learning problem can be solved with a stochastic gradient descent algorithm.
To incorporate the learning of the spatial weight matrix, Equation (6.11) can be
modified as follows:
Zt+1 = ℎ(Zt 𝚯 (0) + (A ⊙ 𝚪)Zt 𝚯 (1) ), (6.13)
where 𝚪 ∈ R𝑛×𝑛 is a matrix to be learned, A is a pre-defined set of observed relations,

and ⊙ is the element-wise multiplication between two matrices. Here, A can be
a simple adjacency matrix where elements may represent connections, proximity,
distance, etc. Then, the model learns the optimal weight of mutual influence between
the connected sources. In the paper, this model is denoted STNN-R(efining).
Then, the optimization problem over 𝑑, Z, 𝚯 (0) , 𝚯 (1) , 𝚪 can be expressed as
1 ∑︁
𝑑 ∗ , Z∗ , 𝚯 (0)∗ , 𝚯 (1)∗ , 𝚪∗ = arg min Δ(𝑑 (Zt ), 𝑋𝑡 ) + 𝛾|𝚪|+
𝑑,Z,𝚯 (0) ,𝚯 (1) ,𝚪 𝑇 𝑡
𝑇−1 (6.14)
1 ∑︁
𝜆 ||Zt+1 − ℎ(Zt 𝚯 (0) + (A ⊙ 𝚪)Zt 𝚯 (1) )|| 2 ,
𝑇 𝑡=1
where |𝚪| is a 𝑙1 regularizing term, and 𝛾 is a hyper-parameter for tuning the

regularization.
If no prior is available, then removing A from Equation (6.13) gives
Zt+1 = ℎ(Zt 𝚯 (0) + 𝚪Zt 𝚯 (1) ), (6.15)
where 𝚪 represents both the relational structure and the relational weights. This
version of the model is named STNN-D(iscovery), and can be estimated by replacing
the dynamic transition function, ℎ(.) in Equation (6.14) with Equation (6.15). This
specification is the most similar to Equation (6.10) in the latent space. An extension
of the model to multiple relations and further specifications for the experiments can
be found in Ziat et al. (2017).
The model was evaluated on different forecasting problems, such as wind speed,
disease, and car-traffic prediction. The paper shows that the STNN and STNN-R
perform superior to the Vector Autoregressive Model (VAR) and various neural
network architectures, such as recurrent neural networks and dynamic factor graphs
(DFG). However, STNN-D, which does not use any prior information on proximity,
performs worse than STNN and STNN-R in all use cases. The authors also describe
experiments showing the ability of this approach to extract relevant spatial relations.
In particular, when estimating the STTN-D model (when the spatial organization of
the series is not provided), the model can still learn the spatial proximity by assigning
a strong correlation to neighboring observations. These correlations are reflected in
the estimated 𝚪 parameter in Equation (6.15).
Learning the spatial autocorrelation with convolutional neural networks. Spatial
autocorrelation may be learned with convolutional layers in a deep neural network
architecture. The basics of deep neural networks and convolution neural networks
(CNNs) are described in Chapter 4. As pointed out by Dewan, Ganti, Srivatsa and
Stein (2019), the main idea for using convolutional layers comes from the fact that the
SAR model, presented in Equation (6.9), can also be formulated as given in Equation
(6.16), under the assumption that ||𝜌A|| < 1.
∞
∑︁
𝑦= 𝜌 𝑖 A𝑖 (𝛽X + 𝜖). (6.16)
𝑖=0
This mutatis mutandis reflects the invertibility of AR and MA processes. The

expression in Equation (6.16) intuitively means that the explanatory variables and
shocks at any location affect all other directly or indirectly connected locations.
Furthermore, these spatial effects decrease in magnitude with distance from the
location in the network. Using neural networks, this learning can be approximated
with convolution filters of different sizes. As described in Chapter 4, convolution
filters can extract features of the underlying data, including spatial dependencies
between pairs of locations.
Using convolutional layers, Dewan et al. (2019) present a novel Convolutional
AutoEncoder (CAE) model, called the NN-SAR, that can learn the spatiotemporal
structure for prediction. In the model, convolutional layers capture the spatial
dependencies, and the autoencoder retains the most important input variables for
prediction. Autoencoders are discussed in detail in Chapter 4.
Their modeling pipeline follows two steps. For convolutional neural networks, it is
necessary to represent the input data as images. Geographical data can be transformed
into images using geohashing, which creates a rectangular grid of all locations while
keeping the spatial locality of observations. Therefore, images representing the input
and output variables can be constructed for each time period. Then, these images
can be used for time series forecasting using neural networks where the historical
images are used as inputs, and the output is a single image in a given time period.
Time series forecasting using machine learning is further discussed in Chapter 4.
Second, the authors built a deep learning pipeline for predicting the output
image. It consists of an encoder that applies convolutional and max pool layers with
Rectifier Linear Unit (ReLU) activations and a decoder that includes convolutional
and deconvolutional layers. The encoder aims to obtain a compressed representation
of the input variables, while the decoder transforms this representation into the output
image. Additionally, the network consists of skip connections to preserve information
that might have been lost in the encoding process. Their model is similar to the
well-known U-Net architecture by Ronneberger, Fischer and Brox (2015), commonly
used for image segmentation problems.
In Dewan et al. (2019), the authors used data over a large spatial region and time
range to predict missing spatial and temporal values of sensor data. In particular, they
used data on particulate matter, an indicator of air pollution levels, but their method
can also be applied to other spatiotemporal variables. Their results indicate that the
NN-SAR approach is 20% superior on average to the SAR models in predicting
outcomes when the sample size is large.
Dewan et al. (2019) and Ziat et al. (2017) provide examples of how different
neural network architectures can be used to learn spatial dependencies from the data.
While understanding the spatial correlation structure from Dewan et al. (2019) is
more difficult, Ziat et al. (2017) provide a method for explicitly estimating the spatial
weight matrix from the data. As these papers concentrate on forecasting, further
research needs to be conducted on their applicability in cross-sectional settings.
6.6.2 Gravity Models for Flow Prediction
In this section, we focus on a specific domain of spatial models, gravity models.

Several economic studies use gravity models for flow projections, including the World
Economic Outlook by the International Monetary Fund (International Monetary
Fund, 2022). While machine learning methods have not gained widespread use in
such studies, many papers discuss how machine learning models have achieved higher
prediction accuracy. This section provides a rationale for using such methods in
economic studies. First, we briefly discuss the econometric formulation of gravity
models. Second, we summarize results from the machine learning literature regarding
mobility flow prediction using gravity models.
Gravity models have been the workhorse of trade and labor economics as they can
quantify the determinants that affect trade or migration flows between geographic
areas. In gravity models, the adjacency matrix can be treated as if it contained the
reciprocals of distances. A higher weight in the adjacency matrix corresponds to a
lower distance in the gravity model. The general formulation of the gravity model is
expressed as
𝑌𝑖 𝑗𝑡 = 𝑔(𝑋𝑖𝑡 , 𝑋 𝑗𝑡 , 𝑖, 𝑗, 𝑡),
where 𝑌𝑖 𝑗𝑡 is the bilateral outcome between country 𝑖 and country 𝑗 at time 𝑡 (response
variable), 𝑋𝑖𝑡 and 𝑋 𝑗𝑡 are the sets of possible predictors from both countries, and
the set {𝑖, 𝑗, 𝑡} refers to a variety of controls on all three dimensions (Anderson &
van Wincoop, 2001). With some simplification, in most specifications, the gravity
models assume that the flows between two locations increase with the population of
locations but decrease with the distance between them.
Matyas (1997) introduced the most widely used specification of the gravity model
in trade economics:
𝑙𝑛𝐸 𝑋 𝑃𝑖 𝑗𝑡 = 𝛼𝑖 + 𝛾 𝑗 + 𝜆 𝑡 + 𝛽1 𝑙𝑛𝑌𝑖𝑡 + 𝛽2 𝑙𝑛𝑌 𝑗𝑡 + 𝛽3 𝐷 𝐼𝑆𝑇𝑖 𝑗 + ... + 𝑢 𝑖 𝑗𝑡 ,
where 𝐸 𝑋 𝑃𝑖 𝑗𝑡 is the volume of trade (exports) from country 𝑖 to country 𝑗 at time 𝑡;

𝑌𝑖𝑡 is the GDP in country 𝑖 at time 𝑡, and the same for 𝑌 𝑗𝑡 for country j; 𝐷 𝐼𝑆𝑇𝑖 𝑗 is the
distance between the countries 𝑖 and 𝑗; 𝛼𝑖 and 𝛾 𝑗 are the origin and target country
fixed effects respectively; 𝜆 𝑡 is the time (business cycle) effect; and 𝑢 𝑖 𝑗𝑡 is a white
noise disturbance term.
The model specified can be estimated
Í by OLS
Í when weÍinclude a constant term
and set the identifying restrictions: 𝛼𝑖 = 1, 𝛾 𝑗 = 1 and 𝜆 𝑡 = 1. The model can
also be estimated with several other techniques, including the penalized regression
(discussed in Chapter 1 and empirically tested by H. Lin et al., 2019).
Gravity models have also been studied in the machine learning literature. Machine
learning methods can capture non-linear relationships between the explanatory
variables and trade/mobility flows, which may better characterize the underlying
structure. These models are also more flexible and capable of generating more realistic
trade and mobility flows. The literature on the topic has mainly focused on comparing
standard econometric methods with neural networks. The concept of neural networks
and their mathematical formulation are discussed in Chapter 4.
Fischer and Gopal (1994), Gopal and Fischer (1996), and Fischer (1998) compare
how accurately gravity models and neural networks can predict inter-regional tele-
communications flows in Austria. To evaluate the models’ performance, the authors
use two different measures – the average relative variance (ARV) and the coefficient
of determination (𝑅 2 ) – and also perform residual analysis. The papers show that the
neural network model approach achieves higher predictive accuracy than the classical
regression. In Fischer (1998), the neural network-based approach achieved 14% lower
ARV and 10% higher 𝑅 2 , stable across different trials.
In a different domain, Tillema, Zuilekom and van Maarseveen (2006) compare the
performance of neural networks and gravity models in trip distribution modeling. The
most commonly used methods in the literature to model trip distributions between
origins and destinations have been gravity models – they are the basis for estimating
trip distribution in a four-step model of transportation (McNally, 2000). The authors
use synthetic and real-world data to perform statistical analyses, which help determine
the necessary sample sizes to obtain statistically significant results. The results show
that neural networks attain higher prediction accuracy than gravity models with small
sample sizes, irrespective of using synthesized or real-world data. Furthermore, the
authors establish that the necessary sample size for statistically significant results is
forty times lower for neural networks.
Pourebrahim, Sultana, Thill and Mohanty (2018) also study the predictive per-
formance of neural networks and gravity models when forecasting trip distribution.
They contribute to the literature by utilizing social media data – the number of tweets
posted in origin and destination locations – besides using standard input variables
such as employment, population, and distance. Their goal is to predict commuter trip
distribution on a small spatial scale – commuting within cities – using geolocated
Twitter post data. Gravity models have traditionally used population and distance as
key predicting factors. However, mobility patterns within cities may not be predicted
reliably without additional predictive factors. The paper compares the performance of
neural networks and gravity models, both trying to predict home-work flows using the
same data set. The results suggest that social media data can improve the modeling
of commuter trip distribution and is a step forward to developing dynamic models
besides using static socioeconomic factors. Furthermore, standard gravity models are
outperformed by neural networks in terms of 𝑅 2 . This indicates that neural networks
better fit the data and are superior for prediction in this application.
Simini, Barlacchi, Luca and Pappalardo (2020) propose a neural network-based
deep gravity model, which predicts mobility flows using geographic data. Specifically,
the authors estimate mobility flows between regions with input variables such as
areas of different land use classes, length of road networks, and the number of health
and education facilities. In brief, the deep gravity model uses these input features
to compute the probability 𝑝 𝑖, 𝑗 that a trip originated at location 𝑙𝑖 has a destination
location 𝑙 𝑗 for all possible locations in the region. The authors use the common
part of commuters (CPC) evaluation metric to evaluate the model, which computes
the similarity between actual and generated flows. Their results suggest that neural
network-based models can predict mobility flows significantly better than traditional
gravity models, shallow neural network models, and models that do not use extensive
geographic data. In areas with high population density, where prediction is more
difficult due to the many locations, the machine learning-based approach outperforms
the gravity model by 350% in CPC. Finally, the authors also show that the deep gravity
model generalizes well to other locations not used in training. This is achieved by
the model being tested to predict flows in a region non-overlapping with the training
regions.
The results from the machine learning literature provide strong evidence that
neural network-based approaches are better suited for flow prediction than stand-
ard econometric techniques. Therefore, such techniques should be part of applied
economists’ toolbox when incorporating flow projections in economic studies.
6.6.3 The Geographically Weighted Regression Model and ML
This section presents the econometric method of geographically weighted regression

(GWR). Our discussion closely follows Fotheringham, Brunsdon and Charlton (2002).
Then, we discuss a paper from the machine learning literature that, by incorporating
this method, improves the predictive performance of several machine learning
techniques.
The geographically weighted regression is an estimation procedure that takes into
account the spatial distance of the data points in the sample to estimate local variations
of the regression coefficients (Brunsdon, Fotheringham & Charlton, 1996). Contrary
to traditional estimation methods like the OLS, GWR applies local regressions
on neighboring points for each observation in the sample. Therefore, it allows for
estimating spatially varying coefficients and uncovers local features that the global
approach cannot measure. If the local coefficients vary by the different estimations in
space and move away from their global values, this may indicate that, for example,
non-stationarity is present.
In GWR estimation, a separate regression is conducted for all data points. Then,
the number of regression equations and estimates equals the number of territorial
units. The sample consists of the regression point and the neighboring data points
within a defined distance in each regression. Their distance from the regression point
weights these data points – observations that are spatially closer to the regression
point are assigned a larger weight (Brunsdon et al., 1996).
To estimate a geographically weighted regression model, researchers have to
choose two parameters that govern local regressions. First, one must define the
weighting scheme that handles how nearby observations are weighted in the local
regression. Generally, data points closer to the regression point are assigned larger
weights, continuously decreasing by distance. Second, researchers have to choose
a kernel that may either be fixed, using a fixed radius set by the econometrician,
or adaptive, using a changing radius. Applying an adaptive kernel is particularly
useful if there are areas in which data points are more sparsely located. Such a kernel
provides more robust estimates in areas with less density as its bandwidth changes
with the number of neighboring data points. Notably, the estimated coefficients are
particularly sensitive to the choice of kernel shape and bandwidth as they determine
which points are included in the local regression.
Weighting scheme options for GWR. The simplest option for local weighting is to
include only those observations in the local regression which are within a 𝑏 radius of
the examined point. This is called a uniform kernel with constrained support and can
be expressed as
𝑤 𝑖 𝑗 = 1 if 𝑑𝑖 𝑗 < 𝑏; 𝑤 𝑖 𝑗 = 0 otherwise, (6.17)
where 𝑖 is the examined point in a local regression, 𝑑𝑖 𝑗 is the distance of point 𝑗 from
point 𝑖 in space, and 𝑏 is the bandwidth chosen by the researcher.
However, this approach is very sensitive to the choice of bandwidth, and estimated
parameters may significantly change if borderline observations are included or
excluded. A more robust approach to this issue weights observations based on their
Euclidean distance from the analyzed point and is called the Gaussian kernel:
1
𝑤 𝑖 𝑗 = 𝑒𝑥 𝑝(− (𝑑𝑖 𝑗 /𝑏) 2 ),
2
where the parameters are as defined in Equation (6.17). Here, the weights are
continuously decreasing as distance increases.
A combination of the two methods above is given by
𝑤 𝑖 𝑗 = [1 − (𝑑𝑖 𝑗 /𝑏) 2 ] 2 if 𝑑𝑖 𝑗 < 𝑏; 𝑤 𝑖 𝑗 = 0 otherwise.
This formulation assigns decreasing weights until the bandwidth is reached and zero
weight beyond.
Adaptive kernels consider the density around the regression point to estimate each
local regression with the same sample size. When we order observations by their
distance to the regression point, an adaptive kernel may be formulated as
𝑤 𝑖 𝑗 = 𝑒𝑥 𝑝(−𝑅𝑖 𝑗 /𝑏),
where 𝑅𝑖 𝑗 is the rank number of point 𝑗 from point 𝑖 and 𝑏 is the bandwidth. Then
the weighting scheme disregards the actual distance between data points and is only a
function of their ordering relative to each other.
Finally, we may also define the weight by taking into account the 𝑁 number of
nearest neighbors:
𝑤 𝑖 𝑗 = [1 − (𝑑𝑖 𝑗 /𝑏) 2 ] 2 if 𝑗 is one of the 𝑁 nearest neighbors of 𝑖,

and 𝑏 is the distance of the N-th nearest neighbor,
𝑤 𝑖 𝑗 = 0 otherwise.
where 𝑁 is a parameter and 𝑑𝑖 𝑗 is as defined in Equation (6.17). Therefore, a fixed

number of nearest neighbors is included in the regression. The parameter 𝑁 needs to
be defined or calibrated for this weighting scheme.
Finding the optimal bandwidth for GWR. The optimal bandwidth can be calibrated
using cross validation (CV). One commonly used method is to minimize the following
function with respect to the bandwidth 𝑏:
𝑛
∑︁
𝐶𝑉 = [𝑦 𝑖 − 𝑦ˆ 𝑗≠𝑖 (𝑏)] 2 ,
𝑖=1
where 𝑏 is the bandwidth and 𝑦ˆ 𝑗≠𝑖 is the estimated value of the dependent variable if
point 𝑖 is left out of the calibration. This condition is necessary to avoid zero optimal
bandwidth (Fotheringham et al., 2002).
Other approaches to finding the optimal 𝑏 parameter include minimizing the
Akaike information criterion (AIC) or the Schwarz criterion (SC). Such metrics
also penalize the model for having many explanatory variables, which helps reduce
overfitting. As discussed in Chapters 1 and 2, these criteria can also be used for model
selection in Ridge, LASSO, and Elastic Net models.
Estimating the GWR model. Once the sample has been constructed for each local
regression, the GWR specification can be estimated. The equation of the geographically
weighted regression is expressed as
∑︁
𝑦 𝑖 = 𝛽0 (𝑢 𝑖 , 𝑣 𝑖 ) + 𝛽 𝑘 (𝑢 𝑖 , 𝑣 𝑖 )𝑥𝑖𝑘 + 𝜖𝑖 ,
𝑘
where (𝑢 𝑖 , 𝑣 𝑖 ) is the geographical coordinates of point 𝑖 and 𝛽 𝑘 (𝑢 𝑖 , 𝑣 𝑖 ) is the estimated

value of the continuous function 𝛽 𝑘 (𝑢, 𝑣) in point 𝑖 (Fotheringham et al., 2002).
With the weighted least squares method, the following solution can be obtained:
ˆ 𝑖 , 𝑣 𝑖 ) = (X𝑇 W(𝑢 𝑖 , 𝑣 𝑖 )X) −1 X𝑇 W(𝑢 𝑖 , 𝑣 𝑖 )𝑦,
𝛽(𝑢 (6.18)
where 𝑦 is the output vector and X is the input matrix of all examples. Furthermore,
W(𝑢 𝑖 , 𝑣 𝑖 ) is the weight matrix of the target location, (𝑢 𝑖 , 𝑣 𝑖 ), whose off-diagonal
elements are zero and diagonal elements measure the spatial weights defined by the
weighting scheme.
The geographically weighted machine learning model. L. Li (2019) combines the
geographically weighted regression model (GWR) with neural networks, XGBoost,
and random forest classifiers to improve high-resolution spatiotemporal wind speed
predictions in China. Due to the size of the country and the heterogeneity between its
regions, modeling meteorological factors at high spatial resolution is a particularly
challenging problem. By incorporating the GWR method, which captures local

variability, the authors show that prediction accuracy can be improved compared to
the performance of the base learners.
The outcome variable in the paper is daily wind speed data collected from several
wind speed monitoring stations across the country. The covariates in the estimation
included coordinates, elevation, day of the year, and region. Furthermore, the author
also utilized coarse spatial resolution reanalysis data (i.e., climate data), which
provides reliable estimates of wind speed and planetary boundary layer height (a
factor for surface wind gradient and related to wind speed) at a larger scale.
The author developed a two-stage approach. They built a geographically weighted
ensemble machine learning model in the first stage. They first trained three base
learners, an autoencoder-based deep residual network, XGBoost, and random forest,
predicting the outcome with the collected covariates. These models are based on
three weakly correlated and fundamentally different algorithms. Therefore, their
combination can provide better ensemble predictions in theory. The author combined
the projections with the GWR method so that the model also incorporates spatial
autocorrelation and heterogeneity for boosting forecasts. In brief, the GWR helps
obtain the optimal weights of the three models. Furthermore, it provides spatially
varying coefficients for the base learners, which allows for spatial autocorrelation
and heterogeneity in the estimation.
In the GWR estimation, there are three explanatory variables (the predictions of
the three base learners), and so the geographically weighted regression equation is
written as
3
∑︁
𝑦 𝑖 = 𝛽0 (𝑢 𝑖 , 𝑣 𝑖 ) + 𝛽 𝑘 (𝑢 𝑖 , 𝑣 𝑖 )𝑥𝑖𝑘 + 𝜖𝑖 ,
𝑘=1
where (𝑢 𝑖 , 𝑣 𝑖 ) are the coordinates of the 𝑖th sample, 𝛽 𝑘 (𝑢 𝑖 , 𝑣 𝑖 ) is the regression

coefficient for the 𝑘th base prediction, 𝑥 𝑖𝑘 is the predicted value by the 𝑘th base
learner, and 𝜖𝑖 is random noise (𝜖𝑖 ∼ 𝑁 (0, 1)).
The Gaussian kernel was used to quantify the weight matrix for this study. These
weights can be applied for the weighted least squares method in Equation (6.18):
𝑤 𝑖 𝑗 = 𝑒𝑥 𝑝(−(𝑑𝑖 𝑗 /𝑏) 2 ),
where 𝑏 is the bandwidth and 𝑑𝑖 𝑗 is the distance between locations 𝑖 and 𝑗.

In the second stage, the author used a deep residual network to perform a
downscaling that matches the wind speed from coarse resolution meteorological
reanalysis data with the average predicted wind speed at high resolution inferred from
the first stage. This reduces bias and yields more realistic spatial variation (smoothing)
in the predictions. Furthermore, it also ensures that projections are consistent with
the observed data at a more coarse resolution. The description of the method is out of
the scope of this chapter and therefore left out of this discussion.
The geographically weighted regression ensemble achieved a 12–16% improvement
in 𝑅 2 and lower Root Mean Square Error (RMSE) compared to individual learners’
predictions. Therefore, GWR can effectively capture the local variation of the target
variable at high-resolution and account for the spatial autocorrelation present in

the data. Overall, the paper provides a clear example of how machine learning and
econometrics can be combined to improve performance on challenging prediction
problems.
This chapter has focused on overcoming the challenges of analyzing network structured
data. The high dimensionality of such data sets is one of the fundamental problems
when working with them. Applied works relying on networks often solve this issue by
including simple aggregations of the adjacency matrix in their analyses. In Section
6.4 we discuss an alternative to this approach, namely graph dimensionality reduction.
While these algorithms have a proven track record in the computer science literature,
they have received limited attention from economists and econometricians. Analyzing
whether and how information extracted using these algorithms changes our knowledge
about economic and social networks is a promising area for future research.
Next, in Section 6.5, we turned our attention towards questions related to sampling
such data sets. Many modern-day networks (e.g., web graphs or social networks)
have hundreds of millions of nodes and billions of edges. The size of these data sets
renders the corresponding analysis computationally intractable. To deal with this
issue, we discuss the best-known sampling methods presented in the literature. They
can be used to preserve the salient information encoded in networks. Choosing the
proper sampling method should always be driven by the analysis. However, there
is no comprehensive study on the choice of the appropriate approach to the best
of our knowledge. Theoretical properties of these methods are usually derived by
assuming a specific network structure. However, these assumptions are unlikely to
hold for real-life networks. Rozemberczki et al. (2020) present a study evaluating
many sampling algorithms using empirical data. Their results show that there is no
obvious choice, although some algorithms consistently outperform others in terms
of preserving population statistics. The way the selection of the sampling method
affects estimation outcomes is also a potential future research domain.
In Section 6.6, we present three research areas in spatial econometrics that have
been analyzed in both the econometrics and the machine learning literature. We aim to
show how machine learning can enhance econometric analysis related to network data.
First, learning spatial autocorrelations from the data is essential when the strength of
spatial interactions cannot be explicitly measured. As estimates of spatial econometrics
models are sensitive to the choice of the spatial weight matrix, researchers may fail
to recover the spatial covariance structure when such spatial interactions are not
directly learned from the data. In the econometrics literature, estimating the spatial
weight matrix relies on identifying assumptions, such as structural constraints or the
sparsity of the weight matrix. This chapter shows advances in the machine learning
literature that rely on less stringent assumptions to extract such information from the
data. Incorporating these algorithms into econometric analyses has vast potential and
remains an important avenue for research.
Second, forecasting in the spatial domain has been a widely studied area in
economics. There are several studies by international institutions that use econometric
methods to predict, for example, future trade patterns, flows of goods, and mobility. On
the other hand, machine learning methods have not yet gained widespread recognition
among economists who work on forecasting macroeconomic variables. We present
papers that show how machine learning methods can achieve higher prediction accur-
acy. In forecasting problems, where establishing estimators’ econometric properties
is less important, machine learning algorithms often prove superior to standard
econometric approaches. Finally, we have introduced the geographically weighted
regression and presented an example where an econometric estimation technique
can improve the forecasting performance of machine learning algorithms applied
to spatiotemporal data. This empirical application suggests that spatially related
economic variables predicted by machine learning models may be modeled with
econometric techniques to improve prediction accuracy. The approaches presented in
the last section suggest that combining machine learning and traditional econometrics
has an untapped potential in forecasting variables relevant to policy makers.
References
Acemoglu, D., Carvalho, V. M., Ozdaglar, A. & Tahbaz-Salehi, A. (2012). The

network origins of aggregate fluctuations. Econometrica, 80(5), 1977–2016.
Adamic, L. A., Lukose, R. M., Puniyani, A. R. & Huberman, B. A. (2001). Search in
power-law networks. Phys. Rev. E, 64, 046135.
Ahmed, N. K., Neville, J. & Kompella, R. (2013). Network sampling: From static
to streaming graphs. ACM Transactions on Knowledge Discovery from Data
(TKDD), 8(2), 1–56.
Ahrens, A. & Bhattacharjee, A. (2015). Two-step lasso estimation of the spatial
weights matrix. Econometrics, 3(1), 1–28.
Anderson, J. E. & van Wincoop, E. (2001). Gravity with gravitas: A solution to
the border puzzle (Working Paper No. 8079). National Bureau of Economic
Research.
Anselin, L. (1980). Estimation methods for spatial autoregressive structures: A study
in spatial econometrics. Program in Urban and Regional Studies, Cornell
University.
Anselin, L. (2003). Spatial econometrics. In A companion to theoretical econometrics
(pp. 310–330). John Wiley & Sons.
Balasubramanian, M. & Schwartz, E. L. (2002). The isomap algorithm and topological
stability. Science, 295(5552), 7–7.
Ballester, C., Calvo-Armengol, A. & Zenou, Y. (2006). Who’s who in networks.
wanted: The key player. Econometrica, 74(5), 1403–1417.
References 211
Belkin, M. & Niyogi, P. (2002). Laplacian eigenmaps and spectral techniques for
embedding and clustering. In Advances in neural information processing
systems (pp. 585–591).
Bhattacharjee, A. & Jensen-Butler, C. (2013). Estimation of the spatial weights
matrix under structural constraints. Regional Science and Urban Economics,
43(4), 617–634.
Breza, E. & Chandrasekhar, A. G. (2019). Social networks, reputation, and
commitment: Evidence from a savings monitors experiment. Econometrica,
87(1), 175–216.
Brunsdon, C., Fotheringham, A. S. & Charlton, M. E. (1996). Geographically weighted
regression: A method for exploring spatial nonstationarity. Geographical
Analysis, 28(4), 281–298.
Cai, H., Zheng, V. W. & Chang, K. (2018). A comprehensive survey of graph
embedding: Problems, techniques, and applications. IEEE Transactions on
Knowledge and Data Engineering, 30(09), 1616–1637.
Chandrasekhar, A. G. & Lewis, R. (2016). Econometrics of sampled networks.
(Retrieved: 02.02.2022 from https://stanford.edu/ arungc/CL.pdf)
Crépon, B., Devoto, F., Duflo, E. & Parienté, W. (2015). Estimating the impact of
microcredit on those who take it up: Evidence from a randomized experiment
in morocco. American Economic Journal: Applied Economics, 7(1), 123–50.
de Lara, N. & Edouard, P. (2018). A simple baseline algorithm for graph classification.
arXiv, abs/1810.09155.
Dell, M. (2015). Trafficking networks and the mexican drug war. American Economic
Review, 105(6), 1738–1779.
Dewan, P., Ganti, R., Srivatsa, M. & Stein, S. (2019). Nn-sar: A neural network
approach for spatial autoregression. In 2019 ieee international conference on
pervasive computing and communications workshops (percom workshops) (pp.
783–789).
Doerr, C. & Blenn, N. (2013). Metric convergence in social network sampling. In
Proceedings of the 5th acm workshop on hotplanet (p. 45–50). New York, NY,
USA: Association for Computing Machinery.
Donaldson, D. (2018). Railroads of the raj: Estimating the impact of transportation
infrastructure. American Economic Review, 108(4-5), 899–934.
Easley, D. & Kleinberg, J. (2010). Networks, crowds, and markets (Vol. 8). Cambridge
university press Cambridge.
Faber, B. (2014). Trade integration, market size, and industrialization: Evidence
from china’s national trunk highway system. The Review of Economic Studies,
81(3), 1046–1070.
Feigenberg, B., Field, E. & Pande, R. (2013). The economic returns to social
interaction: Experimental evidence from microfinance. The Review of Economic
Studies, 80(4), 1459–1483.
Ferrali, R., Grossman, G., Platas, M. R. & Rodden, J. (2020). It takes a village: Peer
effects and externalities in technology adoption. American Journal of Political
Science, 64(3), 536–553.
Fischer, M. (1998). Computational neural networks: An attractive class of mathemat-
ical models for transportation research. In V. Himanen, P. Nijkamp, A. Reggiani

& J. Raitio (Eds.), Neural networks in transport applications (pp. 3–20).
Fischer, M. & Gopal, S. (1994). Artificial neural networks: A new approach to
modeling interregional telecommunication flows. Journal of Regional Science,
34, 503 – 527.
Fotheringham, A., Brunsdon, C. & Charlton, M. (2002). Geographically weighted
regression: The analysis of spatially varying relationships. John Wiley &
Sons.
Gjoka, M., Kurant, M., Butts, C. T. & Markopoulou, A. (2010). Walking in facebook:
A case study of unbiased sampling of osns. In 2010 proceedings ieee infocom
(pp. 1–9).
Gonzalez, J. E., Low, Y., Gu, H., Bickson, D. & Guestrin, C. (2012). Powergraph:
Distributed graph-parallel computation on natural graphs. In Presented as
part of the 10th {USENIX} symposium on operating systems design and
implementation ({OSDI} 12) (pp. 17–30).
Goodman, L. A. (1961). Snowball sampling. The annals of mathematical statistics,
148–170.
Gopal, S. & Fischer, M. M. (1996). Learning in single hidden-layer feedforward
network models: Backpropagation in a spatial interaction modeling context.
Geographical Analysis, 28(1), 38–55.
Grover, A. & Leskovec, J. (2016). node2vec: Scalable feature learning for networks.
In Proceedings of the 22nd acm sigkdd international conference on knowledge
discovery and data mining (pp. 855–864).
Hu, P. & Lau, W. C. (2013). A survey and taxonomy of graph sampling. arXiv,
abs/1308.5865.
Hübler, C., Kriegel, H.-P., Borgwardt, K. & Ghahramani, Z. (2008). Metropolis
algorithms for representative subgraph sampling. In 2008 eighth ieee interna-
tional conference on data mining (pp. 283–292).
International Monetary Fund. (2022). World economic outlook, april 2022: War sets
back the global recovery. USA: International Monetary Fund.
Jackson, M. O. (2010). Social and economic networks. Princeton University Press.
Jackson, M. O., Rodriguez-Barraquer, T. & Tan, X. (2012). Social capital and
social quilts: Network patterns of favor exchange. American Economic Review,
102(5), 1857–97.
Kang, U., Tsourakakis, C. E. & Faloutsos, C. (2009). Pegasus: A peta-scale
graph mining system implementation and observations. In 2009 ninth ieee
international conference on data mining (pp. 229–238).
Krishnamurthy, V., Faloutsos, M., Chrobak, M., Lao, L., Cui, J. H. & Percus, A. G.
(2005). Reducing large internet topologies for faster simulations. In Proceedings
of the 4th ifip-tc6 international conference on networking technologies, services,
and protocols; performance of computer and communication networks; mobile
and wireless communication systems (p. 328–341). Berlin, Heidelberg: Springer-
Verlag.
Lam, C. & Souza, P. C. (2020). Estimation and selection of spatial weight matrix in a
spatial lag model. Journal of Business & Economic Statistics, 38(3), 693–710.
References 213
Lazarsfeld, P. F. & Merton, R. K. (1954). Friendship as a social process: a substantive

and method-ological analysis. Freedom and Control in Modern Society, 18–66.
Lee, C.-H., Xu, X. & Eun, D. Y. (2012). Beyond random walk and metropolis-hastings
samplers: why you should not backtrack for unbiased graph sampling. ACM
SIGMETRICS Performance evaluation review, 40(1), 319–330.
Lee, L.-F. (2007). Gmm and 2sls estimation of mixed regressive, spatial autoregressive
models. Journal of Econometrics, 137(2), 489–514.
Leskovec, J. & Faloutsos, C. (2006). Sampling from large graphs. In Proceedings
of the 12th acm sigkdd international conference on knowledge discovery and
data mining (pp. 631–636).
Leskovec, J., Kleinberg, J. & Faloutsos, C. (2005). Graphs over time: densification
laws, shrinking diameters and possible explanations. In Proceedings of the
eleventh acm sigkdd international conference on knowledge discovery in data
mining (pp. 177–187).
Li, L. (2019). Geographically weighted machine learning and downscaling for
high-resolution spatiotemporal estimations of wind speed. Remote Sensing,
11(11).
Li, R.-H., Yu, J. X., Qin, L., Mao, R. & Jin, T. (2015). On random walk based graph
sampling. In 2015 ieee 31st international conference on data engineering (pp.
927–938).
Lin, H., Hong, H. G., Yang, B., Liu, W., Zhang, Y., Fan, G.-Z. & Li, Y. (2019).
Nonparametric Time-Varying Coefficient Models for Panel Data. Statistics in
Biosciences, 11(3), 548–566.
Lin, X. & Lee, L.-F. (2010). Gmm estimation of spatial autoregressive models with
unknown heteroskedasticity. Journal of Econometrics, 157(1), 34–52.
Liu, X., Lee, L.-f. & Bollinger, C. R. (2010). An efficient GMM estimator of spatial
autoregressive models. Journal of Econometrics, 159(2), 303–319.
Maiya, A. S. & Berger-Wolf, T. Y. (2010). Sampling community structure. In
Proceedings of the 19th international conference on world wide web (pp.
701–710).
Manski, C. F. (1993). Identification of endogenous social effects: The reflection
problem. The Review of Economic Studies, 60(3), 531–542.
Matyas, L. (1997). Proper econometric specification of the gravity model. The World
Economy, 20(3), 363-368.
McNally, M. (2000). The four step model. In D. Hensher & K. Button (Eds.),
Handbook of transport modelling (Vol. 1, pp. 35–53). Elsevier.
Mikolov, T., Sutskever, I., Chen, K., Corrado, G. & Dean, J. (2013). Distributed
representations of words and phrases and their compositionality.
Newman, M. E. J. (2002). Assortative mixing in networks. Phys. Rev. Lett., 89,
208701.
Ord, K. (1975). Estimation methods for models of spatial interaction. Journal of the
American Statistical Association, 70(349), 120–126.
Ortega, F. & Peri, G. (2013). The effect of income and immigration policies on
international migration. Migration Studies, 1(1), 47–74.
Page, L., Brin, S., Motwani, R. & Winograd, T. (1999). The pagerank citation
ranking: Bringing order to the web. (Technical Report No. 1999-66). Stanford
InfoLab. (Previous number = SIDL-WP-1999-0120)
Perozzi, B., Al-Rfou, R. & Skiena, S. (2014). Deepwalk: Online learning of social
representations. In Proceedings of the 20th acm sigkdd international conference
on knowledge discovery and data mining (pp. 701–710).
Pinkse, J. & Slade, M. E. (2010). The future of spatial econometrics. Journal of
Regional Science, 50(1), 103–117.
Pourebrahim, N., Sultana, S., Thill, J.-C. & Mohanty, S. (2018). Enhancing trip
distribution prediction with twitter data: Comparison of neural network and
gravity models. In Proceedings of the 2nd acm sigspatial international
workshop on ai for geographic knowledge discovery (p. 5–8). New York, NY,
USA: Association for Computing Machinery.
Qu, X. & fei Lee, L. (2015). Estimating a spatial autoregressive model with an
endogenous spatial weight matrix. Journal of Econometrics, 184(2), 209–232.
Qu, X., fei Lee, L. & Yang, C. (2021). Estimation of a sar model with endogenous
spatial weights constructed by bilateral variables. Journal of Econometrics,
221(1), 180–197.
Rezvanian, A. & Meybodi, M. R. (2015). Sampling social networks using shortest
paths. Physica A: Statistical Mechanics and its Applications, 424, 254–268.
Ribeiro, B. & Towsley, D. (2010). Estimating and sampling graphs with multidimen-
sional random walks. In Proceedings of the 10th acm sigcomm conference on
internet measurement (pp. 390–403).
Ribeiro, B., Wang, P., Murai, F. & Towsley, D. (2012). Sampling directed graphs
with random walks. In 2012 proceedings ieee infocom (p. 1692-1700).
Ronneberger, O., Fischer, P. & Brox, T. (2015). U-net: Convolutional networks for
biomedical image segmentation. In N. Navab, J. Hornegger, W. M. Wells
& A. F. Frangi (Eds.), Medical image computing and computer-assisted
intervention – miccai 2015 (pp. 234–241). Cham: Springer International
Publishing.
Rozemberczki, B., Kiss, O. & Sarkar, R. (2020). Little Ball of Fur: A Python Library
for Graph Sampling. In Proceedings of the 29th acm international conference
on information and knowledge management (cikm ’20).
Rozemberczki, B. & Sarkar, R. (2018). Fast sequence-based embedding with diffusion
graphs. In International workshop on complex networks (pp. 99–107).
Rozemberczki, B., Scherer, P., He, Y., Panagopoulos, G., Riedel, A., Astefanoaei,
M., . . . Sarkar, R. (2021). Pytorch geometric temporal: Spatiotemporal
signal processing with neural machine learning models. In Proceedings of the
30th acm international conference on information and knowledge management
(p. 4564—4573). New York, NY, USA: Association for Computing Machinery.
Rozemberczki, B., Scherer, P., Kiss, O., Sarkar, R. & Ferenci, T. (2021). Chickenpox
cases in hungary: a benchmark dataset for spatiotemporal signal processing
with graph neural networks. arXiv, abs/2102.08100.
Sacerdote, B. (2001). Peer Effects with Random Assignment: Results for Dartmouth
Roommates. The Quarterly Journal of Economics, 116(2), 681–704.
References 215
Simini, F., Barlacchi, G., Luca, M. & Pappalardo, L. (2020). Deep gravity: en-
hancing mobility flows generation with deep neural networks and geographic
information. arXiv, abs/2012.00489.
Stumpf, M. P. H., Wiuf, C. & May, R. M. (2005). Subnets of scale-free networks are
not scale-free: Sampling properties of networks. Proceedings of the National
Academy of Sciences, 102(12), 4221–4224.
Stutzbach, D., Rejaie, R., Duffield, N., Sen, S. & Willinger, W. (2008). On unbiased
sampling for unstructured peer-to-peer networks. IEEE/ACM Transactions on
Networking, 17(2), 377–390.
Tang, J., Qu, M., Wang, M., Zhang, M., Yan, J. & Mei, Q. (2015). Line: Large-scale
information network embedding. In Proceedings of the 24th international
conference on world wide web (pp. 1067–1077).
Tillema, F., Zuilekom, K. M. V. & van Maarseveen, M. (2006). Comparison of
neural networks and gravity models in trip distribution. Comput. Aided Civ.
Infrastructure Eng., 21, 104–119.
Torres, L., Chan, K. S. & Eliassi-Rad, T. (2020). Glee: Geometric laplacian eigenmap
embedding. Journal of Complex Networks, 8(2).
Ziat, A., Delasalles, E., Denoyer, L. & Gallinari, P. (2017). Spatio-temporal neural
networks for space-time series forecasting and relations discovery. 2017 IEEE
International Conference on Data Mining (ICDM), 705–714.
Chapter 7
Fairness in Machine Learning and Econometrics
Samuele Centorrino, Jean-Pierre Florens and Jean-Michel Loubes
Abstract A supervised machine learning algorithm determines a model from a

learning sample that will be used to predict new observations. To this end, it
aggregates individual characteristics of the observations of the learning sample.
But this information aggregation does not consider any potential selection on
unobservables and any status quo biases which may be contained in the training
sample. The latter bias has raised concerns around the so-called fairness of machine
learning algorithms, especially towards disadvantaged groups. In this chapter, we
review the issue of fairness in machine learning through the lenses of structural
econometrics models in which the unknown index is the solution of a functional
equation and issues of endogeneity are explicitly taken into account. We model
fairness as a linear operator whose null space contains the set of strictly fair indexes.
A fair solution is obtained by projecting the unconstrained index into the null space of
this operator or by directly finding the closest solution of the functional equation into
this null space. We also acknowledge that policymakers may incur costs when moving
away from the status quo. Approximate fairness is thus introduced as an intermediate
set-up between the status quo and a fully fair solution via a fairness-specific penalty
in the objective function of the learning model.
Samuele Centorrino
Stony Brook University, Stony Brook, NY, USA, e-mail: samuele.centorrino@stonybrook.edu
Jean-Pierre Florens
Toulouse School of Economics, University of Toulouse Capitole, Toulouse, France, e-mail: jean
-pierre.florens@tse-fr.eu
Jean-Michel Loubes B
Université Paul Sabatier, Institut de Mathématiques de Toulouse, Toulouse, France, e-mail: loubes@
math.univ-toulouse.fr
218 Centorrino et al.
7.1 Introduction
Fairness has been a growing field of research in Machine Learning, Statistics, and
Economics over recent years. Such work aims to monitor predictions of machine-
learning algorithms that rely on one or more so-called sensitive variables. That
is variables containing information (such as gender or ethnicity) that could create
distortions in the algorithm’s decision-making process. In many situations, the
determination of these sensitive variables is driven by ethical, legal, or regulatory
issues. From a moral point of view, penalizing a group of individuals is an unfair
decision. From a legal perspective, unfair algorithmic decisions are prohibited for a
large number of applications, including access to education, the welfare system, or
microfinance.1 To comply with fairness regulations, institutions may either change
the decision-making process to remove biases using affirmative actions or try to base
their decision on a fair version of the outcome.
Forecasting accuracy has become the gold standard in evaluating machine-learning
models. However, especially for more complex methods, the algorithm is often a black
box that provides a prediction without offering insights into the process which led to it.
Hence, when bias is present in the learning sample, the algorithm’s output can differ
for different subgroups of populations. At the same time, regulations may impose
that such groups ought to be treated in the same way. For instance, discrimination
can occur on the basis of gender or ethnic origin.
A typical example is one of the automatic human resources (HR) decisions that are
often influenced by gender. In available databases, men and women may self-select
in some job categories due to past or present preferences or cultural customs. Some
jobs are considered male-dominant, while other jobs are female-dominant. In such
unbalanced datasets, the machine-learning procedure learns that gender matters and
thus transforms correlation into causality by using gender as a causal variable in
prediction. From a legal point of view, this biased decision leads to punishable
gender discrimination. We refer to De-Arteaga et al. (2019) for more insights on this
gender gap. Differential treatment for university admissions suffers from the same
problems. We point out the example of law school admissions described in McIntyre
and Simkovic (2018), which is used as a common benchmark to evaluate the bias of
algorithmic decisions.
Imposing fairness is thus about mitigating this unwanted bias and preventing the
sensitive variable from influencing decisions. We divide the concept of fairness into
two main categories. The first definition of fairness is to impose that the algorithm’s
output is, on average, the same for all groups. Hence, the sensitive variable does not
play any role in the decision. Such equality of treatment is referred to as statistical
parity.
An alternative fairness condition is to impose that two individuals who are identical
except for their value of the sensitive variables are assigned the same prediction.
This notion of fairness is known as equality of odds. In this case, we wish to ensure
1 Artificial Intelligence European Act, 2021.

7 Fairness in Machine Learning and Econometrics 219
that the algorithm performs equally across all possible subgroups. Equality of odds
is violated, for instance, by the predictive justice algorithm described in Angwin,
Larson, Mattu and Kirchner (2016). Everything else being equal, this algorithm was
predicting a higher likelihood of future crimes for African-American convicts.
Several methods have been developed to mitigate bias in algorithmic decisions.
The proposed algorithms are usually divided into three categories. The first method
is a post-processing method that removes the bias from the learning sample to
learn a fair algorithm. The second way consists of imposing a fairness constraint
while learning the algorithm and balancing the desired fairness with the model’s
accuracy. This method is an in-processing method. Finally, the last method is a
post-processing method where the output of a possibly unfair algorithm is processed
to achieve the desired level of fairness, modeled using different fairness measures.
All three methodologies require a proper definition of fairness and a choice of
fairness measures to quantify it. Unfortunately, a universal definition of fairness is not
available. Moreover, as pointed out by Friedler, Scheidegger and Venkatasubramanian
(2021), complying simultaneously with multiple restrictions has proven impossible.
Therefore, different fairness constraints give rise to different fair models.
Achieving full fairness consists in removing the effect of the sensitive variables
completely. It often involves important changes with respect to the unfair case and
comes at the expense of the algorithm’s accuracy when the accuracy is measured
using the biased distribution of the data set. When the loss of accuracy is considered
too critical by the designer of the model, an alternative consists in weakening the
fairness constraint. Hence the stakeholder can decide, for instance, to build a model
for which the fairness level will be below a certain chosen threshold. The model will
thus be called approximately fair.
To sum up, fairness with respect to some sensitive variables is about controlling
the influence of its distribution and preventing its influence on an estimator. We refer
to Barocas and Selbst (2016), Chouldechova (2017), Menon and Williamson (2018),
Gordaliza, Del Barrio, Fabrice and Loubes (2019), Oneto and Chiappa (2020) and
Risser, Sanz, Vincenot and Loubes (2019) or Besse, del Barrio, Gordaliza, Loubes
and Risser (2021) and references therein for deeper insights on the notion of bias
mitigation and fairness.
In the following, we present the challenges of fairness constraints in econometrics.
Some works have studied the importance of fairness in economics (see, for instance,
Rambachan, Kleinberg, Mullainathan and Ludwig (2020), Lee, Floridi and Singh
(2021), Hoda, Loi, Gummadi and Krause (2018), Hu and Chen (2020), Kasy and
Abebe (2021), and references therein). As seen previously, the term fairness is
polysemic and covers various notions. We will focus on the role and on the techniques
that can be used to impose fairness in a specific class of econometrics models.
Let us consider the example in which an institution must decide on a group of

individuals. For instance, this could be a university admitting new students based on
their expected performance in a test; or a company deciding the hiring wage of new
employees. This decision is made by an algorithm, which we suppose works in the

following way. For a given vector of individual characteristics, denoted by 𝑋, this
algorithm computes a score ℎ(𝑋) ∈ R and makes a decision based on the value of
this score, which is determined by a functional D of ℎ. We are not specific about
the exact form of D (ℎ). For instance, this could be a threshold function in which
students are admitted if the score is higher than or equal to some values 𝐶, and they
are not admitted otherwise. The algorithm is completed by a learning model, which
is written as follows
𝑌 = ℎ(𝑋) + 𝑈, (7.1)
where 𝑌 is the outcome, and 𝑈 is a statistical error (see equation 2.1).2 For instance,
𝑌 could be the test result from previous applicants. We let 𝑋 = (𝑍, 𝑆) ∈ R 𝑝+1 and
X = Z × S to be the support of the random vector 𝑋. This learning model is used to
approximate the score, ℎ(𝑋), which is then used in the decision model.
Let us assume that historical data show that students from private high schools
obtain higher test scores than students in public high schools. The concern with
fairness in this model is twofold. On the one hand, if the distinction between public
and private schools is used as a predictor, students from private schools will always
have a higher probability of being admitted to a given university. On the other hand,
the choice of school is an endogenous decision that is taken by the individual and may
be determined by variables that are unobservable to the econometrician. Hence the
bias will be reflected both in the lack of fairness in past decision-making processes
and the endogeneity of individual choices in observational data. Predictions and
admission decisions may be unfair towards the minority class and bias the decision
process, possibly leading to discrimination. To overcome this issue, we consider
that decision-makers can embed in their learning model a fairness constraint. This
fairness constraint limits the relationship between the score ℎ(𝑋) and 𝑆. Imposing a
fairness constraint directly on ℎ and not on D (ℎ) is done for technical convenience, as
D (ℎ) is often nonlinear, which complicates the estimation and prediction framework
substantially.
More generally, we aim to study the consequences of incorporating a fairness
constraint in the estimation procedure when the score, ℎ, solves a linear inverse
problem of the type
𝐾 ℎ = 𝑟, (7.2)
where 𝐾 is a linear operator. Equation 7.2 can be interpreted as an infinite-dimensional
linear system of equations, where the operator 𝐾 is tantamount to an infinite-
dimensional matrix, whose properties determine the existence and uniqueness of
a solution to 7.2. A leading example of this setting is nonparametric instrumental
regressions (Newey & Powell, 2003; Hall & Horowitz, 2005; Darolles, Fan, Florens
& Renault, 2011). Nonetheless, many other models, such as linear and nonlinear
parametric regressions and additive nonparametric regressions, can fit this general
framework (Carrasco, Florens & Renault, 2007).
2 The learning model is defined as the statistical specification fitted to the learning sample to estimate
the unknown index ℎ.
Let P𝑋 be the distribution function of 𝑋, and E be the space of functions of 𝑋,

which are square-integrable with respect to P𝑋 . That is,
∫
2
E := ℎ : ℎ (𝑥)𝑑P𝑋 (𝑥) < ∞ .
Similarly, let G be the space of functions of 𝑋, which satisfy a fairness constraint.

We model the latter as another linear operator 𝐹 : E → G such that
𝐹 ℎ = 0. (7.3)
That is, the null space (or kernel) of the operator 𝐹 is the space of those functions
that satisfy a fairness restriction, N (𝐹) = {𝑔 ∈ E, 𝐹𝑔 = 0}. The full fairness constraint
implies restricting the solutions to the functional problem to the kernel of the operator.
To weaken this requirement, we also consider relaxations of the condition and define
an approximate fairness condition as
∥𝐹 ℎ∥ ≤ 𝜌
for a fixed parameter 𝜌 ≥ 0.

In this work, we consider fairness according to the following definitions.
Definition 7.1 (Statistical Parity) The algorithm ℎ maintains statistical parity if, for
every 𝑠 ∈ S,
𝐸 [ℎ(𝑍, 𝑠)|𝑆 = 𝑠] = 𝐸 [ℎ(𝑋)] .
Definition 7.2 (Irrelevance in prediction) The algorithm ℎ does not depend on 𝑆.

That is for all 𝑠 ∈ S,
𝜕ℎ(𝑥)
= 0.
𝜕𝑠
Definition 7.1 states that the function ℎ is fair when individuals are treated the
same, on average, irrespective of the value of the sensitive attribute, 𝑆. For instance,
if 𝑆 is a binary characteristic of the population, with 𝑆 = 1 being the protected group,
Definition 7.1 implies that the average score for group 𝑆 = 0 and the average score for
group 𝑆 = 1 are the same. Notice that this definition of fairness does not ensure that
two individuals with the same vector of characteristics 𝑍 = 𝑧 but with different values
of 𝑆 are treated in the same way. The latter is instead true for our second definition of
fairness. In this case, fairness is defined as the lack of dependence of ℎ on 𝑆, which
implies the equality of odds for individuals with the same vector of characteristics
𝑍 = 𝑧. However, we want to point out that both definitions may fail to deliver fairness
if the correlation between 𝑍 and 𝑆 is very strong. In our example above, if students
going to private schools have a higher income than students going to public schools,
and income positively affects the potential score, then discrimination would still
occur based on income.
Other definitions of fairness are possible. In particular, definitions that impose
restrictions on the entire distribution of ℎ given 𝑆. These constraints are nonlinear
and thus more cumbersome to deal with in practice, and we defer their study to future
work.
7.2 Examples in Econometrics
We let F1 and F2 be the set of square-integrable functions which satisfy definitions

7.1 and 7.2, respectively. We consider below examples in which the function ℎ 𝐹
satisfies
ℎ 𝐹 = arg min E (𝑌 − 𝑓 (𝑋)) 2 |𝑊 = 𝑤 ,

𝑓 ∈ F𝑗
with 𝑗 = {1, 2}, and where 𝑊 is a vector of exogenous variables.
7.2.1 Linear IV Model
Consider the example of a linear model in which ℎ(𝑋) = 𝑍 ′ 𝛽 + 𝑆 ′ 𝛾, with 𝑍, 𝛽 ∈ R 𝑝 ,

and 𝑆, 𝛾 ∈ R𝑞 . We take both 𝑍 and 𝑆 to be potentially endogenous, and we have a
vector of instruments 𝑊 ∈ R 𝑘 , such that 𝑘 ≥ 𝑝 + 𝑞 and 𝐸 [𝑊 ′𝑈] = 0.
We let 𝑋 = (𝑍 ′, 𝑆 ′) ′ be the vector of covariates, and ℎ = (𝛽 ′, 𝛾 ′) ′ be the vector of
unknown coefficients.
For simplicity of exposition, we maintain the assumption that the vector
 Σ Σ′ 
   
𝑋
  ∼ 𝑁 ©0 𝑝+𝑞+𝑘 ,  𝑋 𝑋𝑊  ª® ,
   
𝑊  Σ𝑋𝑊 𝐼 𝑘 
  «  ¬
where 0 𝑝+𝑞+𝑘 is a vector of zeroes of dimension 𝑝 + 𝑞 + 𝑘, 𝐼 𝑘 is the identity matrix of
dimension 𝑘, and
 Σ 𝑍 Σ 𝑍′ 𝑆 
  h i
Σ𝑋 =

,
 Σ𝑋𝑊 = Σ 𝑍𝑊 Σ𝑆𝑊 .
|{z} Σ 𝑍 𝑆 Σ𝑆  |{z}
( 𝑝+𝑞)×( 𝑝+𝑞)   𝑘×( 𝑝+𝑞)
The unconstrained value of ℎ is therefore given by

−1
′
ℎ = Σ𝑋𝑊 Σ𝑋𝑊 ′
Σ𝑋𝑊 𝐸 [𝑊𝑌 ] = (𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟.
Because of the assumption of joint normality, we have that 𝐸 [𝑍 |𝑆] = Π𝑆, where
Π = Σ𝑆−1 Σ 𝑍 𝑆 is a 𝑝 × 𝑞 matrix. Statistical parity, as defined in 7.1, implies that
𝑆 ′ (Π𝛽 + 𝛾) = 0,
which is true as long as Π𝛽 + 𝛾 = 0𝑞 . In the case of Definition 7.2, the fairness

constraint is simply given by 𝛾 = 0.
7.2.2 A Nonlinear IV Model with Binary Sensitive Attribute
Let 𝑍 ∈ R 𝑝 be a continuous variable and 𝑆 = {0, 1} 𝑞 a binary random variable. For

instance, 𝑆 can characterize gender, ethnicity, or a dummy for school choice (public
vs. private). Because of the binary nature of 𝑆
ℎ(𝑋) = ℎ0 (𝑍) + ℎ1 (𝑍)𝑆.
Definition 7.1 implies that we are looking for functions {ℎ0 , ℎ1 } such that
𝐸 [ℎ0 (𝑍)|𝑆 = 0] = 𝐸 [ℎ0 (𝑍) + ℎ1 (𝑍)|𝑆 = 1] .
That is
𝐸 [ℎ1 (𝑍)|𝑆 = 1] = 𝐸 [ℎ0 (𝑍)|𝑆 = 0] − 𝐸 [ℎ0 (𝑍)|𝑆 = 1] .
Definition 7.2 instead simply implies that ℎ1 = 0, almost surely. In particular, under
the fairness restriction, 𝑌 = ℎ0 (𝑍) + 𝑈. We discuss this example in more detail in
Section 7.6.
7.2.3 Fairness and Structural Econometrics
More generally, supervised machine learning models are often about predicting a
conditional moment or a conditional probability. However, in many leading examples
in structural econometrics, the score function ℎ does not correspond directly to a
conditional distribution or a conditional moment of the distribution of the learning
variable 𝑌 . Let Γ be the probability distribution generating the data. Then the function
ℎ is the solution to the following equation
𝐴 (ℎ, Γ) = 0.
A leading example is one of the Neyman-Fisher-Cox-Rubin potential outcome models,

in which 𝑋 represents a treatment and, for 𝑋 = 𝜉, we can write
𝑌 𝜉 = ℎ(𝜉) + 𝑈 𝜉 .

If 𝐸 𝑈 𝜉 |𝑊 = 0, this model leads to the nonparametric instrumental regression
model mentioned above, in which the function 𝐴(ℎ, Γ) = 𝐸 [𝑌 − ℎ(𝑋)|𝑊] = 0, and
the fairness condition is imposed directly on the function ℎ. However, this potential
outcome model can lead to other objects of interest. For instance, if we assume for
simplicity that (𝑋,𝑊) ∈ R2 , and under a different set of identification assumptions, it
can be proven that

𝑑𝐸 [𝑌 |𝑊 ]
𝑑ℎ(𝑋) 𝑑𝑊
𝐴(ℎ, Γ) = 𝐸 |𝑊 − 𝑑𝐸 [𝑍 |𝑊 ]
= 0,
𝑑𝑋
𝑑𝑊
which is a linear equation in ℎ that combines integral and differential operators (see
Florens, Heckman, Meghir & Vytlacil, 2008). In this case, the natural object of
interest is the first derivative of ℎ(𝑥), which is the marginal treatment effect. The
fairness constraint is therefore naturally imposed on 𝑑ℎ( 𝑥)
𝑑𝑥 .
Another class of structural models not explicitly considered in this work is one of
nonlinear nonseparable models. In these models, we have that
𝑌 = ℎ(𝑋,𝑈), with 𝑈 ⊥⊥ 𝑊 and 𝑈 ∼ U [0, 1],
and ℎ(𝜉, ·) monotone increasing in its second argument. In this case, ℎ is the solution
to the following nonlinear inverse problem
∫
𝑃 (𝑌 ≤ ℎ(𝑥, 𝑢)|𝑋 = 𝑥,𝑊 = 𝑤) 𝑓 𝑋 |𝑊 (𝑥|𝑤)𝑑𝑥 = 𝑢.
The additional difficulty lies in imposing a distributional fairness constraint in this

setting. We defer the treatment of this case to future research.
7.3 Fairness for Inverse Problems
Recall that the nonparametric instrumental regression (NPIV) model amounts to

solving an inverse problem. Let 𝑊 be a vector of instrumental variables. The NPIV
regression model can be written as
𝐸 (𝑌 |𝑊 = 𝑤) = 𝐸 (ℎ(𝑍, 𝑆)|𝑊 = 𝑤).
We let 𝑋 = (𝑍, 𝑆) ∈ 𝑅 𝑝+𝑞 and X = Z × S to be the support of the random vector 𝑋.

We further restrict ℎ ∈ 𝐿 2 (𝑋), with 𝐿 2 being the space of square-integrable functions
with respect to some distribution P. We further assume that this distribution P is
absolutely continuous with respect to the Lebesgue measure, and it therefore admits a
density, 𝑝 𝑋𝑊 . Using a similar notation, we let 𝑝 𝑋 and 𝑝 𝑊 , be the marginal densities
of 𝑋 and 𝑊, respectively.
If we let 𝑟 = 𝐸 (𝑌 |𝑊 = 𝑤) and 𝐾 ℎ = 𝐸 (ℎ(𝑋)|𝑊 = 𝑤), where 𝐾 is the conditional
expectation operator, then the NPIV framework amounts to solving an inverse problem.
That is, estimating the true function ℎ† ∈ E, defined as the solution of
𝑟 = 𝐾 ℎ† . (7.4)
Equation (7.4) can be interpreted as an infinite-dimensional system of linear equations.

In parallel with the finite-dimensional case, the properties of the solution depend on
the properties of the operator 𝐾. In 1923, Hadamard postulated three requirements for
problems of this type in mathematical physics: a solution should exist, the solution
should be unique, and the solution should depend continuously on 𝑟. That is, ℎ† is
stable to small changes in 𝑟. A problem satisfying all three requirements is called
well-posed. Otherwise, it is called ill-posed. The existence of the solution is usually
established by restricting 𝑟 to belong to the range of the operator 𝐾. In our case, it is
sufficient to let 𝐾 : E → 𝐿 2 (𝑊), and 𝑟 ∈ 𝐿 2 (𝑊), where 𝐿 2 (𝑊) is the space of square-
integrable functions of 𝑊. The uniqueness of a solution to (7.4) is guaranteed by the
so-called completeness condition (or strong identification, see Florens, Mouchart
& Rolin, 1990; Darolles et al., 2011). That is, 𝐾 ℎ = 0, if and only if, ℎ = 0, where
equalities are intended almost surely. In particular, Let ℎ1 and ℎ2 two solutions to
(7.4), then 𝐾 (ℎ1 − ℎ2 ) = 0, which implies that ℎ1 = ℎ2 , by the strong identification
assumption.
In the following, for two square-integrable functions of 𝑋, {ℎ1 , ℎ2 }, we let
∫
⟨ℎ1 , ℎ2 ⟩ = ℎ1 (𝑥)ℎ2 (𝑥)𝑑P,
∥ℎ1 ∥ 2 =⟨ℎ1 , ℎ1 ⟩.
The adjoint operator, 𝐾 ∗ , is obtained as

∫ ∫
⟨𝐾 ℎ, 𝑟⟩ = ℎ(𝑥) 𝑝 𝑋 |𝑊 (𝑥|𝑤)𝑑𝑥 𝑟 (𝑤) 𝑝 𝑊 (𝑤)𝑑𝑤
𝑊 𝑋
∫ ∫
= 𝑟 (𝑤) 𝑝 𝑊 |𝑋 (𝑤|𝑥)𝑑𝑤 ℎ(𝑥) 𝑝 𝑋 (𝑥)𝑑𝑥 = ⟨ℎ, 𝐾 ∗ 𝑟⟩.
𝑋 𝑊
If the operator 𝐾 ∗ 𝐾 is invertible, the solution of (7.4) is given by
ℎ† = (𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟. (7.5)
The ill-posedness of the inverse problem in (7.4) comes from the fact that, when
the distribution of (𝑋,𝑊) is continuous, the eigenvalues of the operator 𝐾 ∗ 𝐾 have
zero an as accumulation point. That is, small changes in 𝑟 may correspond to large
changes in the solution as written in (7.5). The most common solution to deal with
the ill-posedness of the inverse problem is to use a regularization technique (see
Engl, Hanke & Neubauer, 1996, and references therein). In this chapter, we consider
Tikhonov regularization, which imposes an 𝐿 2 -penalty on the function ℎ (Natterer,
1984). Heuristically, Tikhonov regularization can be considered as the functional
extension of Ridge regressions, which are used in linear regression models to obtain
an estimator of the parameters when the design matrix is not invertible. In this respect,
Tikhonov regularization imposes an 𝐿 2 -penalty on the functions of interest (see
Chapters 1 and 2 of this manuscript for additional details).
A regularized version of ℎ, as presented in (Engl et al., 1996), is defined as the

solution of a penalized optimization program
ℎ 𝛼 = arg min ∥𝑟 − 𝐾 ℎ∥ 2 + 𝛼∥ℎ∥ 2 ,

ℎ∈E
which can be explicitly written as
ℎ 𝛼 = (𝛼Id + 𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟 = 𝑅 𝛼 (𝐾)𝐾 ∗ 𝑟 (7.6)
where 𝑅 𝛼 (𝐾) = (𝛼Id + 𝐾 ∗ 𝐾) −1 is a Tikhonov regularized operator. The solution

depends on the choice of the regularization parameter 𝛼, which helps bound the
eigenvalues of 𝐾 ∗ 𝐾 away from zero. Therefore, precisely as in Ridge regressions,
a positive value of 𝛼 introduces a bias in estimation but it allows us to bound the
variance of the estimator away from infinity. As 𝛼 → 0, in a suitable way, then
ℎ 𝛼 → ℎ† .
We consider the estimation of the function ℎ from the following noisy observational
model
𝑟ˆ = 𝐾 ℎ† + 𝑈𝑛 , (7.7)
where 𝑈𝑛 is an unknown random function with bounded norm. That is, ∥𝑈𝑛 ∥ 2 = 𝑂 (𝛿 𝑛 )
for a given sequence 𝛿 𝑛 which tends to 0 as 𝑛 goes to infinity. The operators 𝐾 and
𝐾 ∗ are taken to be known for simplicity. This estimation problem has been widely
studied in the econometrics literature, and we provide details on the estimation of
the operator in Section 7.5. We refer, for instance, to Darolles et al. (2011) for the
asymptotic properties of the NPIV estimator when the operator 𝐾 is estimated from
data.
To summarise, we impose the following regularity conditions on the operator 𝐾
and the true solution ℎ† .
• [A1] 𝑟 ∈ R (𝐾) where R (𝐾) stands for the range of the operator 𝐾.
• [A2] The operator 𝐾 ∗ 𝐾 is a one-to-one operator. This condition implies the
completeness condition as defined above and ensures the identifiability of ℎ† .
• [A3] Source Condition : we assume that there exists 𝛽 ≤ 2 such that
𝛽
ℎ† ∈ R (𝐾 ∗ 𝐾) 2 .
This condition relates the smoothness of the solution of equation (7.4) to the decay
of the eigenvalues of the SVD decomposition of the operator 𝐾. It is commonly
used in statistical inverse problems. In particular, it guarantees that the Tikhonov
regularized solution ℎ 𝛼 converges to the true solution ℎ† at a rate of convergence
given by
∥ℎ 𝛼 − ℎ† ∥ 2 = 𝑂 (𝛼 𝛽 ).
We refer to Loubes and Rivoirard (2009) for a review of the different smoothness
conditions for inverse problems.
7.4 Full Fairness IV Approximation
In this model, full fairness of a function 𝜓 ∈ E is achieved when 𝐹𝜓 = 0, i.e. when the
function belongs to the null space of a fairness operator, 𝐹. Hence imposing fairness
amounts to considering functions that belong to the null space, N (𝐹), and that are
approximate solutions of the functional equation (7.4). The full fairness condition
may be seen as a very restrictive way to impose fairness. Actually, if the functional
equation does not have a solution in N (𝐹), full fairness will induce a loss of accuracy,
which we refer to as price for fairness. The projection to fairness has been studied in
the regression framework in Le Gouic, Loubes and Rigollet (2020), Chzhen, Denis,
Hebiri, Oneto and Pontil (2020), and Jiang, Pacchiano, Stepleton, Jiang and Chiappa
(2020), for the classification task.
The full fairness solution can be achieved in two different ways: either by looking at
the solution of the inverse problem and then imposing a fair condition on the solution;
or by directly solving the inverse problem under the restriction that the solution is fair.
We prove that the two procedures are not equivalent, leading to different estimators
having different properties.
•φ
K
N (F )
•
φF
•r
G
•
KφF
Fig. 7.1: Example of projection onto the space of fair functions.

Figure 7.1 illustrates the situation where either the solution can be solved and then
the fairness condition can be imposed. or the solution is directly approximated in the
set of fair functions.
7.4.1 Projection onto Fairness
The first way consists in first considering the regularized solution to the inverse
problem ℎˆ 𝛼 defined as the Tikhonov regularized solution of the inverse problem

ℎˆ 𝛼 = arg min ∥ 𝑟ˆ − 𝐾 ℎ∥ 2 + 𝛼∥ℎ∥ 2
ℎ∈ E
which can be computed as
ℎˆ 𝛼 = (𝛼Id + 𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟ˆ = 𝑅 𝛼 (𝐾)𝐾 ∗ 𝑟.
ˆ
Then the fair solution is defined as the projection onto the set which models the
fairness condition N ,
ℎˆ 𝛼,𝐹 = arg min ∥ ℎˆ 𝛼 − ℎ∥ 2
ℎ∈N (𝐹)
In this framework, denote by 𝑃 : E → N (𝐹) the projection operator onto the kernel
of the fairness operator. Hence we have
ℎˆ 𝛼,𝐹 = 𝑃 ℎˆ 𝛼 .
Example 7.1 (Linear Model, continued.) Recall that the constraint of statistical parity,
as in Definition 7.1, implies that
𝑆 ′ (Π𝛽 + 𝛾) = 0,
which is true as long as Π𝛽 + 𝛾 = 0𝑞 . Thus, we have that

h i
𝐹 = Π 𝐼𝑞 ,
|{z}
𝑞×( 𝑝+𝑞)
and
Π ′ 𝐼𝑞 + ΠΠ ′ −1 Π Π ′ 𝐼𝑞 + ΠΠ ′ −1 
 
′ ′ −1
𝑃 = 𝐼 𝑝+𝑞 − 𝐹 (𝐹𝐹 ) 𝐹 = 𝐼 𝑝+𝑞 − 
 −1  ,
 𝐼𝑞 + ΠΠ ′ −1 Π ′

𝐼𝑞 + ΠΠ 
 
which immediately gives 𝐹𝑃 = 0𝑞 . Hence, the value of ℎ 𝐹 = 𝑃ℎ is the projection of
the vector ℎ onto the null space of 𝐹.
In the case of definition 7.2, the fairness constraint is simply given by 𝛾 = 0. Let
′ −1 ′
𝑀𝑍𝑊 = 𝐼 𝑘 − Σ 𝑍𝑊 Σ 𝑍𝑊 Σ 𝑍𝑊 Σ 𝑍𝑊 ,
and −1
′ ′
𝐴 𝑍 𝑆 = Σ 𝑍𝑊 Σ 𝑍𝑊 Σ 𝑍𝑊 Σ𝑆𝑊 .
When one wants to project the unconstrained estimator onto the constrained space by
the block matrix inversion lemma, we notice that
 ′ −1 −1 −1  "
+ 𝐴 𝑍𝑆 Σ′𝑆𝑊 𝑀𝑍𝑊 Σ𝑆𝑊 𝐴′𝑍𝑆 −𝐴 𝑍𝑆 Σ′𝑆𝑊 𝑀𝑍𝑊 Σ𝑆𝑊
#
 Σ′ 𝐸 [𝑊𝑌 ]
 Σ 𝑍𝑊 Σ 𝑍𝑊

ℎ= −1  𝑍𝑊

−1
 − Σ′𝑆𝑊 𝑀𝑍𝑊 Σ𝑆𝑊 𝐴′𝑍𝑆 Σ′𝑆𝑊 𝑀𝑍𝑊 Σ𝑆𝑊  Σ′ 𝐸 [𝑊𝑌 ]
  𝑆𝑊
 
 ′ −1 ′ −1 
 𝑍𝑊 Σ 𝑍𝑊
 Σ Σ 𝑍𝑊 𝐸 [𝑊𝑌 ] − 𝐴 𝑍𝑆 Σ′𝑆𝑊 𝑀𝑍𝑊 Σ𝑆𝑊 Σ′𝑆𝑊 𝐸 [𝑊𝑌 ] − 𝐴′𝑍𝑆 Σ′𝑍𝑊 𝐸 [𝑊𝑌 ] 
= −1 

 Σ′𝑆𝑊 𝑀𝑍𝑊 Σ𝑆𝑊 Σ′𝑆𝑊 𝐸 [𝑊𝑌 ] − 𝐴′𝑍𝑆 Σ′𝑍𝑊 𝐸 [𝑊𝑌 ] 

 
−1
" #
Σ′𝑍𝑊 Σ 𝑍𝑊 Σ′𝑍𝑊 𝐸 [𝑊𝑌 ] − 𝐴 𝑍𝑆 𝛾

= .
𝛾
Therefore, we have that

 
 𝛽 + 𝐴𝑍 𝑆 𝛾
ℎ 𝐹 = 𝑃ℎ =  .

 0𝑞 
 
The behavior of the projection of the unfair solution onto the space of fair functions
is given by the following theorem
Theorem 7.1 Under Assumptions [A1] to [A3], the fair projection estimator is such
that
1
∥ ℎˆ 𝛼,𝐹 − 𝑃ℎ† ∥ 2 = 𝑂 + 𝛼𝛽 (7.8)
𝛼𝛿 𝑛
Proof
∥ ℎˆ 𝛼,𝐹 − 𝑃ℎ† ∥ ≤ ∥𝑃 ℎˆ 𝛼 − 𝑃ℎ† ∥

≤ ∥ ℎˆ 𝛼 − ℎ† ∥
since 𝑃 is a projection. The term ∥ ℎˆ 𝛼 − ℎ† ∥ is the usual estimation term for the
structural IV inverse problem. As proved in (Darolles et al., 2011) this term converges
at the following rate
1
∥ ℎˆ 𝛼 − ℎ† ∥ 2 = 𝑂 + 𝛼𝛽 ,
𝛼𝛿 𝑛
which proves the result. □
The estimator converges towards the fair part of the function ℎ† , i.e. its projection
onto the kernel of the fairness operator 𝐹. If we consider the difference with respect
to the unconstrained solution, we have that

ˆ 2 1 𝛽 2
∥ ℎ 𝛼 − ℎ† ∥ = 𝑂 + 𝛼 + ∥ℎ† − 𝑃ℎ† ∥ .
𝛼𝛿 𝑛
Hence the difference ∥ℎ† − 𝑃ℎ† ∥ 2 corresponds to the price to pay for ensuring
fairness, which is equal to zero only if the true function satisfies the fairness constraint.
This difference between the underlying function ℎ† and its fair representation is the
necessary change of the model that would enable a fair decision process minimizing
the quadratic distance between the fair and the unfair functions.
7.4.2 Fair Solution of the Structural IV Equation
A second and alternative solution to impose fairness is to solve the structural IV

equation directly on the space of fair functions, N (𝐹). We denote by 𝐾 𝐹 the operator
𝐾 restricted to N (𝐹), 𝐾 𝐹 : N (𝐹) ↦→ F . Since N (𝐹) is a convex closed space, the
projection onto this space is well-defined and unique. We will write 𝑃 the projection
onto N (𝐹) and 𝑃⊥ the projection onto its orthogonal complement in E, N (𝐹) ⊥ .
With these notations, we get that 𝐾 𝐹 = 𝐾 𝑃.
Definition 7.3 Define ℎ 𝐾𝐹 as the solution of the structural equation 𝐾 ℎ = 𝑟 in the set
of fair functions defined as the kernel of the operator 𝐹, i.e.

ℎ 𝐾𝐹 = arg min ∥𝑟 − 𝐾 ℎ∥ 2 .
ℎ∈N (𝐹)
Note that ℎ 𝐾𝐹 is the projection of ℎ† onto N (𝐹) with the metric defined by 𝐾 ∗ 𝐾,
since
ℎ 𝐾𝐹 = arg min ∥𝐾 ℎ† − 𝐾 ℎ∥ 2 .
ℎ∈N (𝐹)
Note that this approximation depends not only on 𝐾 but on the properties of the
fair kernel 𝐾 𝐹 = 𝐾 𝑃. Therefore, fairness is quantified here through its effect on
the operator 𝐾 and we denote this solution ℎ 𝐾𝐹 to highlight its dependence on the
operators 𝐾 and 𝐹.
The following proposition proposes an explicit expression of ℎ 𝐾𝐹 .
Proposition 7.1
ℎ 𝐾𝐹 = (𝐾 𝐹∗ 𝐾 𝐹 ) −1 𝐾 𝐹∗ 𝑟.
Proof First, ℎ 𝐾𝐹 belongs to N (𝐹). For any function 𝑔 ∈ E, 𝑃𝐾 ∗ 𝐾𝑔 ∈ N (𝐹) so the
operator (𝐾 𝐹∗ 𝐾 𝐹 ) −1 = (𝑃𝐾 ∗ 𝐾 𝑃) −1 is defined from N (𝐹) ↦→ N (𝐹).
Let 𝜓 ∈ N (𝐹) so 𝑃𝜓 = 𝜓. We have that
0 =< 𝑟 − 𝐾 ℎ 𝐾𝐹 , 𝐾𝜓 >
=< 𝐾 ∗ 𝑟 − 𝐾 ∗ 𝐾 ℎ 𝐾𝐹 , 𝜓 >
=< 𝐾 ∗ 𝑟 − 𝐾 ∗ 𝐾 𝑃ℎ 𝐾𝐹 , 𝑃𝜓 >
=< 𝑃𝐾 ∗ 𝑟 − 𝑃𝐾 ∗ 𝐾 𝑃ℎ 𝐾𝐹 , 𝜓 >
which holds for 𝑃𝐾 ∗ 𝑟 − 𝑃𝐾 ∗ 𝐾 𝑃ℎ 𝐾𝐹 = 0, which leads to ℎ 𝐾𝐹 = (𝑃𝐾 ∗ 𝐾 𝑃) −1 𝑃𝐾 ∗ 𝑟.□

Example 7.2 (Linear model, continued.) For both our definitions of fairness in 7.1
and 7.2, we have that
′ −1 ′
ℎ 𝐾𝐹 = 𝑃Σ𝑋𝑊 Σ𝑋𝑊 𝑃 𝑃Σ𝑋𝑊 𝐸 [𝑊𝑌 ] ,
which simply restricts the conditional expectation operators onto the null space of 𝐹.
In the case of definition 7.2, the closed-form expression of this estimator is easy
to obtain and is equal to
′
−1 ′
© Σ 𝑍𝑊 Σ 𝑍𝑊 Σ 𝑍𝑊 𝐸 [𝑊𝑌 ] ª ′ −1 ′
ℎ 𝐾𝐹 = ® = 𝑃Σ𝑋𝑊 Σ𝑋𝑊 𝑃 𝑃Σ𝑋𝑊 𝐸 [𝑊𝑌 ] ,
« 0 𝑞 ¬
which is equivalent to excluding 𝑆 from the second stage estimation of the IV model,
and where
 
0 𝑝× 𝑝 0 𝑝×𝑞 
𝐹=  , and 𝑃 = 𝐼 𝑝+𝑞 − 𝐹.
 
 0𝑞× 𝑝 𝐼𝑞 
 
Now consider the fair approximation of the solution of (7.4) as the solution of the
following minimization program

ℎˆ 𝐾𝐹 , 𝛼 = arg min ∥ 𝑟ˆ − 𝐾 ℎ∥ 2 + 𝛼∥ℎ∥ 2 .
ℎ∈N (𝐹)
Proposition 7.2 The fair solution of the IV structural equation has the following
expression
ℎˆ 𝐾𝐹 , 𝛼 = (𝛼Id + 𝐾 𝐹∗ 𝐾 𝐹 ) −1 𝐾 𝐹∗ 𝑟.
ˆ
It converges to ℎ 𝐾𝐹 when 𝛼 goes to zero as soon as 𝛼 is chosen such that 𝛼𝛿 𝑛 → +∞.
Proof As previously, ℎˆ 𝐾𝐹 , 𝛼 minimizes in N (𝐹), ∥ 𝑟ˆ − 𝐾 ℎ∥ 2 + 𝛼∥ℎ∥ 2 . Hence the
first-order condition is that for all 𝑔 ∈ N (𝐹) we have
< −𝐾𝑔, 𝑟ˆ − 𝐾 ℎ > +𝛼 < 𝑔, ℎ > = 0

< 𝑔, 𝐾 ∗ 𝐾 ℎ − 𝐾 ∗ 𝑟ˆ > +𝛼 < 𝑔, ℎ > = 0
< 𝑔, 𝑃𝐾 ∗ 𝐾 ℎ − 𝑃𝐾 ∗ 𝑟ˆ + 𝛼ℎ > = 0.
Hence using 𝐾 𝐹∗ = 𝑃𝐾 ∗ , and since ℎ is in N (𝐹) and thus 𝑃ℎ = ℎ, we obtain the

expression of the theorem.
Using this expression, we can compute the estimator as follows :
ℎˆ 𝐾𝐹 , 𝛼 − ℎ 𝐾𝐹 =
(𝛼Id + 𝐾 𝐹∗ 𝐾 𝐹 ) −1 𝐾 𝐹∗ (𝑟ˆ − 𝐾 ℎ† ) + ((𝛼Id + 𝐾 𝐹∗ 𝐾 𝐹 ) −1 − (𝐾 𝐹∗ 𝐾 𝐹 ) −1 )𝐾 𝐹∗ 𝐾 ℎ†
= (𝐼) + (𝐼 𝐼).
The first term is a variance term which is such that


1
∥(𝐼)∥ 2 = 𝑂 .
𝛼𝛿 𝑛
Recall that for two operators
𝐴−1 − 𝐵−1 = 𝐴−1 (𝐵 − 𝐴)𝐵−1 .
Hence, the second term can be written as
(𝐼 𝐼) = −𝛼(𝛼Id + 𝐾 𝐹∗ 𝐾 𝐹 ) −1 ℎ 𝐾𝐹 .
This tern is the bias of Tikhonov’s regularization of the operator 𝐾 𝐹∗ 𝐾 𝐹 = 𝑃𝐾 ∗ 𝐾 𝑃,

which goes to zero when 𝛼 goes to zero. □
When 𝛼 decreases to zero, the rate of consistency of the projected fair estimator can
be made precise if we assume some Hilbert scale regularity for both the fair part of
ℎ† and the remaining unfair part 𝑃⊥ ℎ† .
Assume that
𝛽
• [E1] 𝑃ℎ† ∈ R (𝑃𝐾 ∗ 𝐾 𝑃) 2 for 𝛽 ≤ 2
𝛾
• [E2] 𝑃⊥ ℎ† ∈ R (𝑃𝐾 ∗ 𝐾 𝑃) 2 for 𝛾 ≤ 2.
These assumptions are analogous to the source condition in [A3] adapted to the fair
operator 𝐾 𝐹 .
Theorem 7.2 Under Assumptions [E1] and [E2], the estimator ℎˆ 𝐾𝐹 converges
towards ℎ 𝐾𝐹 at the following rate

1
∥ ℎˆ 𝐾𝐹 − ℎ 𝐾𝐹 ∥ 2 = 𝑂 + 𝛼 𝑚𝑖𝑛(𝛽,𝛾)
𝛼𝛿 𝑛
We recognize the usual rate of convergence of the Tikhonov’s regularized estimator.

The main change is given here by the fact that the rate is driven by the fair source
conditions [E1] and [E2]. These conditions relate the smoothness of the function with
the decay of the SVD of the operator restricted to the kernel of the fairness operator.
Proof The rate of convergence depends on the term (𝐼 𝐼) previously defined. We
decompose it into two terms.
(𝐼 𝐼) = −𝛼(𝛼Id + 𝐾 𝐹∗ 𝐾 𝐹 ) −1 (𝐾 𝐹∗ 𝐾 𝐹 ) −1 𝐾 𝐹∗ (𝐾 𝑃ℎ† + 𝐾 𝑃⊥ ℎ† )
= ( 𝐴) + (𝐵).
First remark that since 𝑃 = 𝑃2
( 𝐴) = −𝛼(𝛼Id + 𝐾 𝐹∗ 𝐾 𝐹 ) −1 (𝐾 𝐹∗ 𝐾 𝐹 ) −1 𝐾 𝐹∗ 𝐾 𝐹 𝑃ℎ†
= −𝛼(𝛼Id + 𝐾 𝐹∗ 𝐾 𝐹 ) −1 𝑃ℎ†
Assumption [E1] provides the rate of decay of this term ∥ ( 𝐴) ∥ 2 and enables to prove
that it is of order 𝛼 𝛽 .
For the second term (𝐵), consider the SVD of the operator 𝐾 𝐹 = 𝐾 𝑃 denoted by
𝜆 𝑗 , 𝜓 𝑗 , 𝑒 𝑗 for all 𝑗 ≥ 1. So we have that
∥ (𝐵) ∥ 2 = ∥𝛼(𝛼Id + 𝐾 𝐹∗ 𝐾 𝐹 ) −1 (𝐾 𝐹∗ 𝐾 𝐹 ) −1 𝐾 𝐹∗ 𝐾 𝑃⊥ ℎ† ∥ 2
∑︁ 𝜆2𝑗
= 𝛼2 | < 𝐾 𝑃 ⊥ ℎ† , 𝑒 𝑗 > | 2
𝜆4 (𝛼 + 𝜆2𝑗 ) 2
𝑗 ≥1 𝑗
2𝛾
∑︁ 𝜆𝑗 | < 𝐾 𝑃 ⊥ ℎ† , 𝑒 𝑗 > | 2
= 𝛼2
𝑗 ≥1
(𝛼 + 𝜆2𝑗 ) 2 𝜆𝑗
2(1+𝛾)
= 𝑂 (𝛼 𝛾 )
To ensure that
∑︁ | < 𝐾 𝑃⊥ ℎ† , 𝑒 𝑗 > | 2
2(1+𝛾)
< +∞
𝑗 ≥1 𝜆𝑗
we assume that
∑︁ | < 𝑃⊥ ℎ† , 𝜆 𝑗 𝜓 𝑗 > | 2 ∑︁ | < 𝑃⊥ ℎ† , 𝜓 𝑗 > | 2
2(1+𝛾)
= 2𝛾
< +∞
𝑗 ≥1 𝜆𝑗 𝑗 ≥1 𝜆𝑗
where 𝐾 ∗ 𝑒 𝑗 = 𝜆 𝑗 𝜓 𝑗 , which is ensured under Assumption [E2]. Finally the two terms
are of order 𝑂 (𝛼 𝛽 + 𝛼 𝛾 ), which proves the result. □
To summarize, we have defined two fair approximations of the function ℎ† . The
first one is its fair projection ℎ 𝐹 = 𝑃ℎ† , while the other is the solution of the fair
kernel ℎ 𝐾𝐹 . The two solutions coincide as soon as
ℎ 𝐾𝐹 − 𝑃ℎ† = (𝐾 𝐹∗ 𝐾 𝐹 ) −1 𝐾 𝐹∗ 𝐾 𝑃⊥ ℎ† = 0.
Under assumption [A2], 𝐾 𝐹∗ 𝐾 𝐹 is also one to one. Hence the difference between both
approximations is null only if
𝐾 𝑃⊥ ℎ† = 0. (7.9)
If we consider the case of (IV) regression. This condition is met when
𝐸 (ℎ(𝑍, 𝑆)|𝑊) − 𝐸 (𝐸 (ℎ(𝑍, 𝑆)|𝑍)|𝑊) = 0.
This is the case when the sensitive variable 𝑆 is independent w.r.t to the instrument
𝑊 conditionally to the characteristics 𝑍. Yet, in the general case, both functions are
different.
7.4.3 Approximate Fairness
Imposing (7.3) is a way to ensure complete fairness of the solution of (7.4). In many
cases, this full fairness leads to bad approximation properties. Hence, we replace it
with a constraint on the norm of 𝐹 ℎ. Namely, we look for the estimator defined as the
solution of the optimization

ℎˆ 𝛼,𝜌 = arg min ∥ 𝑟ˆ − 𝐾 ℎ∥ 2 + 𝛼∥ℎ∥ 2 + 𝜌∥𝐹 ℎ∥ 2 (7.10)
ℎ∈E
This estimator corresponds to the usual Tikhonov regularized estimator with an

additional penalty term 𝜌∥𝐹 ℎ∥ 2 . The penalty enforces fairness since it enforces ∥𝐹 ℎ∥
to be small, which corresponds to a relaxation of the full fairness constraint 𝐹 ℎ = 0.
The parameter 𝜌 provides a trade-off between the level of fairness which is imposed
and the closeness to the usual estimator of the NPIV model. We study its asymptotic
behavior in the following theorem.
Note first that the solution of (7.10) has a close form and can be written as
ℎˆ 𝛼,𝜌 = (𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟.
ˆ
The asymptotic behavior of the estimator is provided by the following theorem. It

also ensures that the limit solution of (7.10), i.e. when 𝜌 → +∞, is fair in the sense
that lim𝜌→+∞ ∥𝐹 ℎ 𝛼,𝜌 ∥ = 0.. That is, it converges to the structural solution restricted
to the set of fair functions ℎ 𝐾𝐹 .
We will use the following notations. Consider the collection of operators
𝐿 𝛼 = (𝛼Id + 𝐾 ∗ 𝐾) −1 𝐹 ∗ 𝐹
𝐿 = (𝐾 ∗ 𝐾) −1 𝐹 ∗ 𝐹.
• [A4] R (𝐹 ∗ 𝐹) ⊂ R (𝐾 ∗ 𝐾). This condition guarantees that the operators 𝐿 and 𝐿 𝛼
are well-defined operators.
𝐿 is an operator 𝐿 : E → E, which is not self-adjoint.
Consider also the operator
𝑇 = (𝐾 ∗ 𝐾) −1/2 𝐹 ∗ 𝐹 (𝐾 ∗ 𝐾) −1/2
which is an self-adjoint operator, and is well-defined as long as

• [A5] R (𝐹 ∗ 𝐹) ⊂ R (𝐾 ∗ 𝐾) 1/2 .
If we assume a source condition on the form
• [A6] There exists 𝛾 ≥ 𝛽
𝛾+1
𝐹 ∗ 𝐹𝑃⊥ ℎ† ∈ R (𝐾 ∗ 𝐾) 2
Theorem 7.3 (Consistency of fair IV estimator) The approximated fair IV estimator

ℎˆ 𝛼,𝜌 is an estimator of the fair projection of the structural function, i.e. ℎ 𝐾𝐹 . Its rate
of consistency under assumptions [A1] to [A6] is given by

ˆ 2 𝛽 1 1
∥ ℎ 𝛼,𝜌 − ℎ 𝐾𝐹 ∥ = 𝑂 𝛼 + 2 + . (7.11)
𝜌 𝛼𝛿 𝑛
The rate of convergence is consistent in the following sense. When we increase the
level of imposed fairness to the full fairness constraint, i.e when 𝜌 goes to infinity,
for appropriate choices of the smoothing parameter 𝛼, the estimator converges to a
fully fair function. The rate in 𝜌12 corresponds to the fairness part of the rate. If 𝛽 the
1
Source condition parameter can be chosen large enough such that 𝛼 𝛽 = 𝜌2
, hence we
− 1
recover, for an optimal choice of 𝛼opt of order 𝛿 𝑛 𝛽+1 ,
the usual rate of convergence of
the NPIV estimates 𝛽
−
∥ ℎˆ 𝛼,𝜌 − ℎ 𝐾𝐹 ∥ 2 = 𝑂 𝛿 𝑛 𝛽+1 .
Example 7.3 (Linear model, continued.) In the linear IV model, let

−1
ℎ𝜌 = 𝜌𝐹 ′ 𝐹 + Σ𝑋𝑊
′
Σ𝑋𝑊 ′
Σ𝑋𝑊 𝐸 [𝑊𝑌 ] ,
the estimator which imposes the approximate fairness constraint. Notice that
−1
𝜌𝐹 ′ 𝐹 + Σ′𝑋𝑊 Σ𝑋𝑊
−1 −1 ′ −1 ′ −1 −1
= Σ′𝑋𝑊 Σ𝑋𝑊 − 𝜌 Σ′𝑋𝑊 Σ𝑋𝑊 𝐹 𝐼𝑞 + 𝜌𝐹 Σ′𝑋𝑊 Σ𝑋𝑊 𝐹 𝐹 Σ′𝑋𝑊 Σ𝑋𝑊
−1 ′ −1

−1 −1 ′ 1 −1
= Σ′𝑋𝑊 Σ𝑋𝑊 − Σ′𝑋𝑊 Σ𝑋𝑊 𝐹 𝐼𝑞 + 𝐹 Σ′𝑋𝑊 Σ𝑋𝑊 𝐹 𝐹 Σ′𝑋𝑊 Σ𝑋𝑊 .
𝜌
This decomposition implies that
′ −1 −1 ′ −1
lim ℎ𝜌 = ℎ − Σ𝑋𝑊 Σ𝑋𝑊 𝐹 ′ 𝐹 Σ𝑋𝑊
′
Σ𝑋𝑊 𝐹 𝐹 ℎ,
𝜌→∞
which directly gives

lim 𝐹 ℎ𝜌 = 0.
𝜌→∞
Therefore, as implied by our general theorem, as 𝜌 diverges to ∞, the full fairness

constraint is imposed.
Remark 7.1 Previous theorems enable us to understand the asymptotic behavior

of the fair regularized IV estimator. When 𝛼 goes to zero, but 𝜌 is fixed, this
estimator converges towards a function ℎ𝜌 which differs from the original function
ℎ† . Interestingly, we point out that the fairness constraint enables one to obtain a
fair solution, but the latter does not coincide with the fair approximation of the
true function, ℎ† . Instead, the fair solution is obtained by considering the set of
approximate solutions that satisfy the fairness constraint.
Remark 7.2 The theorem requires an additional assumption denoted by [A6]. This
assumption aims at controlling the regularity of the unfair part of the function ℎ† . It
is analogous to a source condition imposed on the part of the solution which does not
lie in the kernel of the operator 𝐹, namely 𝑃⊥ ℎ† . This condition is obviously fulfilled
if ℎ† is fair since 𝑃⊥ ℎ† = 0.
Remark 7.3 The smoothness assumptions we impose in this paper are source con-
ditions with regularity smaller than 2. Such restrictions come from the choice of
standard Tikhonov’s regularization method. Other regularization approaches, such
as Landwebers’s iteration or iterated Tikhonov’s regularization, would enable to
deal with more regular functions without changing the results presented in this work
(Florens, Racine & Centorrino, 2018).
Proof (Theorem (7.3)) Note that the fair estimator can be decomposed into a bias
and a variance term that will be studied separately
ℎˆ 𝛼,𝜌 = (𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟ˆ
= (𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟 + (𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 𝐾 ∗𝑈𝑛
= (𝐵) + (𝑉).
Then the bias term can be decomposed as
(𝐵) = [(𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 − (𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 ]𝐾 ∗ 𝑟 + (𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟

= (𝐵1 ) + (𝐵2 ).
The operator (𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 can be written as
(𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 = (𝑅 −1 ∗
𝛼 (𝐾) + 𝜌𝐹 𝐹)
−1
= (Id + 𝜌𝑅 𝛼 (𝐾)𝐹 ∗ 𝐹) −1 𝑅 𝛼 (𝐾)
Note that condition [A4] ensures that
𝐿 𝛼 := 𝑅 𝛼 (𝐾)𝐹 ∗ 𝐹 = (𝐾 ∗ 𝐾 + 𝛼Id) −1 𝐹 ∗ 𝐹
is a well-defined operator on E. Moreover, condition [A2] ensures that 𝑅 𝛼 (𝐾) is

one-to-one. Hence, the kernel of the operator 𝐿 𝛼 is the kernel of 𝐹. Using the
Tikhonov approximation (7.6), we thus have
(𝐵1) = [(Id + 𝜌𝐿 𝛼 ) −1 𝑅 𝛼 (𝐾) − (Id + 𝜌𝐿) −1 (𝐾 ∗ 𝐾) −1 ]𝐾 ∗ 𝑟

= (Id + 𝜌𝐿 𝛼 ) −1 (ℎ 𝛼 − ℎ† ) + [(Id + 𝜌𝐿 𝛼 ) −1 − (Id + 𝜌𝐿) −1 ]ℎ†
We will study each term separately.

• Since ∥ (Id + 𝜌𝐿 𝛼 ) −1 ∥ is bounded, we get that the first term is of the same order as
ℎ 𝛼 − ℎ† . Hence, under the source condition in [A3], we have that
∥(Id + 𝜌𝐿 𝛼 ) −1 (ℎ 𝛼 − ℎ† )∥ 2 = 𝑂 (𝛼 𝛽 ).
• Using that for two operators
𝐴−1 − 𝐵−1 = 𝐴−1 (𝐵 − 𝐴)𝐵−1
we obtain for the second term that

(Id + 𝜌𝐿 𝛼 ) −1 − (Id + 𝜌𝐿) −1 ℎ† = 𝜌(Id + 𝜌𝐿 𝛼 ) −1 (𝐿 − 𝐿 𝛼 ) (Id + 𝜌𝐿) −1 ℎ† .
Note that (𝐿 − 𝐿 𝛼 )𝑃ℎ† = 0 and (Id + 𝜌𝐿) −1 𝑃ℎ† = 𝑃ℎ† .

Hence we can replace ℎ† in the last expression by the projection onto the orthogonal
space of N (𝐹). Namely, 𝑃⊥ ℎ† . Hence

∥ (Id + 𝜌𝐿 𝛼 ) −1 − (Id + 𝜌𝐿) −1 ℎ† ∥ 2 = 𝑂 𝜌 2 ∥𝐿 − 𝐿 𝛼 ∥ 2 ∥ (Id + 𝜌𝐿) −1 𝑃⊥ ℎ† ∥ 2 .
We have that (Id + 𝜌𝐿) −1 𝑃⊥ ℎ† ∥ 2 = 𝑂 (1/𝜌 2 ). Then
𝐿 − 𝐿 𝛼 = 𝛼(𝛼Id + 𝐾 ∗ 𝐾) −1 (𝐾 ∗ 𝐾) −1 𝐹 ∗ 𝐹.
Under Assumption [E6], We obtain that (𝐾 ∗ 𝐾) −1 𝐹 ∗ 𝐹𝑃⊥ ℎ† is of regularity 𝛾 so
∥(𝐿 − 𝐿 𝛼 )𝑃⊥ ℎ† ∥ 2 = 𝑂 (𝛼 𝛾 ) .
Hence we can conclude that

∥ (Id + 𝜌𝑇𝛼 ) −1 − (Id + 𝜌𝑇) −1 ℎ† ∥ 2 = 𝑂 (𝛼 𝛾 ) .
The second term (𝐵2 ) is such that (𝐵2 ) = (𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟. We can write
−1
(𝐵2 ) = (𝐾 ∗ 𝐾) 1/2 (Id + 𝜌(𝐾 ∗ 𝐾) −1/2 𝐹 ∗ 𝐹 (𝐾 ∗ 𝐾) −1/2 ) (𝐾 ∗ 𝐾) 1/2 𝐾 ∗ 𝐾 ℎ†
= (𝐾 ∗ 𝐾) −1/2 (Id + 𝜌𝑇) −1 (𝐾 ∗ 𝐾) 1/2 ℎ† ,
where 𝑇 := (𝐾 ∗ 𝐾) −1/2 𝐹 ∗ 𝐹 (𝐾 ∗ 𝐾) −1/2 is a self-adjoint operator well-defined using

Assumption [A5]. Let
ℎ𝜌 = (𝐾 ∗ 𝐾) −1/2 (Id + 𝜌𝑇) −1 (𝐾 ∗ 𝐾) 1/2 ℎ† .
• Note first that ℎ𝜌 converges when 𝜌 → +∞ to the projection of 𝜓 := (𝐾 ∗ 𝐾) 1/2 ℎ†

onto Ker(𝑇). We can write the SVD of 𝑇 as 𝜆2𝑗 and 𝑒 𝑗 for 𝑗 ≥ 1. So we get that
∑︁ 1
(Id + 𝜌𝑇) −1 𝜓 = < 𝜓, 𝑒 𝑗 >
𝑗 ≥1
1 + 𝜌𝜆2𝑗
∑︁ 1 ∑︁
= < 𝜓, 𝑒 𝑗 > 𝑒 𝑗 + < 𝜓, 𝑒 𝑗 > 𝑒 𝑗 .
𝑗 ≥1,𝜆 𝑗 ≠0
1 + 𝜌𝜆2𝑗 𝑗 ≥1,𝜆 𝑗 =0
The last quantity converges when 𝜌 → +∞ towards the projection of 𝜓 onto the
kernel of 𝑇. Applying the operator (𝐾 ∗ 𝐾) −1/2 does not change the limit since
𝐾 ∗ 𝐾 is one to one.
• Note then that the kernel of the operator 𝑇 can be identified as follows
{𝜓 ∈ Ker(𝑇)} = {𝜓, 𝐹 (𝐾 ∗ 𝐾) −1/2 𝜓 = 0}

= {𝜓, (𝐾 ∗ 𝐾) −1/2 𝜓 ∈ Ker(𝐹)}
= {𝜓 = (𝐾 ∗ 𝐾) 1/2 ℎ, ℎ ∈ Ker(𝐹)}.
Hence ℎ𝜌 converges towards the projection of (𝐾 ∗ 𝐾) 1/2 ℎ† onto the functions

(𝐾 ∗ 𝐾) 1/2 ℎ with ℎ ∈ Ker(𝐹).
• Characterization of the projection. Note that the projection can be written as
arg min ∥ (𝐾 ∗ 𝐾) 1/2 ℎ† − (𝐾 ∗ 𝐾) 1/2 ℎ∥ 2

ℎ∈Ker(𝐹)
= arg min ∥(𝐾 ∗ 𝐾) 1/2 (ℎ† − ℎ)∥ 2

ℎ∈Ker(𝐹)
= arg min < (𝐾 ∗ 𝐾) 1/2 (ℎ† − ℎ), (𝐾 ∗ 𝐾) 1/2 (ℎ† − ℎ) >

ℎ∈Ker(𝐹)
= arg min < ℎ† − ℎ, (𝐾 ∗ 𝐾)(ℎ† − ℎ) >

ℎ∈Ker(𝐹)
= arg min ∥𝐾 (ℎ† − ℎ)∥ 2

ℎ∈Ker(𝐹)
= arg min ∥𝑟 − 𝐾 ℎ∥ 2
ℎ∈Ker(𝐹)
= ℎ 𝐾𝐹
as defined previously.
• Usual bounds enable us to prove that

1
∥ℎ𝜌 − ℎ 𝐾𝐹 ∥ 2 = 𝑂 .
𝜌2
Using all previous bounds, we can write

1
∥(𝐵) − 𝑃ℎ† ∥ 2 = 𝑂 ( + 𝛼 𝛽 + 𝛼 𝛾 ). (7.12)
𝜌2
Finally, we prove that the variance term (𝑉) is such that

∗ ∗ −1 ∗ 2 1
∥ (𝛼Id + 𝜌𝐹 𝐹 + 𝐾 𝐾) 𝐾 𝑈𝑛 ∥ = 𝑂
𝛼𝛿 𝑛
Using previous notations, we obtain
∥(𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 𝐾 ∗𝑈𝑛 ∥ = ∥ (Id + 𝜌𝐿 𝛼 ) −1 (𝛼Id + 𝐾 ∗ 𝐾) −1 𝐾 ∗𝑈𝑛 ∥

≤ ∥ (Id + 𝜌𝐿 𝛼 ) −1 ∥ ∥ (𝛼Id + 𝐾 ∗ 𝐾) −1 𝐾 ∗ ∥ ∥𝑈𝑛 ∥
1 1
≤ ∥(Id + 𝜌𝐿 𝛼 ) −1 ∥ .
𝛼 𝛿1/2
𝑛
Using that (Id + 𝜌𝐿 𝛼 ) −1 is bounded leads to the desired result.

Both bounds prove the final result for the theorem. □
Choosing the fairness constraint implies modifying the usual estimator. The
following theorem quantifies, for fixed parameters 𝜌 and 𝛼, the deviation of the fair
estimator (7.10) with respect to the unfair solution of the linear inverse problem.
Theorem 7.4 (Price for fairness)
𝜌
∥ℎ 𝛼 − ℎ 𝛼,𝜌 ∥ = 𝑂
𝛼2
Proof
∥ℎ 𝛼 − ℎ 𝛼,𝜌 ∥
≤ ∥(𝛼Id + 𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟 − (𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 𝐾 ∗ 𝑟 ∥
≤ ∥ [(𝛼Id + 𝐾 ∗ 𝐾) −1 − (𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 ]𝐾 ∗ 𝑟 ∥.
Using that for two operators
𝐴−1 − 𝐵−1 = 𝐴−1 (𝐵 − 𝐴)𝐵−1
we obtain
∥ℎ 𝛼 − ℎ 𝛼,𝜌 ∥ ≤ ∥(𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 𝜌𝐹 ∗ 𝐹 (𝛼Id + 𝐾 ∗ 𝐾) −1 ∥ (7.13)
Now using that

1
∥(𝛼Id + 𝐾 ∗ 𝐾) −1 ∥ ≤
𝛼
1
(𝛼Id + 𝜌𝐹 ∗ 𝐹 + 𝐾 ∗ 𝐾) −1 ≤
𝛼
and since
∥𝐾 ∗ 𝑟 ∥ ≤ 𝑀
leads to the result. □
The previous theorem suggests that in a decision-making process, the stakeholder’s
choice can be cast as a function of the parameter 𝜌. For fixed 𝛼, as the parameter
𝜌 diverges to ∞, the fully fair solution is imposed. However, there is also a loss of
accuracy in prediction, which increases as 𝜌 diverges. Imposing anapproximately
fair solution can help balance the benefit of fairness, which in certain situations can
have a clear economic and reputational cost, with the statistical loss associated with a
worse prediction. This trade-off could also be used to determine an optimal choice
for the parameter 𝜌. We illustrate this procedure in Section 7.6.
7.5 Estimation with an Exogenous Binary Sensitive Attribute
We discuss the estimation and the finite sample implementation of our method in the
simple case when 𝑆 is an exogenous binary random variable (for instance, gender or
race), and 𝑍 ∈ R 𝑝 only contains continuous endogenous regressors. This framework
can be easily extended to the case when 𝑆 is an endogenous multivariate categorical
variable and to include additional exogenous components in 𝑍 (Hall & Horowitz,
2005; Centorrino, Fève & Florens, 2017; Centorrino & Racine, 2017). Our statistical
model can be written as
𝑌 = ℎ0 (𝑍) + ℎ1 (𝑍)𝑆 + 𝑈 = S′ ℎ(𝑍) + 𝑈, (7.14)

] ′,
where ℎ = [ℎ0 ℎ1 and S = [1 𝑆] ′.
This model is a varying coefficient model (see, among others, Hastie & Tibshirani,
1993; Fan & Zhang, 1999; Li, Huang, Li & Fu, 2002). Adopting the terminology that
is used in this literature, we refer to S as the ‘linear’ variables (or predictors), and
to the 𝑍’s as the ‘smoothing’ variables (or covariates) (Fan & Zhang, 2008). When
𝑍 is endogenous, Centorrino and Racine (2017) have studied the identification and
estimation of this model with instrumental variables. That is, we assume there is a
random vector 𝑊 ∈ R𝑞 , such that 𝐸 [S𝑈|𝑊] = 0, and
𝐸 [SS′ ℎ(𝑍)|𝑊] = 0 ⇒ ℎ = 0, (7.15)
where equalities are intended almost surely.3 The completeness condition in equation
(7.15) is necessary for identification, and it is assumed to hold. As proven in Centorrino
and Racine (2017), this condition is implied by the injectivity of the conditional
expectation operator (see our Assumption A2), and by the matrix 𝐸 [SS′ |𝑧, 𝑤] being
full rank for almost every (𝑧, 𝑤).
We would like to obtain a nonparametric estimator of the functions {ℎ0 , ℎ1 } when
a fairness constraint is imposed. We use the following operator’s notations
(𝐾𝑠 ℎ) (𝑤) =𝐸 [SS′ ℎ(𝑍)|𝑊 = 𝑤]

𝐾𝑠∗ 𝜓 (𝑧) =𝐸 [SS′𝜓(𝑊)|𝑍 = 𝑧]

(𝐾 ∗ 𝜓) (𝑧) =𝐸 [𝜓(𝑊)|𝑍 = 𝑧] ,
for every ℎ ∈ 𝐿 2 (𝑍), and 𝜓 ∈ 𝐿 2 (𝑊).

When no fairness constraint is imposed, the regularized approximation to the pair
{ℎ0 , ℎ1 } is given by
ℎ 𝛼 = arg min ∥𝐾𝑠 ℎ − 𝑟 ∥ 2 + 𝛼∥ℎ∥ 2 , (7.16)
ℎ∈𝐿 2 (𝑍)
where ∥ℎ∥ 2 = ∥ℎ0 ∥ 2 + ∥ℎ1 ∥ 2 . That is
3 Notice that the moment conditions 𝐸 [S𝑈 |𝑊 ] = 0 are implied by the assumption that 𝐸 [𝑈 |𝑊 , 𝑆 ] =
0, although they allow to exploit the semiparametric structure of the model and reduce the curse of
dimensionality (see Centorrino & Racine, 2017)
−1
ℎ 𝛼 = 𝛼𝐼 + 𝐾𝑠∗ 𝐾𝑠 𝐾𝑠∗ 𝑟, (7.17)
with 𝑟 (𝑤) = 𝐸 [S𝑌 |𝑊 = 𝑤].

As in Centorrino and Racine (2017), the quantities in equation (7.17) can be
replaced by consistent estimators. Let {(𝑌𝑖 , 𝑋𝑖 ,𝑊𝑖 ), 𝑖 = 1, . . . , 𝑛} be an iid sample
from the joint distribution of (𝑌 , 𝑋,𝑊). We denote by
 
𝑌1 
 
 
𝑌2  h i
Y𝑛 =  .  , S𝑛 = 𝐼𝑛 𝑑𝑖𝑎𝑔(𝑆1 , 𝑆2 , . . . , 𝑆 𝑛 ) ,
 .. 
 
 
𝑌𝑛 
 
the 𝑛 × 1 vector, which stacks the observations of the dependent variable and the
𝑛 × 2𝑛 matrix of predictors, where 𝐼𝑛 is the identity matrix of dimension 𝑛, and
𝑑𝑖𝑎𝑔(𝑆1 , 𝑆2 , . . . , 𝑆 𝑛 ) is a 𝑛 × 𝑛 diagonal matrix whose diagonal elements are equal to
the sample observations of the sensitive attribute 𝑆. Similarly, we let
 1 − 𝑆1 
   
 𝑆1 
   
 1 − 𝑆2 
   
 𝑆2 
D1,𝑛 =  .  , and D0,𝑛 =  ..  ,
 
 ..   . 
   
   
1 − 𝑆 𝑛 
𝑆 𝑛 
 
 
two 𝑛 × 1 vectors stacking the sample observations of 𝑆 and 1 − 𝑆. ∫
Finally, let 𝐶 (·) a univariate kernel function, such that 𝐶 (·) ≥ 0, and 𝐶 (𝑢)𝑑𝑢 = 1,
h i′
and C(·) be a multivariate product kernel. That is, for a vector u = 𝑢 1 𝑢2 . . . 𝑢 𝑝 ,
with 𝑝 ≥ 1, C(u) = 𝐶 (𝑢 1 ) × 𝐶 (𝑢 2 ) × · · · × 𝐶 (𝑢 𝑝 ).
As detailed in Centorrino et al. (2017), the operators 𝐾 and 𝐾 ∗ can be approximated
by finite-dimensional matrices of kernel weights. In particular, we have that
h i 𝑛 h i 𝑛
𝑊 −𝑊 c∗ = C 𝑍𝑖 −𝑍 𝑗
𝐾ˆ = C 𝑖𝑎 𝑗 and 𝐾 𝑎𝑍
,
|{z} 𝑊 𝑖, 𝑗=1 |{z} 𝑖, 𝑗=1
𝑛×𝑛 𝑛×𝑛
where 𝑎 𝑊 and 𝑎 𝑍 are bandwidth parameters, chosen in such a way that 𝑎 𝑊 , 𝑎 𝑍 → 0,

as 𝑛 → ∞. Therefore,
ˆ 𝑛′ Y𝑛

𝑟ˆ =𝑣𝑒𝑐 (𝐼2 ⊗ 𝐾)S
ˆ 𝑛′ S𝑛
𝐾ˆ 𝑠 =(𝐼2 ⊗ 𝐾)S
c∗ )S𝑛′ S𝑛 ,
c∗ 𝑠 =(𝐼2 ⊗ 𝐾
𝐾
in a way that
h i −1
ℎˆ 𝛼 = ℎˆ 0, 𝛼 ℎˆ 1, 𝛼 = (𝑣𝑒𝑐(𝐼𝑛 ) ′ ⊗ 𝐼𝑛 ) 𝐼𝑛 ⊗ 𝛼𝐼 + 𝐾
c∗ 𝑠 𝐾ˆ 𝑠 c∗ 𝑠 𝑟ˆ .
𝐾 (7.18)
As explained above, the fairness constrain can be characterized by a linear operator

𝐹 𝑗 , such that 𝐹 𝑗 ℎ = 0, where 𝑗 = {1, 2}. In the case of Definition 7.1, and exploiting
the binary nature of 𝑆, the operator 𝐹1 can be approximated by
 
 0𝑛 0𝑛 
𝐹1,𝑛 = 
 −1 −1 −1 
′ D

′ − D′ D ′ ′ D ′ 
,
|{z} 𝜄𝑛 D1,𝑛 1,𝑛 D1,𝑛 0,𝑛 0,𝑛 D0,𝑛 𝜄𝑛 D1,𝑛 1,𝑛 D1,𝑛 
2𝑛×2𝑛  
where 𝜄𝑛 is a 𝑛 × 1 vector of ones, and 0𝑛 is a 𝑛 × 𝑛 matrix of zeroes.
In the case of definition 7.2, the fairness operator can be approximated by
 
0 𝑛 0 𝑛 
𝐹2,𝑛 =   ,
|{z} 0𝑛 𝐼𝑛 
2𝑛×2𝑛
 
In both cases, when the function ℎ ∈ F 𝑗 , we obviously have that 𝐹 𝑗 𝑣𝑒𝑐(ℎ) = 0, with
𝑗 = {1, 2}.
As detailed in Section 7.4, and for 𝑗 = {1, 2}, the estimator consistent with the
fairness constraint can be obtained in several ways
1) By projecting the unconstrained estimator in (7.18) onto the null space of 𝐹 𝑗 . Let
𝑃 𝑗,𝑛 be the estimator of such projection, then we have that

ℎˆ 𝛼,𝐹, 𝑗 = (𝑣𝑒𝑐(𝐼𝑛 ) ′ ⊗ 𝐼𝑛 ) 𝐼𝑛 ⊗ 𝑃 𝑗,𝑛 𝑣𝑒𝑐( ℎˆ 𝛼 ) , (7.19)
2) By restricting the conditional expectation operator to project onto the null space
of 𝐹 𝑗 . Let
𝐾ˆ 𝐹, 𝑗,𝑠 = 𝐾ˆ 𝑠 𝑃 𝑗.𝑛 , and 𝐾

c∗ 𝐹, 𝑗,𝑠 = 𝑃 𝑗,𝑛 𝐾
c∗ 𝑠 ,
then
−1
ˆℎ 𝛼,𝐾𝐹 , 𝑗 = (𝑣𝑒𝑐(𝐼𝑛 ) ′ ⊗ 𝐼𝑛 ) 𝐼𝑛 ⊗ 𝛼𝐼 + 𝐾
c∗ 𝐹, 𝑗,𝑠 𝐾ˆ 𝐹, 𝑗,𝑠 ∗
𝐾 𝐹, 𝑗,𝑠 𝑟ˆ ,
c (7.20)
3) By modifying the objecting function to include an additional term which penalizes

deviations from fairness. That is, we let
ℎˆ 𝛼,𝜌, 𝑗 = arg min ∥ 𝐾ˆ 𝑠 ℎ − 𝑟ˆ ∥ 2 + 𝛼∥ℎ∥ 2 + 𝜌∥𝐹 𝑗,𝑛 ℎ∥ 2 ,

ℎ∈ F 𝑗
in a way that
−1
ℎˆ 𝛼,𝜌, 𝑗 = 𝛼𝐼𝑛 + 𝜌𝐹 𝑗,𝑛 𝐹 𝑗,𝑛 + 𝐾
c∗ 𝑠 𝐾ˆ 𝑠 c∗ 𝑠 𝑟.
𝐾 ˆ (7.21)
For 𝜌 = 0, this estimator is equivalent to the unconstrained estimator, ℎˆ 𝛼 , and,

for 𝜌 sufficiently large, it imposes the full fairness constraint.
To implement the estimators above, we need to select several smoothing, {𝑎 𝑊 , 𝑎 𝑍 },
and regularization, {𝛼, 𝜌}, parameters. For the choice of the tuning parameters
{𝑎 𝑊 , 𝑎 𝑍 , 𝛼}, we follow Centorrino (2016) and use a sequential leave-one-out cross-
validation approach. We instead select the regularization parameter 𝜌, for 𝑗 = {1, 2}
as
𝜌 ∗𝑗 = arg min ∥ ℎˆ 𝛼,𝜌, 𝑗 − ℎˆ 𝛼 ∥ 2 + 𝜍 ∥𝐹 𝑗,𝑛 ℎˆ 𝛼,𝜌, 𝑗 ∥ 2 , (7.22)

𝜌
with 𝜍 > 0 a constant. The first term of this criterion function is a statistical
loss that we incur when we impose the fairness constraint. The second term instead
represents the distance of our estimator to full fairness. The smaller the norm of
the second term, the closer we are to obtaining a fair estimator. For instance, if our
unconstrained estimator, ℎˆ 𝛼 is fair, then the second term will be identically zero for
any value of 𝜌, while the first term will be zero for 𝜌 = 0, and then would increase as
𝜌 → ∞. The constant 𝜍 serves as a subjective weight for fairness. In principle, one
could set 𝜍 = 1. Values of 𝜍 higher than 1 imply that the decision-maker considers
deviations from fairness to be costly and thus prefers them to be penalized more
heavily. The opposite is true for values of 𝜍 < 1.
7.6 An Illustration
We consider the following illustration of the model described in the previous Section.
We generate a random vector 𝜏 = (𝜏1 , 𝜏2 ) ′ from a bivariate normal distribution with
mean (0, 0.5) ′ and covariance matrix equal to
 
 1 2 sin(𝜋/12) 
Σ 𝜏 =  .

2 sin(𝜋/12) 1 
 
Then, we fix
𝑊 = − 1 + 2Φ(𝜏1 )
𝑆 =𝐵(Φ(𝜏2 )),
where 𝐵(·) is a Bernoulli distribution with probability parameter equal to Φ(𝜏2 ), and
Φ is the cdf of a standard normal distribution.
We then let 𝜂 and 𝑈 to be independent normal random variables with mean 0 and
variances equal to 0.16 and 0.25, respectively, and we generate
𝑍 = −1 + 2Φ (𝑊 − 0.5𝑆 − 0.5𝑊 𝑆 + 0.5𝑈 + 𝜂) ,

and
𝑌 = ℎ0 (𝑍) + ℎ1 (𝑍)𝑆 + 𝑈,
where ℎ0 (𝑍) = 3𝑍 2 , and
ℎ1 (𝑍) = 1 − 5𝑍 3 .
In this illustration, the random variable 𝑍 can be thought to be an observable
characteristic of the individual, while 𝑆 could be a sensitive attribute related, for
instance, to gender or ethnicity. Notice that the true regression function is not fair in
the sense of either Definition 7.1 or Definition 7.2. This reflects the fact that real data
may contain a bias with respect to the sensitive attribute, which is often the case in
practice. We fix the sample size at 𝑛 = 1000, and we use Epanechnikov kernels for
estimation.
1.0
0.8
0.6
0.4
0.2
0.0
−0.5 0.0 0.5
Fig. 7.2: Empirical CDF of the endogenous regressor 𝑍, conditional of the sensitive
attribute 𝑆. CDF of 𝑍 |𝑆 = 0, solid gray line; CDF of 𝑍 |𝑆 = 1, solid black line.
In Figure 7.2, we plot the empirical cumulative distribution function (CDF) of 𝑍

given 𝑆 = 0 (solid gray line), and of 𝑍 given 𝑆 = 1 (solid black line). We can see that
the latter stochastically dominates the former. This can be interpreted as the fact that
systematic differences in group characteristics can generate systematic differences in
the outcome, 𝑌 , even when the sensitive attribute 𝑆 is not directly taken into account.
We compare the unconstrained estimator, ℎˆ 𝛼 , with the fairness-constrained
estimators in the sense of Definitions 7.1 and 7.2.
In Figures 7.3 and 7.4, we plot the estimators of the functions {ℎ0 , ℎ1 }, under the
fairness constraints in Definitions 7.1 and 7.2, respectively. Notice that, as expected,
the estimator which imposes approximate fairness through the penalization parameter
𝜌 lays somewhere in between the unconstrained estimator and the estimators which
impose full fairness.
In Figure 7.5, we depict the objective function in equation (7.22) for the optimal
choice of 𝜌, using both Definition 7.1 (left panel) and Definition 7.2 (right panel).
The optimal value of 𝜌 is obtained in our case by fixing 𝜍 = 1 (solid black line).
4
3
2
2
0
−2
1
−4
0
−0.5 0.0 0.5 −0.5 0.0 0.5
(a) ℎ0 ( 𝑥) = 3𝑥 2 (b) ℎ1 ( 𝑥) = 1 − 5𝑥 3
Fig. 7.3: Estimation using the definition of fairness in 7.1. Solid black line, true
function; dotted black line, true function with fairness constraint; solid gray line, ℎˆ 𝛼 ;
dashed gray line, ℎˆ 𝛼,𝐹 ; dashed-dotted gray line, ℎˆ 𝛼,𝐾𝐹 ; dashed-dotted light-gray line,
ℎˆ 𝛼,𝜌 .
4
4
3
2
2
0
−2
1
−4
0
−0.5 0.0 0.5 −0.5 0.0 0.5
(a) ℎ0 ( 𝑥) = 3𝑥 2 (b) ℎ1 ( 𝑥) = 1 − 5𝑥 3
Fig. 7.4: Estimation using the definition of fairness in 7.2. Solid black line, true
function; dotted black line, true function with fairness constraint; solid gray line, ℎˆ 𝛼 ;
dashed gray line, ℎˆ 𝛼,𝐹 ; dashed-dotted gray line, ℎˆ 𝛼,𝐾𝐹 ; dashed-dotted light-gray line,
ℎˆ 𝛼,𝜌 .
2.0
2.0
1.5
1.5
1.0
1.0
0.5
0.5
0.00 ρ*1 0.05 0.10 0.15 0.20 0.00 ρ*2 0.05 0.10 0.15 0.20
(a) Definition 7.1 (b) Definition 7.2
Fig. 7.5: Choice of the optimal value of 𝜌.

However, if a decision-maker wished to impose more fairness, this could be achieved

by setting 𝜍 > 1. For illustrative purposes, we also report the objective function when
𝜍 = 2 (solid gray line). It can be seen that this leads to a larger value of 𝜌 ∗ , but also
that the objective function tends to flatten out.
2000
2000
1500
1500
1000
1000
500
500
0
0
0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20
Fig. 7.6: Cost and benefit of fairness as a function of the penalization parameter 𝜌.
We also present in Figure 7.6 the trade-off between the statistical loss (solid black
line), ∥ ℎˆ 𝛼,𝜌, 𝑗 − ℎˆ 𝛼 ∥ 2 , which can be interpreted as the cost of imposing a fair solution,
and the benefit of fairness (solid gray line), which is measured by the squared norm of
𝐹𝑛, 𝑗 ℎˆ 𝛼,𝜌, 𝑗 , when 𝑗 = {1, 2} to reflect both Definitions 7.1 (left panel) and 7.2 (right
panel). In both cases, we fix 𝜍 = 1. The upward-sloping line is the squared deviation
from the unconstrained estimator, which increases with 𝜌. The downward-sloping
curve is the norm of the projection of the estimator onto the space of fair functions,
which converges to zero as 𝜌 increases.
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0
0.0
−2 0 2 4 6 −2 0 2 4 6
Fig. 7.7: Density of the predicted values from the constrained models. Solid lines
represent group 𝑆 = 0, and dashed-dotted lines group 𝑆 = 1. Black lines are the
densities of the observed data; dark-gray lines are from constrained model 1; gray
lines from constrained model 2; and light-gray from constrained model 3.
Finally, it is interesting to assess how the different definitions of fairness and

the different implementations affect the distribution of the predicted values. This
prediction is made in-sample as its goal is not to assess the predictive properties of
our estimator but rather to determine how the different definitions of fairness and the
various ways to impose the fairness constraint in estimation affect the distribution of
the model predicted values.
The black lines in Figure 7.7 represent the empirical CDF of the dependent
variable 𝑌 for 𝑆 = 0 (solid black line), and 𝑆 = 1 (dashed-dotted black line). This is
compared with the predictions using estimators 1 (dark-gray lines), 2 (gray lines),
and 3 (light-gray lines). In the data, the distribution of 𝑌 given 𝑆 = 1 stochastically
dominates the distribution of 𝑌 given 𝑆 = 0.
Notice that in the case of fairness as defined in 7.1, the estimator which modifies
the conditional expectation operator to project directly onto the space of fair functions
seems to behave best in terms of fairness, as the distribution of the predicted values
for groups 0 and 1 are very similar. The estimator, which imposes approximate
fairness, obviously lies somewhere in between the data and the previous estimator.
The projection of the unconstrained estimator onto the space of fair functions does
not seem to deliver an appropriate distribution of the predicted values. What happens
is that this estimator penalizes people in group 1 with low values of 𝑍, to maintain
fairness on the average while preserving a substantial difference in the distribution of
the two groups.
Differently, in the case of fairness, as defined in 7.2, the projection of the
unconstrained estimator seems to behave best. However, this may be because the
distribution of 𝑍 given 𝑆 = 0 and 𝑆 = 1 are substantially similar. If, however, there is
more difference in the observable characteristics by group, this estimator may not
behave as intended.
7.7 Conclusions
In this chapter, we consider the issue of estimating a structural econometrics model

when a fairness constraint is imposed on the solution. We focus our attention on
models when the function is the solution to a linear inverse problem, and the
fairness constraint is imposed on the included covariates and can be expressed as
a linear restriction on the function of interest. We also discuss how to construct an
approximately fair solution to a linear functional equation and how this notion can
be implemented to balance accurate predictions with the benefits of a fair machine
learning algorithm. We further present regularity conditions under which the fair
approximation converges towards the projection of the true function onto the null
space of the fairness operator. Our leading example is a nonparametric instrumental
variable model, in which the fairness constraint is imposed. We detail the example
of such a model when the sensitive attribute is binary and exogenous (Centorrino &
Racine, 2017).
The framework introduced in this chapter can be extended in several directions.

The first significant extension would be to consider models in which the function ℎ†
is the solution to a nonlinear equation. The latter can arise, for instance, when the
conditional mean independence restriction is replaced with full independence between
the instrumental variable and the structural error term (Centorrino, Fève & Florens,
2019; Centorrino & Florens, 2021). Moreover, one can potentially place fairness
restrictions directly on the decision algorithm or on the distribution of predicted
values. These restrictions usually imply that the fairness constraint is nonlinear, and a
different identification and estimation approach should be employed.
In this work, we restrict to fairness as a group notion and do not consider fairness
at an individual level as in Kusner, Loftus, Russell and Silva (2017), or De Lara,
González-Sanz, Asher and Loubes (2021). This framework could enable a deeper
understanding of fairness in econometrics from a causal point of view.
Finally, the fairness constraint imposed in this paper is limited to the regression
function, ℎ. However, other constraints may be imposed directly on the functional
equation. For instance, on the selection of the instrumental variables, which will be
the topic of future work.
Acknowledgements Jean-Pierre Florens acknowledges funding from the French National Research
Agency (ANR) under the Investments for the Future program (Investissements d’Avenir, grant
ANR-17-EURE-0010).
References
Angwin, J., Larson, J., Mattu, S. & Kirchner, L. (2016). Machine bias risk assessments
in criminal sentencing. ProPublica, May, 23.
Barocas, S. & Selbst, A. D. (2016). Big data’s disparate impact. Calif. L. Rev., 104,
671.
Besse, P., del Barrio, E., Gordaliza, P., Loubes, J.-M. & Risser, L. (2021). A survey of
bias in machine learning through the prism of statistical parity. The American
Statistician, 1–11.
Carrasco, M., Florens, J.-P. & Renault, E. (2007). Linear inverse problems in structural
econometrics estimation based on spectral decomposition and regularization.
In J. Heckman & E. Leamer (Eds.), Handbook of econometrics (p. 5633-5751).
Elsevier.
Centorrino, S. (2016). Data-Driven Selection of the Regularization Parameter
in Additive Nonparametric Instrumental Regressions. Mimeo - Stony Brook
University.
Centorrino, S., Fève, F. & Florens, J.-P. (2017). Additive Nonparametric Instrumental
Regressions: a Guide to Implementation. Journal of Econometric Methods,
6(1).
Centorrino, S., Fève, F. & Florens, J.-P. (2019). Nonparametric Instrumental
Regressions with (Potentially Discrete) Instruments Independent of the Error
References 249
Term. Mimeo - Stony Brook University.

Centorrino, S. & Florens, J.-P. (2021). Nonparametric estimation of accelerated
failure-time models with unobservable confounders and random censoring.
Electronic Journal of Statistics, 15(2), 5333 – 5379.
Centorrino, S. & Racine, J. S. (2017). Semiparametric Varying Coefficient Models
with Endogenous Covariates. Annals of Economics and Statistics(128), 261–
295.
Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in
recidivism prediction instruments. Big data, 5(2), 153–163.
Chzhen, E., Denis, C., Hebiri, M., Oneto, L. & Pontil, M. (2020). Fair regression
with wasserstein barycenters. arXiv preprint arXiv:2006.07286.
Darolles, S., Fan, Y., Florens, J. P. & Renault, E. (2011). Nonparametric Instrumental
Regression. Econometrica, 79(5), 1541–1565.
De-Arteaga, M., Romanov, A., Wallach, H., Chayes, J., Borgs, C., Chouldechova, A.,
. . . Kalai, A. T. (2019). Bias in bios: A case study of semantic representation
bias in a high-stakes setting. In Proceedings of the Conference on Fairness,
Accountability, and Transparency (pp. 120–128).
De Lara, L., González-Sanz, A., Asher, N. & Loubes, J.-M. (2021). Transport-based
counterfactual models. arXiv preprint arXiv:2108.13025.
Engl, H. W., Hanke, M. & Neubauer, A. (1996). Regularization of inverse problems
(Vol. 375). Springer Science & Business Media.
Fan, J. & Zhang, W. (1999). Statistical Estimation in Varying Coefficient Models.
Ann. Statist., 27(5), 1491–1518.
Fan, J. & Zhang, W. (2008). Statistical Methods with Varying Coefficient Models.
Statistics and Its Interface, 1, 179–195.
Florens, J. P., Heckman, J. J., Meghir, C. & Vytlacil, E. (2008). Identification
of Treatment effects using Control Functions in Models with Continuous,
Endogenous Treatment and Heterogenous Effects. Econometrica, 76(5), 1191–
1206.
Florens, J. P., Mouchart, M. & Rolin, J. (1990). Elements of Bayesian Statistics. M.
Dekker.
Florens, J.-P., Racine, J. & Centorrino, S. (2018). Nonparametric Instrumental
Variable Derivative Estimation. Jounal of Nonparametric Statistics, 30(2),
368-391.
Friedler, S. A., Scheidegger, C. & Venkatasubramanian, S. (2021, mar). The
(im)possibility of fairness: Different value systems require different mechanisms
for fair decision making. Commun. ACM, 64(4), 136–143. Retrieved from
https://doi.org/10.1145/3433949 doi: 10.1145/3433949
Gordaliza, P., Del Barrio, E., Fabrice, G. & Loubes, J.-M. (2019). Obtaining
fairness using optimal transport theory. In International conference on machine
learning (pp. 2357–2365).
Hall, P. & Horowitz, J. L. (2005). Nonparametric Methods for Inference in the
Presence of Instrumental Variables. Annals of Statistics, 33(6), 2904–2929.
Hastie, T. & Tibshirani, R. (1993). Varying-Coefficient Models. Journal of the Royal
Statistical Society. Series B (Methodological), 55(4), 757-796.
Hoda, H., Loi, M., Gummadi, K. P. & Krause, A. (2018). A moral framework for
understanding of fair ml through economic models of equality of opportunity.
Machine Learning(a).
Hu, L. & Chen, Y. (2020). Fair classification and social welfare. In Proceedings of the
2020 conference on fairness, accountability, and transparency (pp. 535–545).
Jiang, R., Pacchiano, A., Stepleton, T., Jiang, H. & Chiappa, S. (2020). Wasserstein
fair classification. In Uncertainty in artificial intelligence (pp. 862–872).
Kasy, M. & Abebe, R. (2021). Fairness, Equality, and Power in Algorithmic
Decision-Making. In Proceedings of the 2021 ACM Conference on Fairness,
Accountability, and Transparency (pp. 576–586). New York, NY, USA:
Association for Computing Machinery.
Kusner, M. J., Loftus, J., Russell, C. & Silva, R. (2017). Counterfactual fairness.
Advances in neural information processing systems, 30.
Lee, M. S. A., Floridi, L. & Singh, J. (2021). Formalising trade-offs beyond
algorithmic fairness: lessons from ethical philosophy and welfare economics.
AI and Ethics, 1–16.
Le Gouic, T., Loubes, J.-M. & Rigollet, P. (2020). Projection to fairness in statistical
learning. arXiv e-prints, arXiv–2005.
Li, Q., Huang, C. J., Li, D. & Fu, T.-T. (2002). Semiparametric Smooth Coefficient
Models. Journal of Business & Economic Statistics, 20(3), 412-422.
Loubes, J.-M. & Rivoirard, V. (2009). Review of rates of convergence and regularity
conditions for inverse problems. Int. J. Tomogr. Stat, 11(S09), 61–82.
McIntyre, F. & Simkovic, M. (2018). Are law degrees as valuable to minorities?
International Review of Law and Economics, 53, 23–37.
Menon, A. K. & Williamson, R. C. (2018, 2). The cost of fairness in binary
classification. In S. A. Friedler & C. Wilson (Eds.), Proceedings of the 1st
conference on fairness, accountability and transparency (Vol. 81, pp. 107–118).
New York, NY, USA: PMLR.
Natterer, F. (1984). Error bounds for tikhonov regularization in hilbert scales.
Applicable Analysis, 18(1-2), 29–37.
Newey, W. K. & Powell, J. L. (2003). Instrumental Variable Estimation of Nonpara-
metric Models. Econometrica, 71(5), 1565–1578.
Oneto, L. & Chiappa, S. (2020). Fairness in machine learning. In L. Oneto,
N. Navarin, A. Sperduti & D. Anguita (Eds.), Recent trends in learning
from data: Tutorials from the inns big data and deep learning conference
(innsbddl2019) (pp. 155–196). Cham: Springer International Publishing. doi:
10.1007/978-3-030-43883-8_7
Rambachan, A., Kleinberg, J., Mullainathan, S. & Ludwig, J. (2020). An economic
approach to regulating algorithms (Tech. Rep.). Cambridge, MA: National
Bureau of Economic Research.
Risser, L., Sanz, A. G., Vincenot, Q. & Loubes, J.-M. (2019). Tackling algorithmic
bias in neural-network classifiers using wasserstein-2 regularization. arXiv
preprint arXiv:1908.05783.
Chapter 8
Graphical Models and their Interactions with
Machine Learning in the Context of Economics
and Finance
Ekaterina Seregina
Abstract Many economic and financial systems, including financial markets, financial
institutions, and macroeconomic policy making can be modelled as systems of
interacting agents. Graphical models, which are the main focus of this chapter, are
a means of estimating the relationships implied by such systems. The main goals
of this chapter are (1) acquainting the readers with graphical models; (2) reviewing
the existing research on graphical models for economic and finance problems; (3)
reviewing the literature that merges graphical models with other machine learning
methods in economics and finance.
8.1 Introduction
Technological advances have made large data sets available for scientific discovery.
Extracting information about interdependence between many variables in rich data
sets plays an important role in various applications, such as portfolio management,
risk assessment, forecast combinations, classification, as well as running generalized
least squares regressions on large cross-sections, and choosing an optimal weighting
matrix in the general method of moments (GMM). The goal of estimating variable
dependencies can be formulated as a search for a covariance matrix estimator that
contains pairwise relationships between variables. However, as shown in this chapter,
in many applications what is required is not a covariance matrix, but its inverse, which
is known as a precision matrix. Instead of direct pairwise dependencies, precision
matrix contains the information about partial pairwise dependencies conditional on
the remaining variables.
Originating from the literature on high-dimensional statistics and machine learning,
graphical models search for the estimator of precision matrix. Building on Hastie,
Tibshirani and Friedman (2001); Pourahmadi (2013) and Bishop (2006), we review
Ekaterina SereginaB
Colby College, Waterville, ME, USA e-mail: eseregin@colby.edu
252 Seregina
the terminology used in the literature on graphical models, such as vertices, edges,
directed and undirected graphs, partial correlations, sparse and fully connected graphs.
We explain the connection between entries of the precision matrix, partial correlations,
graph sparsity and the implications of graph structure for common economic and
finance problems. We further review several prominent approaches for estimating a
high-dimensional graph: Graphical LASSO (Friedman, Hastie & Tibshirani, 2007),
nodewise regression (Meinshausen & Bühlmann, 2006) and CLIME (T. Cai, Liu
& Luo, 2011); as well as several computational methods to recover the entries of
precision matrix including stochastic gradient descent and the alternating direction
method of multipliers (ADMM).
Having introduced readers to graphical models, we proceed by reviewing the
literature that applies this tool to tackle economic and finance problems. Such
applications include portfolio construction, see Brownlees, Nualart and Sun (2018),
Barigozzi, Brownlees and Lugosi (2018), Koike (2020), Callot, Caner, Önder and
Ulasan (2019) among others, and forecast combination for macroeconomic forecasting
Lee and Seregina (2021a). We elaborate on what is common between such applications
and why sparse precision matrix estimator is desirable in these settings. Further, we
highlight the drawbacks encountered by early models that used graphical models
in economic or finance context. Noticeably, such drawbacks are caused by the lack
of reconciliation between models borrowed from statistical literature and stylized
economic facts observed in practice. The chapter proceeds by reviewing most recent
studies that propose solutions to overcome these drawbacks.
In the concluding part of the chapter, we aspire to demonstrate that graphical
models are not a stand-alone tool and there are several promising directions how they
can be integrated with other machine learning methods in the context of economics
and finance. Along with reviewing the existing research that has attempted to tackle
the aforementioned issue, we outline several directions that have been examined to a
lesser extent but which we deem to be worth exploring.
8.1.1 Notation
Throughout the chapter S 𝑝 denote the set of all 𝑝 × 𝑝 symmetric matrices, and
S ++
𝑝 denotes the set of all 𝑝 × 𝑝 positive definite matrices. For any matrix C, its
(𝑖, 𝑗)-th element is denoted as 𝑐 𝑖 𝑗 . Given a vector u ∈ R𝑑 and parameter 𝑎 ∈ [1, ∞),
let ∥u∥ 𝑎 denote ℓ𝑎 -norm. Given a matrix U ∈ S 𝑝 , let Λmax (U) ≡ Λ1 (U) ≥ Λ2 (U) ≥
. . . ≥ Λmin (U) ≡ Λ 𝑝 (U) be the eigenvalues of U, and eig𝐾 (U) ∈ R𝐾× 𝑝 denote the
first 𝐾 ≤ 𝑝 normalized eigenvectors corresponding to Λ1 (U), . . . , Λ𝐾 (U). Given
parameters 𝑎, 𝑏 ∈ [1, ∞), let |||U||| 𝑎,𝑏 ≡ max ∥y ∥ 𝑎 =1 ∥Uy∥ 𝑏 denote the induced matrix-
Í 𝑁
operator norm. The special cases are |||U||| 1 ≡ max1≤ 𝑗 ≤ 𝑁 𝑖=1 𝑢 𝑖, 𝑗 for the ℓ1 /ℓ1 -
operator norm; the operator norm (ℓ2 -matrix norm) |||U||| 22 ≡ Λ (UU′) is equal
Í 𝑁 max
to the maximal singular value of U; |||U||| ∞≡ max1≤ 𝑗 ≤𝑁 𝑖=1 𝑢 𝑗,𝑖 for the ℓ∞ /ℓ∞ -
operator norm. Finally, ∥U∥ max ≡ max𝑖, 𝑗 𝑢 𝑖, 𝑗 denotes the element-wise maximum,
8 Graphical Models and Machine Learning in the Context of Economics and Finance 253
and |||U||| 2𝐹 ≡ 𝑖, 𝑗 𝑢 2𝑖, 𝑗 denotes the Frobenius matrix norm. Additional more specific
Í
notations are introduced throughout the chapter.
8.2 Graphical Models: Methodology and Existing Approaches
A detailed coverage of the terminology used in the network theory literature was
provided in Chapter 6. We start with a brief summary to refresh and empasize the
concepts that we use for studying graphical models. A graph consists of a set of
vertices (nodes) and a set of edges (arcs) that join some pairs of the vertices. In
graphical models, each vertex represents a random variable, and the graph visualizes
the joint distribution of the entire set of random variables. Figure 8.1 shows a
simplified example of a graph for five random variables, and Figure 8.2 depicts a
larger network that has several color-coded clusters.
The graphs can be separated into undirected graphs-where the edges have no
directional arrows, and directed graphs-where the edges have a direction associated
with them. Directed graphs are useful for expressing causal relationship between
random variables (see Pearl, 1995; Verma & Pearl, 1990 among others), whereas
undirected graphs are used to study the conditional dependencies between variables.
As shown later, the edges in a graph are parameterized by potentials (values) that
encode the strength of the conditional dependence between the random variables at
the corresponding vertices. Sparse graphs have a relatively small number of edges.
Among the main challenges in working with graphical models are choosing the
structure of the graph (model selection) and estimation of the edge parameters from
the data.
Fig. 8.1: A sample network: zoom in Fig. 8.2: A sample network: zoom out
Let 𝑋 = (𝑋1 , . . . , 𝑋 𝑝 ) be random variables that have a multivariate Gaussian

distribution with an expected value 𝝁 and covariance matrix 𝚺, 𝑋 ∼ N ( 𝝁, 𝚺). We
254 Seregina
note that even though the normality assumption is not required for estimating graphical
models, it helps us illustrate the relationship between partial correlations and the
entries of precision matrix.
The precision matrix 𝚺−1 ≡ 𝚯 contains information about pairwise covariances
between the variables conditional on the rest, which is known as “partial covariances".
For instance, if 𝜃 𝑖 𝑗 , which is the 𝑖 𝑗-th element of the precision matrix, is zero, then
the variables 𝑖 and 𝑗 are conditionally independent, given the other variables. In order
to gain more insight about Gaussian graphical models, let us partition 𝑋 = (𝑍,𝑌 ),
where 𝑍 = (𝑋1 , . . . , 𝑋 𝑝−1 ) and 𝑌 = 𝑋 𝑝 . Given the length of the sample size 𝑇, let x𝑡 =
(𝑥1𝑡 , . . . , 𝑥 𝑝𝑡 ), X = (x1′ , . . . , x′𝑝 ), z𝑡 = (𝑥1𝑡 , . . . , 𝑥 ( 𝑝−1)𝑡 ), 𝑦 𝑡 = 𝑥 𝑝𝑡 , Z = (z1′ , . . . , z′𝑝−1 ),
and y = (𝑦 1 , . . . , 𝑦 𝑇 ) ′ denote realizations of the respective random variables, where
𝑡 = 1, . . . ,𝑇. Then with the partitioned covariance matrix given in (8.1), we can write
the conditional distribution of 𝑌 given 𝑍:
©𝚺 𝑍 𝑍 𝝈 𝑍𝑌 ª
𝚺= ®, (8.1)
𝝈 ′ 𝜎
« 𝑍𝑌 𝑌𝑌 ¬

𝑌 |𝑍 = 𝑧 ∼ N 𝜇𝑌 + (𝑧 − 𝝁 𝑍 ) ′ 𝚺−1
𝑍𝑍 𝝈 𝑍𝑌 , 𝜎𝑌𝑌 − 𝝈 ′
𝚺 −1
𝑍𝑌 𝑍 𝑍 𝝈 𝑍𝑌 . (8.2)
Note that the conditional mean in (8.2) is determined by the regression coefficient of
population multiple linear regression of 𝑌 on 𝑍, denoted as 𝜷𝑌 |𝑍 = (𝛽1 , . . . , 𝛽 𝑝−1 ) ′,
which can be expressed as the familiar OLS estimator introduced in Chapter 1
𝜷𝑌 |𝑍 = 𝚺−1 ′
𝑍 𝑍 𝝈 𝑍𝑌 with a Gram Matrix 𝚺 𝑍 𝑍 = Z Z and 𝝈 𝑍𝑌 = Z y.
′
If 𝛽 𝑗 = 0 for any 𝑗 = 1, . . . , 𝑝 − 1, then 𝑌 and 𝑍 𝑗 are conditionally independent,

given the rest. Therefore, the regression coefficients 𝜷𝑌 |𝑍 determine the conditional
(in)dependence structure of the graph. We reiterate that independence in this case
is a function of normality, when the distribution is no longer normal 𝛽 𝑗 = 0 is not
sufficient for independence. Let us partition 𝚯 in the same way as in (8.1):
©𝚯 𝑍 𝑍 𝜽 𝑍𝑌 ª
𝚯= ®.
′
𝜽
« 𝑍𝑌 𝜃 𝑌𝑌 ¬
Remark 8.1 Let M be an (𝑚 + 1) × (𝑚 + 1) matrix partitioned into a block form:
A
©|{z} b
|{z}ª®
M = 𝑚×𝑚 𝑚×1 ® .

®
b ′ 𝑐
« ¬
We can use the following standard formula for partitioned inverses (see Cullen, 1990;
Eves, 2012 among others):
−1
A − 1
bb ′ − 𝑘1 A−1 bª ©A−1 + 𝑘1 A−1 bb′A−1 − 𝑘1 A−1 bª
M−1 = 𝑐
®= (8.3)
©
®,
« −𝑘b A
1 ′ −1 1
𝑘 ¬ « − 𝑘1 b′A−1 1
𝑘 ¬
where 𝑘 = 𝑐 − b′A−1 b.
Now apply (8.3) and use 𝚺𝚯 = I, where I is an identity matrix, to get:
©𝚯 𝑍 𝑍 𝜽 𝑍𝑌 ª ©𝚯 𝑍 𝑍 −𝜃𝑌𝑌 𝚺−1𝑍 𝑍 𝝈 𝑍𝑌
®= (8.4)
ª
®,
′ ′ −1 −1
« 𝜽 𝑍𝑌 𝜃𝑌𝑌 ¬ « 𝜽 𝑍𝑌 (𝜎𝑌𝑌 − 𝝈 𝑍𝑌 𝚺 𝑍 𝑍 𝝈 𝑍𝑌 ) ¬
where 1/𝜃𝑌𝑌 = 𝜎𝑌𝑌 − 𝝈 ′𝑍𝑌 𝚺−1

𝑍 𝑍 𝝈 𝑍𝑌 . From (8.4):
𝜽 𝑍𝑌 = −𝜃𝑌𝑌 𝚺−1
𝑍 𝑍 𝝈 𝑍𝑌 = −𝜃𝑌𝑌 𝜷𝑌 |𝑍 ,
therefore,
−𝜽 𝑍𝑌
𝜷𝑌 |𝑍 = .
𝜃𝑌𝑌
Hence, zero elements in 𝜷𝑌 |𝑍 correspond to zeroes in 𝜽 𝑍𝑌 and mean that the
corresponding elements of 𝑍 are conditionally independent of 𝑌 , given the rest.
Therefore, 𝚯 contains all the conditional dependence information for the multivariate
Gaussian model.
Let W be the estimate of 𝚺. In practice, W can be any pilot estimator of
covariance matrix. Given a sample {x𝑡 }𝑇𝑡=1 , let S = (1/𝑇) 𝑇𝑡=1 (x𝑡 − x̄𝑡 ) (x𝑡 − x̄𝑡 ) ′
Í
denote the sample covariance matrix, which can be used as a choice for W. Also,
b2 ≡ diag(W). We can write down the Gaussian log-likelihood (up to constants)
let D
𝑙 (𝚯) = log det(𝚯) − trace(W𝚯). When W = S, the maximum likelihood estimator of
𝚯 is 𝚯b = S−1 .
In the high-dimensional settings it is necessary to regularize the precision matrix,
which means that some edges will be zero. In the following subsections we discuss
two most widely used techniques to estimate sparse high-dimensional precision
matrices.
8.2.1 Graphical LASSO
The first approach to induce sparsity in the estimation of precision matrix is to add
penalty to the maximum likelihood and use the connection between the precision
matrix and regression coefficients to maximize the following weighted penalized
log-likelihood (Janková & van de Geer, 2018):
∑︁
b 𝜆 = arg min trace(W𝚯) − log det(𝚯) + 𝜆
𝚯 𝑑b𝑖𝑖 𝑑b𝑗 𝑗 𝜃 𝑖 𝑗 , (8.5)
𝚯 𝑖≠ 𝑗
256 Seregina
over positive definite symmetric matrices, where 𝜆 ≥ 0 is a penalty parameter and

b The subscript 𝜆 in 𝚯
𝑑b𝑖𝑖 , 𝑑b𝑗 𝑗 are the 𝑖-th and 𝑗-th diagonal entries of D. b 𝜆 means that
the solution of the optimization problem in (8.5) will depend upon the choice of the
tuning parameter. More details on the latter are provided in Janková and van de Geer
(2018); Lee and Seregina (2021b) that describe how to choose the shrinkage intensity
in practice. In order to simplify notation, we will omit the subscript.
The objective function in (8.5) extends the family of linear shrinkage estimators
of the first moment studied in Chapter 1 to linear shrinkage estimators of the inverse
of the second moments. Instead of restricting the number of regressors for estimating
conditional mean, equation (8.5) restricts the number of edges in a graph by shrinking
some off-diagonal entries of precision matrix to zero. We draw readers’ attention to
the following: first, shrinkage occurs adaptively with respect to partial covariances
normalized by individual variances of both variables; second, only off-diagonal
partial correlations are shrunk to zero since the goal is to identify variables with
strongest pairwise conditional dependencies.
One of the most popular and fast algorithms to solve the optimization problem in
(8.5) is called the Graphical LASSO (GLASSO), which was introduced by Friedman
et al. (2007). Define the following partitions of W, S and 𝚯:
© W11 w12 S11 s12 © 𝚯11 𝜽 12

|{z} |{z} ª® © |{z} |{z} ®ª |{z} |{z} ª®
W = ( 𝑝−1)×( 𝑝−1) ( 𝑝−1)×1® , S = ( 𝑝−1)×( 𝑝−1) ( 𝑝−1)×1® , 𝚯 = ( 𝑝−1)×( 𝑝−1) ( 𝑝−1)×1® .

® ® ®
′ ′ ′
« w12 𝑤 22 ¬ « s12 𝑠22 ¬ « 𝜽 12 𝜃 22 ¬
Let 𝜷 ≡ −𝜽 12 /𝜃 22 . The idea of GLASSO is to set W = S + 𝜆I in (8.5) and combine

the gradient of (8.5) with the formula for partitioned inverses to obtain the following
ℓ1 -regularized quadratic program
n1 o
𝜷 = arg min 𝜷 ′W11 𝜷 − 𝜷 ′s12 + 𝜆∥ 𝜷∥ 1 ,
b (8.6)
𝜷 ∈R 𝑝−1 2
As shown by Friedman et al. (2007), (8.6) can be viewed as a LASSO regression,

where the LASSO estimates are functions of the inner products of W11 and 𝑠12 .
Hence, (8.5) is equivalent to 𝑝 coupled LASSO problems. Once we obtain b 𝜷, we can
estimate the entries 𝚯 using the formula for partitioned inverses.
The LASSO penalty in (8.5) can produce a sparse estimate of the precision
matrix. However, as was pointed out in Chapter 1 when discussing linear shrinkage
estimators of the first moment, it produces substantial biases in the estimates of
nonzero components. An approach to de-bias regularized estimators was proposed in
Janková and van de Geer (2018). Let 𝚯 b ≡ S−1 be the maximum likelihood estimator
−1
of the precision matrix, 𝚯0 ≡ 𝚺0 is the true value which is assumed to exist. Janková
and van de Geer (2018) show that asymptotic linearity of 𝚯 b follows from the
decomposition:

b − 𝚯0 = 𝚯0 (S − 𝚺0 ) 𝚯
𝚯 b = −𝚯0 (S − 𝚺0 )𝚯0 + − 𝚯0 (S𝚺0 )( 𝚯
b − 𝚯0 ) . (8.7)
| {z }
rem0
√
satisfies ∥rem0 ∥ ∞ = 𝑜(1/ 𝑇)1, where we use
Note that the remainder term in (8.7)
the notation ∥A∥ ∞ = max1≤𝑖, 𝑗 ≤ 𝑝 𝑎 𝑖 𝑗 for the supremum norm of a matrix A.
Now consider the standard formulation of the Graphical LASSO in (8.5). Its’
gradient is given by:
b −1 − S − 𝜆 · b
𝚯 𝚪 = 0, (8.8)
𝚪 is a matrix of component-wise signs of 𝚯:
where b b
𝛾 𝑗 𝑘 = 0 if 𝑗 = 𝑘,
𝛾 𝑗 𝑘 = sign( 𝜃ˆ 𝑗 𝑘 ) if 𝜃ˆ 𝑗 𝑘 ≠ 0,
𝛾 𝑗 𝑘 =∈ [−1, 1] if 𝜃ˆ 𝑗 𝑘 = 0.
Post-multiply (8.8) by 𝚯:
b
I − S𝚯
b − 𝜆b
𝚪𝚯b = 0,
therefore,
b = I − 𝜆b
S𝚯 𝚪𝚯.
b
Consider the following decomposition:
𝚯 b ′𝜂( 𝚯)
b +𝚯 b − 𝚯0 = −𝚯0 (S − 𝚺0 )𝚯0 + rem0 + rem1 ,
where 𝜂( 𝚯)b = 𝜆b 𝚪𝚯
b is the bias term, rem0 is the same as in (8.7), and rem1 =
b − 𝚯0 ) ′𝜂( 𝚯).
(𝚯 b
Proof
− 𝚯0 (S − 𝚺0 ) 𝚯 b − 𝚯0 ) ′𝜂( 𝚯)
b + (𝚯 b
= 𝚯0 (I − 𝜆b b + 𝚯0 𝚺0 𝚯
𝚪𝚯) b ′𝜂( 𝚯)
b +𝚯 b − 𝚯′ 𝜂( 𝚯)
b
0
′
(𝚯0 is symmetric) b +𝚯
=𝚯 b − 𝚯0
b 𝜂( 𝚯)
Provided that the remainder terms rem0 and rem1 are small enough, the de-biased
estimator can be defined as
b≡𝚯
T b ′ + 𝜂( 𝚯)
b +𝚯 b′ − 𝚯
b +𝚯
b =𝚯 b ′ S𝚯 b ′S𝚯.
b −𝚯
b = 2𝚯 b (8.9)
As pointed out by Janková and van de Geer (2018), in order to control the remainder
terms, we need bounds for the ℓ1 -error of 𝚯
b and to control the bias term, it is sufficient
1 See (Janková & van de Geer, 2018) for the proof.

258 Seregina

to control the upper bound 𝜂( 𝚯) = 𝜆b𝚪𝚯 ≤ 𝜆𝚯 , where given a matrix U
b b b
Í 𝑁 ∞ ∞ 1 Í
we used |||U||| ∞ ≡ max1≤ 𝑗 ≤𝑁 𝑖=1𝑢 𝑗,𝑖 and |||U||| 1 ≡ max1≤ 𝑗 ≤ 𝑁 𝑁 𝑢 𝑖, 𝑗 .
𝑖=1
8.2.2 Nodewise Regression
An alternative approach to induce sparsity in the estimation of precision matrix in

equation (8.5) is to solve for 𝚯 b one column at a time via linear regressions, replacing
population moments by their sample counterparts S. When we repeat this procedure
for each variable 𝑗 = 1, . . . , 𝑝, we will estimate the elements of 𝚯b column by column
using {x𝑡 }𝑇𝑡=1 via 𝑝 linear regressions. Meinshausen and Bühlmann (2006) use this
approach to incorporate sparsity into the estimation of the precision matrix. Instead
of running 𝑝 coupled LASSO problems as in GLASSO, they fit 𝑝 separate LASSO
regressions using each variable (node) as the response and the others as predictors to
estimate 𝚯.
b This method is known as the "nodewise“ regression and it is reviewed
below based on van de Geer, Buhlmann, Ritov and Dezeure (2014) and Callot et al.
(2019).
Let x 𝑗 be a 𝑇 × 1 vector of observations for the 𝑗-th regressor, the remaining
covariates are collected in a 𝑇 × 𝑝 matrix X− 𝑗 . For each 𝑗 = 1, . . . , 𝑝 we run the
following LASSO regressions:
2
𝜸 𝑗 = arg min x 𝑗 − X− 𝑗 𝜸 2 /𝑇 + 2𝜆 𝑗 ∥𝜸∥ 1 ,
b (8.10)
𝜸 ∈R 𝑝−1
where b𝜸 𝑗 = {b
𝛾 𝑗,𝑘 ; 𝑗 = 1, . . . , 𝑝, 𝑘 ≠ 𝑗 } is a ( 𝑝 − 1) × 1 vector of the estimated regression
coefficients that will be used to construct the estimate of the precision matrix, 𝚯. b
Define
© 1 −b 𝛾1,2 · · · −b
𝛾1, 𝑝 ª
®
𝛾2,1 1 · · · −b
−b 𝛾2, 𝑝 ®®
b=
C

.. . .. . ®.
. .. . .. ®®
®
«−b𝛾 𝑝,1 −b 𝛾 𝑝,2 · · · 1 ¬
For 𝑗 = 1, . . . , 𝑝, define the optimal value function
2
𝜏ˆ 2𝑗 = x 𝑗 − X− 𝑗 b

𝜸 𝑗 2 /𝑇 + 2𝜆 𝑗 b
𝜸 𝑗 1
and write
b2 = diag( 𝜏ˆ 2 , . . . , 𝜏ˆ𝑝2 ).
T 1
The approximate inverse is defined as
𝚯 b−2 C.
b𝜆 𝑗 = T b
Similarly to GLASSO, the subscript 𝜆 𝑗 in 𝚯 b 𝜆 𝑗 means that the estimated 𝚯 will

depend upon the choice of the tuning parameter: more details are provided in Callot
et al. (2019) which discusses how to choose shrinkage intensity in practice. The
subscript is omitted to simplify the notation.
8.2.3 CLIME
A different approach to recover the entries of precision matrix was motivated by the
compressed sensing and high-dimensional linear regression literature: instead of using
ℓ1 -MLE estimators, it proceeds by using a method of constrained ℓ1 minimization for
inverse covariance matrix estimation (CLIME, T. Cai et al., 2011). To illustrate the
motivation of CLIME and its connection with GLASSO, recall from (8.5) that when
W = S, the solution 𝚯
b GL satisfies:
𝚯b −1
GL − S = 𝜆 Z,
b (8.11)

where Zb is an element of the subdifferential 𝜕 Í𝑖≠ 𝑗 𝑑b𝑖𝑖 𝑑b𝑗 𝑗 𝜃 𝑖 𝑗 , which is a partial
derivative with respect to each off-diagonal entry of precision matrix. This leads
T. Cai et al. (2011) to consider the following optimization problem:
min∥𝚯∥ 1 s.t. ∥𝚯 − S∥ ∞ ≤ 𝜆, 𝚯 ∈ R 𝑝× 𝑝 , (8.12)
Notice that for a suitable choice of the tuning parameter, first-order conditions of
(8.12) coincide with the (8.11), meaning that both approaches would theoretically
lead to the same optimal solution for the precision matrix. However, the feasible set
in (8.12) is complicated, hence, the following relaxation is proposed:
min∥𝚯∥ 1 s.t. ∥S𝚯 − I∥ ∞ ≤ 𝜆. (8.13)
To make the solution 𝚯b in (8.13) symmetric, an additional symmetrization procedure

is performed which selects the values with smaller magnitude from lower and upper
parts of 𝚯.
b
Following T. Cai et al. (2011), Figure 8.3 illustrates the solution for recovering a 2
©𝑥 𝑧 ª
by 2 precision matrix ®, and only the plane 𝑥(= 𝑦) is plotted versus 𝑧 for simplicity.
«𝑧 𝑦¬
The CLIME solution 𝛀 b is located at the tangency of the feasible set (shaded polygon)
and objective function as in (8.13) (dashed diamond).The log-likelihood function of
GLASSO (as in (8.5)) is represented by the dotted line.
260 Seregina
Fig. 8.3: Objective functions of CLIME (dashed diamond) and GLASSO (the dotted
line) with the constrained feasible set (shaded polygon)
8.2.4 Solution Techniques
We now comment on the procedures used to obtain solutions for equations (8.6),
(8.10), and (8.13). T. Cai et al. (2011) use a linear relaxation followed by the
primal dual interior method approach. For GLASSO and nodewise regression, the
classical solution technique proceeds by applying stochastic gradient descent to
(8.6) and (8.10). Another useful technique to solve for precision matrix from (8.6)
and (8.10) is through the alternating direction method of multipliers (ADMM)
– a decomposition-coordination procedure in which the solutions to small local
subproblems are coordinated to find a solution to a large global problem. A detailed
coverage of a general ADMM method is provided in Boyd, Parikh, Chu, Peleato and
Eckstein (2011), we limit the discussion below to a GLASSO-specific procedure.
We now illustrate the use of ADMM to recover the entries of precision matrix
using GLASSO. Let us rewrite our objective function in (8.5):
b = arg min trace(W𝚯) − log det(𝚯) + 𝜆∥𝚯∥ 1 ,

𝚯 (8.14)
Θ≻0
over nonnegative definite matrices (denoted as 𝚯 ≻ 0). We now reformulate the

unconstrained problem in (8.14) as a constrained problem which can be solved using
ADMM:
min trace(W𝚯) − log det(𝚯) + 𝜆∥𝚯∥ 1

𝚯≻0
s.t. 𝚯 = Z,
where Z is an auxiliary variable designed for optimization purposes to track deviations

of the estimated precision matrix from the constraint. Now we can use scaled ADMM
to write down the augmented Lagrangian:
𝜌 𝜌
L𝜌 (𝚯, Z, U) = trace(S𝚯) − log det(𝚯) + 𝜆∥Z∥ 1 + ∥𝚯 − Z + U∥ 2𝐹 − ∥U∥ 2𝐹 .
2 2
The iterative updates are:
𝜌
𝚯 𝑘+1 ≡ arg min trace(W𝚯) − log det(𝚯) + ∥𝚯 − Z 𝑘 + U 𝑘 ∥ 2𝐹 , (8.15)
𝚯 2
𝜌 𝑘+1
𝑘+1 𝑘 2
Z ≡ arg min 𝜆∥Z∥ 1 + ∥𝚯 − Z + U ∥ 𝐹 , (8.16)
Z 2
U 𝑘+1 ≡ U 𝑘 + 𝚯 𝑘+1 − Z 𝑘+1 ,
where ∥·∥ 2𝐹 denotes the Frobenius norm which is calculated as the square root of the
sum of the squares of the entries. The updating rule in (8.16) is easily recognized to
be the element-wise soft thresholding operator:
Z 𝑘+1 ≡ 𝑆𝜆/𝜌 (𝚯 𝑘+1 + U 𝑘 ),
where the soft-thresholding operator is defined as:



 𝑎 − 𝑘, for 𝑎 > 𝑘


𝑆 𝑘 (𝑎) = 0, for |𝑎| ≤ 𝑘

 𝑎 + 𝑘,

for 𝑎 < −𝑘

Take the gradient of the updating rule in (8.15) in order to get a closed-form
solution to this updating rule:

W − 𝚯−1 + 𝜌 𝚯 − Z 𝑘 + U 𝑘 = 0,
Rearranging,

𝜌𝚯 − 𝚯−1 = 𝜌 Z 𝑘 − U 𝑘 − W. (8.17)
262 Seregina

Equation (8.17) implies that 𝚯 and 𝜌 Z 𝑘 − U 𝑘 − W share the same ei-

genvectors.2 Let Q𝚲Q′ be the eigendecomposition of 𝜌 Z 𝑘 − U 𝑘 − W, where
𝚲 = diag(𝜆1 , . . . , 𝜆 𝑁 ), and Q′Q = QQ′ = I. Pre-multiply (8.17) by Q′ and post-
multiply it by Q:
𝜌𝚯 e −1 = 𝚲.
e −𝚯 (8.18)
Now construct a diagonal solution of (8.18):
1
𝜌 𝜃˜ 𝑗 − = 𝜆 𝑗,
𝜃˜ 𝑗
e Solving for 𝜃˜ 𝑗 we get:

where 𝜃˜ 𝑗 denotes the 𝑗-th eigenvalue of 𝚯.
√︃
𝜆 𝑗 + 𝜆2𝑗 + 4𝜌
𝜃˜ = .
2𝜌
Now we can calculate 𝚯 which satisfies the optimality condition in (8.18):
1
√︃
𝚯= Q 𝚲 + 𝚲2 + 4𝜌I Q′ .
2𝜌
Note that the computational cost of the update in (8.15) is determined by the
eigenvalue decomposition of 𝑝 × 𝑝 matrix, which is O ( 𝑝 3 ).
As pointed out by Danaher, Wang and Witten (2014), suppose we determine that
the estimated precision matrix 𝚯 b is block diagonal with {𝑏 𝑙 } 𝐵 blocks, where each
𝑙=1
block contains 𝑝 𝑙 features. Then instead of computing the eigendecomposition of
a 𝑝 × 𝑝 matrix, we only need to compute the eigendecomposition of matrices of
dimension 𝑝 1 × 𝑝 1 , . . . , 𝑝 𝐵 × 𝑝 𝐵 . As a result, the computational complexity decreases
O ( 𝑝 𝑙3 ).
Í𝐵
to 𝑙=1
8.3 Graphical Models in the Context of Finance
This section addresses the questions of why and how graphical models are useful for
finance problems. In particular, we review the use of graphical models for portfolio
allocation and asset pricing.

2 𝚯q𝑖 = 1 𝑘 − U 𝑘 − W + 𝚯−1 q = 𝜃 q , where q is the eigenvector of 𝚯 corresponding to
𝜌 Z𝑖𝑖
𝜌 𝑖𝑖 𝑖𝑖 𝑖 𝑖 𝑖 𝑖

its eigenvalue 𝜃𝑖 . Post-multiply both parts of (8.17) by q𝑖 and rearrange: 𝜌 Z 𝑘 − U 𝑘 − W q𝑖 =

𝜌𝚯q𝑖 − 𝚯−1 q𝑖 = 𝜌 𝜃𝑖 q𝑖 − 𝜃1𝑖 q𝑖 = 𝜌 𝜃𝑖 − 𝜃1𝑖 q𝑖 . See Witten and Tibshirani (2009) for more general
cases.
We start by reviewing the basic problem faced by investors that allocate their
savings by investing in financial markets and forming a portfolio of financial assets.
Hence, they need to choose which stocks to include in a financial portfolio and
how much to invest in these stocks. Suppose we observe 𝑖 = 1, . . . , 𝑁 assets over
𝑡 = 1, . . . ,𝑇 period of time. Let r𝑡 = (𝑟 1𝑡 , 𝑟 2𝑡 , . . . , 𝑟 𝑁 𝑡 ) ′∼D (m, 𝚺) be an 𝑁 × 1 return
vector drawn from a distribution D which can belong to the either sub-Gaussian or
elliptical families.
The investment strategy is reflected by the choice of portfolio weights w =
(𝑤 1 , . . . , 𝑤 𝑁 ) which indicates how much is invested in each asset. Given w the
expected return of an investment portfolio is m′w, and the risk associated with the
investment strategy w is w′ 𝚺w. When a weight is positive, it is said that an investor
has a long position in an asset (i.e. they bought an asset), whereas negative weights
correspond to short positions (they are expected to deliver an asset). When the sum
of portfolio weights is equal to one, it means that an investor allocates all available
budget (normalized to one) for portfolio positions; when the sum is greater than one,
it means that an investors borrows additional amount on top of the initial budget;
when the sum is less than one, it means that an investor keeps some money as cash.
The choice of the investment strategy depends on several parameters and constraints,
including the minimum level of target return an investor desires to achieve (which we
denote as 𝜇), the maximum level of risk an investor is willing to tolerate (denoted as
𝜎), whether an investor is determined to allocate all available budget for portfolio
positions (in this case the portfolio weights are required to sum up to one), whether
short-selling is allowed (i.e. whether weights are allowed to be negative). The
aforementioned inputs give rise to three portfolio formulations that are discussed
below.
The first two formulations originate from Markowitz mean-variance portfolio
theory (Markowitz, 1952) that formulates the search for optimal portfolio weights
as a trade-off of achieving the maximum desired portfolio return while minimizing
the risk. Naturally, riskier strategies are associated with higher expected returns. The
statistics aimed √at capturing such trade-off is called Sharpe Ratio (SR), which is
defined as m′w/ w′ 𝚺w. The aforementioned goal can be formulated as the following
quadratic optimization problem:
1
min w′ 𝚺w



 w 2


s.t. w′ 𝜾 = 1 (8.19)


 m′w ≥ 𝜇,


where w is an 𝑁 × 1 vector of assets weights in the portfolio, 𝜾 is an 𝑁 × 1 vector
of ones, and 𝜇 is a desired expected rate of portfolio return. The first constraint in
(8.19) requires investors to have all available budget, normalized to one, invested in a
portfolio. This assumption can be easily relaxed and we demonstrate the implications
of this constraint on portfolio weights.
Equation (8.19) gives rise to two portfolio formulations. First, when the second
constraint is not binding, then the solution to (8.19) yields the global minimum-
264 Seregina
variance portfolio (GMV) weights w𝐺 :
w𝐺 = (𝜾′𝚯𝜾) −1 𝚯𝜾. (8.20)
Note that portfolio strategy in (8.19) does not depend on the target return or risk
tolerance, in this sense the solution is a global minimizer of portfolio risk for all
levels of portfolio return. Despite its simplicity, the importance of the minimum-
variance portfolio formation strategies as a risk-management tool has been studied
by many researchers. In Exhibit 1, Clarke, de Silva and Thorley (2011) provide
empirical evidence of superior performance of the GMV portfolios compared to
market portfolios for 1,000 largest U.S. stocks over 1968-2009. The most notable
difference occurs at recessions: during the financial crisis of 2007-09 minimum-
variance portfolio outperformed the market by 15-20% on average.
Below we provide a detailed proof showing how to derive an expression for a
portfolio strategy w𝐺 in (8.20). Let 𝜆1 and 𝜆2 denote Lagrange multipliers for the
first and the second constraints in (8.19) respectively.
Proof If m′w > 𝜇, then 𝜇 = (𝜾′Θ𝜾) −1 𝜾′Θm, 𝜆1 = (𝜾′Θ𝜾) −1 , 𝜆2 = 0. If 𝜆2 = 0, the
first-order condition of (8.19) yields:
1
w = − 𝜆1 𝚯𝜾, (8.21)
2
Pre-multiply both sides of (8.21) by 𝜾′ and express 𝜆1 :
1
𝜆1 = −2 , (8.22)
𝜾′𝚯𝜾
Plug-in (8.22) into (8.21):
w𝐺 = (𝜾′𝚯𝜾) −1 𝚯𝜾. (8.23)
The second portfolio formulation arises when the second constraint in (8.19)
is binding (i.e. m′w = 𝜇), meaning that the resulting investment strategy w 𝑀𝑊𝐶
achieves minimum risk for a certain level of target return 𝜇. We refer to this portfolio
strategy as Markowitz Weight-Constrained (MWC) portfolio:
w 𝑀𝑊𝐶 = (1 − 𝑎 1 )w𝐺 𝑀𝑉 + 𝑎 1 w∗𝑀 , (8.24)

w∗𝑀 ′ −1
= (𝜾 𝚯m) 𝚯m, (8.25)
𝜇(m′𝚯𝜾)(𝜾′𝚯𝜾) − (m′𝚯𝜾) 2
𝑎1 = .
(m′𝚯m)(𝜾′𝚯𝜾) − (m′𝚯𝜾) 2
This result is a well-known two-fund separation theorem introduced by Tobin (1958):
MWC strategy can be viewed as holding a GMV portfolio and a proxy for the market
fund w∗𝑀 since the latter capture all mean-related market information. In terms of
the parameters, this strategy requires an additional input asking investors to specify
their target return level. Below we provide a detailed proof showing how to derive an
expression for a portfolio strategy w 𝑀𝑊𝐶 in (8.24).
Proof Suppose the second constraint in (8.19) is binding, i.e. m′w = 𝜇. From the
first-order condition of (8.19) with respect to w we get:
w = 𝚯(𝜆1 𝜾 + 𝜆2 m) = 𝜆 1 𝚯𝜾 + 𝜆2 𝚯m (8.26)
= 𝜆1 (𝜾′𝚯𝜾)w𝐺 + 𝜆2 (𝜾′𝚯m)w 𝑀 .
Rewrite the first constraint in (8.19):
1 = w′𝚯𝚺𝜾 = 𝜆1 𝜾′𝚯𝜾 + 𝜆2 𝜾′𝚯m,
therefore, set
𝜆2 𝜾′𝚯m = 𝑎 1 , and w′𝚯𝚺𝜾 = 𝜆1 𝜾′𝚯𝜾 = 1 − 𝑎 1 .
To solve for 𝜆2 combine both constraints and (8.26):
1 = w′𝚯𝚺𝜾 = 𝜆1 𝜾′𝚯𝜾 + 𝜆 2 𝜾′𝚯m,

𝜇 = m′𝚯𝚺w = 𝜆1 m′𝚯𝜾 + 𝜆2 m′𝚯m,
therefore,
(m′𝚯m) − (m′𝚯𝜾)𝜇
𝜆1 = ,
(m′𝚯m)(𝜾′𝚯𝜾) − (m′𝚯𝜾) 2
(𝜾′𝚯𝜾)𝜇 − m′𝚯𝜾
𝜆2 = ,
(m 𝚯m)(𝜾′𝚯𝜾) − (m′𝚯𝜾) 2
′
𝜇(m′𝚯𝜾)(𝜾′𝚯𝜾) − (m′𝚯𝜾) 2
𝑎1 = .
(m′𝚯m)(𝜾′𝚯𝜾) − (m′𝚯𝜾) 2
It is possible to relax the constraint in (8.19) that requires portfolio weights to sum
up to one: this gives rise to the Markowitz Risk-Constrained (MRC) problem which
maximizes SR subject to either target risk or target return constraints, but portfolio
weights are not required to sum up to one:
m′w
max √ s.t. (i) m′w ≥ 𝜇 or(ii) w′ 𝚺w ≤ 𝜎 2 ,
w w′ 𝚺w
√
when 𝜇 = 𝜎 m′𝚯m, the solution to either of the constraints is given by
𝜎
w 𝑀 𝑅𝐶 = √ 𝚯m. (8.27)
m′𝚯m
Equation (8.27) tells us that once an investor specifies the desired return, 𝜇, and
maximum risk-tolerance level, 𝜎, this pins down the Sharpe Ratio of the portfolio.
The objective function in (8.19) only models risk preferences of investors. If we
want to incorporate investors’ desire to maximize expected return while minimizing
risk, we need to introduce the Markowitz mean-variance function, denoted as
𝑀 (m, 𝚺). According to Fan, Zhang and Yu (2012), if r𝑡 ∼N (m, 𝚺), and the utility
266 Seregina
function is given by 𝑈 (𝑥) = 1 − exp(−𝐴𝑥), where 𝐴 is the absolute risk-aversion

parameter, maximizing the expected utility would be equivalent to maximizing
𝑀 (m, 𝚺) = w′m − 𝛾w′ 𝚺w, where 𝛾 = 𝐴/2. Hence, we can formulate the following
optimization problem:
max w′m − 𝛾w′ 𝚺w
(
w
(8.28)
𝑠.𝑡. w′ 𝜾 = 1.
The solution to the above problem is analogous to the equation (8.24):
w = (1 − 𝑎 2 )w𝐺 + 𝑎 2 w 𝑀 ,
𝑎 2 = 𝛾(m′𝚯𝜾).
We can see from equations (8.23), (8.24), and (8.27) that in order to obtain optimal
portfolio weights, one needs to get an estimate of the precision matrix, 𝚯. As pointed
out by Zhan, Sun, Jakhar and Liu (2020), “from a graph viewpoint, estimating the
covariance using historic returns models a fully connected graph between all assets.
The fully connected graph appears to be a poor model in reality, and substantially
adds to the computational burden and instability of the problem". In the following
sections we will examine existing approaches to solving Markowitz mean-variance
portfolio problem. In addition, we will also propose desirable characteristics in the
estimated precision matrix that are attractive for a portfolio manager.
In practice, solution of (8.28) depends sensitively on the input vectors m and 𝚺,
and their accumulated estimation errors. Summarizing the work by Jagannathan and
Ma (2003), Fan et al. (2012) show that the sensitivity of utility function to estimation
errors is bounded by:
|𝑀 ( m b − 𝑀 (m, 𝚺)| ≤ ∥ m
b , 𝚺) b − 𝚺∥ ∞ ∥w∥ 2 ,
b − m∥ ∞ ∥w∥ 1 + 𝛾∥ 𝚺 (8.29)
1
b − m∥ ∞ and ∥ 𝚺
where ∥ m b − 𝚺∥ ∞ are the maximum componentwise estimation errors3.
The sensitivity problem can be alleviated by considering a modified version of (8.19),
known as the optimal no-short-sale portfolio:
1 ′
 min w 𝚺w




 w 2
′ (8.30)
 𝑠.𝑡. w 𝜾 = 1


 ∥w∥ ≤ 𝑐,
 1
where ∥w∥ 1 ≤ 𝑐 is the gross-exposure constraint for a moderate 𝑐. When 𝑐 = 1 - no

short sales are allowed4. We do not provide a closed-form solution for the
optimization problem in (8.30) since it needs to be solved numerically. Furthermore,
equation (8.24) no longer holds since the portfolio frontier can not be constructed
from a linear combination of any two optimal portfolios due to the restrictions on
weights imposed by the gross-exposure constraint. The constraint specifying the
3 The proof of (8.29) is a straightforward application of the Hölder’s inequality.

4 If w′𝜾 = 1 and 𝑤𝑖 ≥ 0, ∀𝑖, then ∥w ∥ 1 = |𝑤1 | + . . . |𝑤𝑁 | ≤ 1.
target portfolio return is also omitted since there may not exist a no-short-sale
portfolio that reaches this target.
Define a measure of risk 𝑅(w, 𝚺) = w′ 𝚺w, where risk is measured by the variance
of the portfolio. For risk minimization with the gross-exposure constraint we obtain:
b − 𝑅(w, 𝚺)| ≤ ∥ 𝚺
|𝑅(w, 𝚺) b − 𝚺∥ ∞ ∥w∥ 2 . (8.31)
1
The minimum of the right-hand side of (8.31) is achieved under no short-sale

constraint ∥w∥ 1 = 1. Fan et al. (2012) concluded that the optimal no-short-sale
portfolio in (8.30) has smaller actual risk than that for the global minimum-variance
portfolio described by weights in (8.20). However, their empirical studies showed
that the optimal no-short-sale portfolio is not diversified enough and the performance
can be improved by allowing some short positions.
8.3.1 The No-Short-Sale Constraint and Shrinkage
Investment strategies discussed so far did not put any sign-restrictions on portfolio
weights. Jagannathan and Ma (2003) explored the connection between regularizing
the sample covariance matrix and restricting portfolio weights: they showed that the
solution to the short sale constrained problem using the sample covariance matrix
𝑇
1 ∑︁
S= (r𝑡 − r̄𝑡 ) (r𝑡 − r̄𝑡 ) ′ (depicted in the left column of (8.32)) coincides with the
𝑇 𝑡=1
solution to the unconstrained problem (in the right column of (8.32)) if S is replaced
by 𝚺
b𝐽 𝑀 :
 min
 w′Sw  min

 w′ 𝚺
b𝐽 𝑀 w
 w

  w


 𝑠.𝑡. w′ 𝜾 = 1 =⇒  𝑠.𝑡. w′ 𝜾 = 1 (8.32)
 
 𝑤 ≥ 0 ∀𝑖  ′ ′
 b 𝐽 𝑀 = S − 𝝀𝜾 − 𝜾𝝀 ,
 𝑖 𝚺
where 𝝀 ∈ R 𝑁 is the vector of Lagrange multipliers for the short sale constraint.
(8.32) means that each of the no-short-sale constraints is equivalent to reducing the
estimated covariance of the corresponding asset with other assets by a certain amount.
Jagannathan and Ma (2003) interpret the estimator 𝚺b 𝐽 𝑀 as a shrinkage version of the
sample covariance matrix S and argue that it can reduce sampling error even when
no-short-sale constraints do not hold in the population. In order to understand the
relationship between covariance matrix and portfolio weights, consider the following
unconstrained GMV problem:
1
 min w′Sw



w 2 (8.33)
 𝑠.𝑡. w′ 𝜾 = 1.


The first-order condition of (8.33) is:
268 Seregina
𝑁
∑︁
𝑤 𝑖 𝑠 𝑗,𝑖 = 𝜆 ≥ 0, 𝑗 = 1, . . . , 𝑁 (8.34)
𝑖=1
where 𝜆 ∈ R is the Lagrange multiplier. Let us denote the investment strategy obtained
using a short-sale constraint as w𝑆 . Equation (8.34) means that at the optimum the
marginal contribution of stock 𝑗 to the portfolio variance is the same as the marginal
contribution of stock 𝑖 for any 𝑗, 𝑖. This is consistent with well-known asset pricing
models such as Capital Asset Pricing Model (CAPM) and Arbitrage Pricing Theory
(APT): the underlying principle behind them states that under the assumption that the
markets are efficient, total risk of a financial asset can be decomposed into common
and idiosyncratic parts. Common risk stems from similar drivers of volatility for all
assets, such as market movements and changes in macroeconomic indicators. Since
all assets are influenced by common drivers, the risk associated with this component
cannot be reduced by increasing the number of stocks in the portfolio. In contrast,
idiosyncratic (or asset-specific) risk differs among various assets depending on the
specificity of a firm, industry, country associated with an asset. It is possible to
decrease one’s exposure to idiosyncratic risk by increasing the number of stocks in
the portfolio (by doing so, investors reduce their exposure to a particular country or
industry. This is known as diversification – naturally, higher diversification benefits
can be achieved when portfolio includes assets with low or negative correlation.
Suppose stock 𝑗 has higher covariance with other stocks, meaning that the 𝑗-th
row of S has larger elements compared to other rows. This means that stock 𝑗 will
contribute more to the portfolio variance. Hence, to satisfy optimal condition in
(8.34) we need to reduce the weight of stock 𝑗 in the portfolio. If stock 𝑗 has high
variance and is highly correlated with other stocks, then its’ weight can be negative.
According to Green and Hollifield (1992), the presence of dominant factors leads to
extreme negative weights even in the absence of the estimation errors.
Remark 8.2 Consider the short-sale-constrained optimization problem in (8.32).

Suppose the non-negativity constraint for asset 𝑗 is binding. Then its’ covariances
with other assets will be reduced by 𝜆 𝑗 + 𝜆𝑖 for all 𝑖 ≠ 𝑗, and its’ variance is reduced
by 2𝜆 𝑗 . Jagannathan and Ma (2003) argue that since the largest covariance estimates
are more likely caused by upward-biased estimation error, the shrinking of covariance
matrix may reduce the estimation error. On the other hand, Green and Hollifield
(1992) suggest that the short-sale constraint will not, in general, hold in the population
when asset returns have dominant factors.
According to Jagannathan and Ma (2003), given a short-sale-constrained optimal

portfolio weight w𝑆 in (8.32), there exist many covariance matrix estimates that
have w𝑆 as their unconstrained GMV portfolio. Under the joint normality of returns,
Jagannathan and Ma (2003) show that 𝚺 b 𝐽 𝑀 is the constrained MLE of the population
𝑖.𝑖.𝑑.
covariance matrix. Let r𝑡 = (𝑟 1𝑡 , 𝑟 2𝑡 , . . . , 𝑟 𝑁 𝑡 ) ∼ N (m, 𝚺) be an 𝑁 × 1 return vector,
𝑇
1 ∑︁
S= (r𝑡 − r¯𝑡 ) (r𝑡 − r¯𝑡 ) ′ is the unconstrained MLE of 𝚺 and define 𝚺−1 ≡ 𝚯. Then
𝑇 𝑡=1
the log-likelihood as a function of covariance matrix becomes (up to the constants):
𝑙 (𝚯) = − log det 𝚺 − trace(S𝚯)

= log det 𝚯 − trace(S𝚯).
Jagannathan and Ma (2003) show that 𝚺 b 𝐽 𝑀 constructed from the solution to the
constrained GMV problem in (8.32) is the solution of the constrained ML problem
(8.35), where no short sales constraint in (8.32) is translated into regularization of
precision matrix in (8.35):

 max log det 𝚯 − trace(S𝚯)


 𝚯


∑︁ (8.35)
 𝑠.𝑡.
 𝜃 𝑖, 𝑗 ≥ 0,


 𝑗
where 𝜃 𝑖, 𝑗 is the (𝑖, 𝑗)-th element of the precision matrix. Even though the
regularization of weights in (8.35) is translated into the regularization of 𝚯,
Jagannathan and Ma (2003) solve the above problem for 𝚺.
b
The main findings of Jagannathan and Ma (2003) can be summarized as follows:
1. No-short-sale constraint shrinks the large elements of the sample covariance

matrix towards zero which has two effects. On the one hand, if an estimated large
covariance is due to the sampling error, the shrinkage reduces this error. However,
if the population covariance is large, the shrinkage introduces specification error.
The net effect is determined by the trade-off between sampling and specification
errors.
2. No-short-sale constraint deteriorates performance for factor models and shrinkage
covariance estimators (such as Ledoit and Wolf (2003), which is formed from a
combination of the sample covariance matrix and the 1-factor (market return)
covariance matrix);
3. Under no-short-sale constraint, minimum-variance portfolios constructed using
the sample covariance matrix perform comparable to factor models and shrinkage
estimators;
4. GMV outperform MWC portfolio which implies that the estimates of the mean
returns are very noisy
Let us now elaborate on the idea of incorporating no-short-sale constraint and,
consequently, shrinking the sample covariance estimator. We will examine the case
studied in Jagannathan and Ma (2003) when the asset returns are normally
distributed. Consider the first-order condition in (8.34). Suppose that stock 𝑗 has
high covariance with other stocks. As a result, the weight of this stock can be
negative and large in the absolute value. In order to reduce the impact of the 𝑗-th
stock on portfolio variance, no-short-sale approach will set the weight of such asset
to zero by shrinking the corresponding entries of the sample covariance matrix. The
main motivation of such shrinkage comes from the assumption that high covariance
is caused by the estimation error. However, as pointed out by Green and Hollifield
270 Seregina
(1992), extreme negative weights can be a result of the dominant factors rather than
the estimation errors. Hence, imposing no-short-sale assumption will fail to account
for important structural patterns in the data.
Furthermore, the approach discussed by Jagannathan and Ma (2003) only studies

the impact of the covariance structure. Assume we have 10 assets. Let asset 𝑗 = 1 be
highly correlated with all other assets. We can calculate partial correlation of asset
𝑗 = 1 and 𝑗 −1 = 2, . . . , 10. Suppose we find out that once we condition on 𝑗 = 2, the
partial correlation of stock 𝑗 with all assets except 𝑗 = 2 becomes zero. That would
mean that high covariance of the 𝑗-th stock with all other assets was caused by its
strong relationship with asset 𝑗 = 2. The standard approaches to the mean-variance
analysis will assume that this high covariance was a result of an estimation error,
and will reduce the weight of asset 𝑗 in the portfolio. We claim that instead of doing
that, we should exploit this relationship between assets 𝑗 = 1 and 𝑗 = 2. However,
the covariance (therefore, correlation) matrix will not be able to detect this structure.
Hence, we need another statistics, such as a matrix of partial correlations (precision
matrix) to help us draw such conclusions.
8.3.2 The 𝑨-Norm Constraint and Shrinkage
DeMiguel, Garlappi, Nogales and Uppal (2009) establish the relationship between
portfolio weights regularization and shrinkage estimator of covariance matrix pro-
posed by Ledoit and Wolf (2003, 2004a, 2004b). The latter developed a linear
shrinkage estimator denoted as 𝚺
b 𝐿𝑊 which is a combination of the sample covariance
matrix S and a low-variance target estimator 𝚺
btarget :
1 𝑣 b
𝚺
b 𝐿𝑊 = S+ 𝚺target ,
1+𝑣 1+𝑣
where 𝑣 ∈ R is a positive constant. Define ∥w∥ 𝐴 = (w′Aw) 1/2 to be an 𝐴-norm, where
A ∈ R 𝑁 ×𝑁 is a positive-definite matrix. DeMiguel et al. (2009) show that for each
𝑣 ≥ 0 there exists a 𝛿 such that the solution to the 𝐴-norm constrained GMV portfolio
problem coincides with the solution to the unconstrained problem if the sample
covariance matrix is replaced by 𝚺 b 𝐿𝑊 :
min w′Sw  min w′ 𝚺

 b 𝐿𝑊 w
  w

 w 


 
 ′
𝑠.𝑡. w′ 𝜾 = 1 =⇒  𝑠.𝑡. w 𝜾 = 1 (8.36)
1

 w′Aw ≤ 𝛿
 
 𝑣
𝚺
 b = S+ A.
 𝐿𝑊 1 + 𝑣

1+𝑣
If A is chosen to be the identity matrix I, then there is a one-to-one correspondence
between the 𝐴-norm-constrained portfolio on the left of (8.36) and the shrinkage
estimator proposed in Ledoit and Wolf (2004b). If A is chosen to be the 1-factor
(market return) covariance matrix 𝚺

b𝐹 , then there is a one-to-one correspondence with
the shrinkage portfolio in Ledoit and Wolf (2003). Therefore, direct regularization
of weights using 𝐴-norm constraint achieves shrinkage of the sample covariance
matrix. In order to understand this result, let us consider the 𝐴-norm-constrained
optimization problem in (8.36). Note that in contrast to (8.32), the 𝐴-norm shrinks
the total norm of the minimum-variance portfolio weights rather than shrinking every
weight. Now suppose the 𝐴-norm constraint in (8.36) binds. In this case 𝑣 > 0 and in
order to ensure the 𝐴-norm constraint is not violated, the sample covariance matrix
will be forced to shrink towards A.
Remark 8.3 When A = I, the 𝐴-norm becomes an ℓ2 -norm and the constraint becomes:
𝑁
∑︁ 1 2 1
𝑤𝑖 − ≤ 𝛿− . (8.37)
𝑖=1
𝑁 𝑁
Equation (8.37) follows from the footnote 10 of DeMiguel et al. (2009):

𝑁 𝑁 𝑁 𝑁 𝑁
∑︁ 1 2 ∑︁ 2 ∑︁ 1 ∑︁ 2𝑤 𝑖 ∑︁ 2 1
𝑤𝑖 − = 𝑤𝑖 + − 2 = 𝑤𝑖 − .
𝑖=1
𝑁 𝑖=1 𝑖=1
𝑁2 𝑖=1
𝑁 𝑖=1
𝑁
Therefore, using an ℓ2 -norm constraint imposes an upper-bound on the deviations of

the minimum-variance portfolio from the equally weighted portfolio. If 𝛿 = 1/𝑁 we
obtain 𝑤 𝑖 = 1/𝑁.
DeMiguel et al. (2009) show that empirically 𝐴-norm constrained portfolios

outperform the portfolio strategies in Jagannathan and Ma (2003); Ledoit and Wolf
(2003, 2004b), factor portfolios, and equally-weighted portfolio in terms of
out-of-sample Sharpe ratio. They study monthly returns for 5 datasets with the
number of assets less than 𝑁 = 50 for four datasets, and 𝑁 = 500 for the last dataset.
Once the number of assets is increased, direct regularization of weights is
computationally challenging. Moreover, one needs to justify the choice of the free
parameter 𝛿 in the 𝐴-norm constraint.
Let us now elaborate on the idea of the 𝐴-norm constraint and shrinkage
estimators, such as in Ledoit and Wolf (2003, 2004b). Note that the 𝐴-norm
constraint is equivalent to constraining the total exposure of the portfolio. Since the
constraint does not restrict individual weights, the resulting portfolio is not sparse.
Therefore, even when some assets have negligible weights, they are still included in
the final portfolio. However, as pointed out by Li (2015), when the number of assets
is large, sparse portfolio rule is desired. This is motivated from two main
perspectives: zero portfolio weights reduce transaction costs as well as portfolio
management costs; and, since the number of historical asset returns, 𝑇, might be
relatively small compared to the number of assets, 𝑁, this increases the estimation
error. Hence, we need some regularization scheme which would induce sparsity in
272 Seregina
the portfolio weights.
Furthermore, Rothman, Bickel, Levina and Zhu (2008) emphasized that shrinkage
estimators of the form proposed by Ledoit and Wolf (2003, 2004b) do not affect the
eigenvectors of the covariance, only the eigenvalues. However, Johnstone and Lu
(2009) showed that the sample eigenvectors are also not consistent in high-dimensions.
Moreover, recall that in the formulas for portfolio weights we need to use an estimator
of precision matrix, 𝚯. Once we shrink the eigenvalues of the sample covariance
matrix, it becomes invertible and we could use it for calculating portfolio weights.
However, shrinking the eigenvalues of the sample covariance matrix will not improve
the spectral behavior of the estimated precision. That is, it might have exploding
eigenvalues which might lead to extreme portfolio positions. In this sense, consistency
of precision in ℓ2 -operator norm implies that the eigenvalues of 𝚯 b consistently
estimate the corresponding eigenvalues of 𝚯. Therefore, we need a sparse estimator
of precision matrix that would be able to handle high-dimensional financial returns
and have a bounded spectrum.
8.3.3 Classical Graphical Models for Finance
Graphical models were shown to provide consistent estimates of the precision matrix
Friedman et al. (2007); Meinshausen and Bühlmann (2006); T. Cai et al. (2011).
Goto and Xu (2015) estimated a sparse precision matrix for portfolio hedging using
graphical models. They found out that their portfolio achieves significant out-of-
sample risk reduction and higher return, as compared to the portfolios based on
equal weights, shrunk covariance matrix, industry factor models, and no-short-sale
constraints. Awoye (2016) used Graphical LASSO to estimate a sparse covariance
matrix for the Markowitz mean-variance portfolio problem to improve covariance
estimation in terms of lower realized portfolio risk. Millington and Niranjan (2017)
conducted an empirical study that applies Graphical LASSO for the estimation of
covariance for the portfolio allocation. Their empirical findings suggest that portfolios
that use Graphical LASSO for covariance estimation enjoy lower risk and higher
returns compared to the empirical covariance matrix. They show that the results are
robust to missing observations: they remove a number of samples randomly from the
training data and compare how the corruption of the data affects the risks and returns
of the portfolios produced on both seen and unseen data. Millington and Niranjan
(2017) also construct a financial network using the estimated precision matrix to
explore the relationship between the companies and show how the constructed
network helps to make investment decisions. Callot et al. (2019) use the nodewise-
regression method of Meinshausen and Bühlmann (2006) to establish consistency of
the estimated variance, weights and risk of high-dimensional financial portfolio. Their
empirical application demonstrates that the precision matrix estimator based on the
nodewise-regression outperforms the principal orthogonal complement thresholding
estimator (POET) (Fan, Liao & Mincheva, 2013) and linear shrinkage (Ledoit &
Wolf, 2004b). T. T. Cai, Hu, Li and Zheng (2020) use constrained ℓ1 -minimization
for inverse matrix estimation (CLIME) of the precision matrix (T. Cai et al., 2011)
to develop a consistent estimator of the minimum variance for high-dimensional
global minimum-variance portfolio. It is important to note that all the aforementioned
methods impose some sparsity assumption on the precision matrix of excess returns.
Having originated from the literature on statistical modelling, graphical models
inherit the properties and assumptions common in that literature, such as sparse
environment and lack of dynamics. Natural questions are (1) whether these statistical
assumptions are justified in economics and finance settings, and (2) how to augment
graphical models and make them suitable for the use in economics and finance?
8.3.4 Augmented Graphical Models for Finance Applications
We start with analysing the sparsity assumption imposed in all graphical models: many
entries of precision matrix are zero, which is a necessary condition to consistently
estimate inverse covariance.
The arbitrage pricing theory (APT), developed by (Ross, 1976), postulates that the
expected returns on securities should be related to their covariance with the common
components or factors only. The goal of the APT is to model the tendency of asset
returns to move together via factor decomposition. Assume that the return generating
process (r𝑡 ) follows a 𝐾-factor model:
r𝑡 = B f𝑡 + 𝜺 𝑡 , 𝑡 = 1, . . . ,𝑇 (8.38)
|{z} |{z}
𝑝×1 𝐾×1
where f𝑡 = ( 𝑓1𝑡 , . . . , 𝑓𝐾𝑡 ) ′ are the factors, B is a 𝑝 × 𝐾 matrix of factor loadings, and
𝜺 𝑡 is the idiosyncratic component that cannot be explained by the common factors.
Without loss of generality, we assume throughout the paper that unconditional means
of factors and idiosyncratic component are zero. Factors in (8.38) can be either
observable, such as in (Fama & French, 1993, 2015), or can be estimated using
statistical factor models. Unobservable factors and loadings are usually estimated
by the principal component analysis (PCA), as studied in Connor and Korajczyk
(1988); Bai (2003); Bai and Ng (2002); Stock and Watson (2002). Strict factor
structure assumes that the idiosyncratic disturbances, 𝜺 𝑡 , are uncorrelated with each
other, whereas approximate factor structure allows correlation of the idiosyncratic
disturbances (see Chamberlain and Rothschild (1983); Bai (2003) among others).
When common factors are present across financial returns, the precision matrix
cannot be sparse because all pairs of the forecast errors are partially correlated given
other forecast errors through the common factors. To illustrate this point, we generated
variables that follow (8.38) with 𝐾 = 2 and 𝜀 𝑡 ∼ N (0, 𝚺 𝜀 ), where 𝜎𝜀,𝑖 𝑗 = 0.4 |𝑖− 𝑗 | is
the 𝑖, 𝑗-th element of 𝚺 𝜀 . The vector of factors f𝑡 is drawn from N (0, I𝐾 /10), and the
entries of the matrix of factor loadings for forecast error 𝑗 = 1, . . . , 𝑝, b 𝑗 , are drawn
from N (0, I𝐾 /100). The full loading matrix is given by B = (b1 , . . . , b 𝑝 ) ′. Let 𝐾 b
274 Seregina
denote the number of factors estimated by the PCA. We set (𝑇, 𝑝) = (200, 50) and
plot the heatmap and histogram of population partial correlations of financial returns
r𝑡 , which are the entries of a precision matrix, in Figure 8.4. We now examine the
performance of graphical models for estimating partial correlations under the factor
structure. Figure 8.5 shows the partial correlations estimated by GLASSO that does
not take into account factors: due to strict sparsity imposed by graphical models
almost all partial correlations are shrunk to zero which degenerates the histogram in
Figure 8.5. This means that strong sparsity assumption on 𝚯 imposed by classical
graphical models (such as GLASSO, nodewise regression, or CLIME discussed in
Section 8.2) is not realistic under the factor structure.
One attempt to integrate factor modeling and high-dimensional precision estimation
was made by (Fan, Liu & Wang, 2018) (Section 5.2): the authors referred to such
class of models as “conditional graphical models". However, this was not the main
focus of their paper which concentrated on covariance estimation through elliptical
factor models. As (Fan et al., 2018) pointed out, “though substantial amount of
efforts have been made to understand the graphical model, little has been done
for estimating conditional graphical model, which is more general and realistic".
One of the studies that examines theoretical and empirical performance of graphical
models integrated with the factor structure in the context of portfolio allocation is
(Lee & Seregina, 2021b). They develop a Factor Graphical LASSO Algorithm that
decomposes precision matrix of stock returns into low-rank and sparse components,
with the latter estimated using GLASSO. To have a better understanding of the
framework, let us introduce some notations. First, rewrite (8.38) in matrix form:
R = B F + E, (8.39)
|{z} |{z}
𝑝×𝑇 𝑝×𝐾
The factors and loadings in (8.39) are estimated by solving the following minimization
problem: ( B,
bb F) = arg minB,F ∥R − BF∥ 2𝐹 s.t. 𝑇1 FF′ = I𝐾 , B′B is diagonal. The
constraints are needed to identify the factors (Fan et al., 2018).
Given a symmetric positive semi-definite matrix U, let Λmax (U) ≡ Λ1 (U) ≥
Λ2 (U) ≥ . . . ≥ Λmin (U) ≡ Λ 𝑝 (U) be the eigenvalues of U, and eig𝐾 (U) ∈ R𝐾× 𝑝
denote the first 𝐾 ≤ 𝑝 normalized eigenvectors √ corresponding to Λ1 (U), . . . , Λ𝐾 (U).
It was shown (Stock & Watson, 2002) that b F = 𝑇eig𝐾 (R′R) and B b = 𝑇 −1 Rb F′.
Given b b define E
F, B, b = R−B bb −1 ′ −1 ′
F. Let 𝚺 𝜀 = 𝑇 EE and 𝚺 𝑓 = 𝑇 FF be covariance
matrices of the idiosyncratic components and factors, and let 𝚯 𝜀 = 𝚺−1 𝜀 and 𝚯 𝑓 = 𝚺 𝑓
−1
be their inverses. Given a sample of the estimated residuals {b 𝜺 𝑡 = r𝑡 − B

bfb𝑡 }𝑇 and
𝑡=1
Í
b 𝜀 = (1/𝑇) 𝑇 b ′ b 𝑓 = (1/𝑇) Í𝑇 b b′
the estimated factors {bf𝑡 }𝑇𝑡=1 , let 𝚺 𝑡=1 𝜺 𝑡 𝜺
b 𝑡 and 𝚺 𝑡=1 f𝑡 f𝑡 be
the sample counterparts of the covariance matrices.
To decompose precision of financial returns into low-rank and sparse components,
the authors apply Sherman-Morrison-Woodbury formula to estimate the final precision
matrix of excess returns:
𝚯 b𝜀 −𝚯
b=𝚯 b 𝚯
b 𝜀 B[ b′𝚯
b 𝑓 +B b −1 B
b 𝜀 B] b′𝚯
b𝜀. (8.40)
The estimated precision matrix from(8.40) is used to compute portfolio weights, risk
and the Sharpe Ratio and establish consistency of these performance metrics.
Let us now revisit the motivating example at the beginning of this section: Figures
8.6-8.8 plot the heatmaps and the estimated partial correlations when precision
matrix is computed using Factor GLASSO with 𝐾 b ∈ {1, 2, 3} statistical factors. The
heatmaps and histograms closely resemble population counterparts in Figure 8.4, and
the result is not very sensitive to over- or under-estimating the number of factors 𝐾. b
This demonstrates that using a combination of classical graphical models and factor
structure via Factor Graphical Models improves upon the performance of classical
graphical models.
We continue with the lack of dynamics in the graphical models literature. First,
recall that a precision matrix represents a network of interacting entities, such as
corporations or genes. When the data is Gaussian, the sparsity in the precision
matrix encodes the conditional independence graph - two variables are conditionally
independent given the rest if and only if the entry corresponding to these variables in
the precision matrix is equal to zero. Inferring the network is important for portfolio
allocation problem. At the same time, the financial network changes over time, that
is, the relationships between companies can change either smoothly, or abruptly (e.g.
as a response to an unexpected policy shock, or in the times of economic downturns).
Therefore, it is important to account for time-varying nature of stock returns. The
time-varying network also implies time-varying second moments of a distribution –
this naturally means that both covariance and precision matrices vary with time.
There are two streams of literature that study time-varying networks. The first one
models dynamics in the precision matrix locally. Zhou, Lafferty and Wasserman (2010)
develop a nonparametric method for estimating time-varying graphical structure for
multivariate Gaussian distributions using an ℓ1 -penalized log-likelihood. They find
out that if the covariances change smoothly over time, the covariance matrix can
be estimated well in terms of predictive risk even in high-dimensional problems.
Lu, Kolar and Liu (2015) introduce nonparanormal graphical models that allow to
model high-dimensional heavy-tailed systems and the evolution of their network
structure. They show that the estimator consistently estimates the latent inverse
Pearson correlation matrix. The second stream of literature allows the network to vary
with time by introducing two different frequencies. Hallac, Park, Boyd and Leskovec
(2017) study time-varying Graphical LASSO with smoothing evolutionary penalty.
One of the works that combines latent factor modeling, time-variation and graphical
modeling is a paper by Zhan et al. (2020). To capture the latent space distribution, the
authors use PCA and autoencoders. To model temporal dependencies they employ
variational autoencoders with Gaussian and Cauchy priors. A graphical model relying
on GLASSO is associated to each time interval and the graph is updated when moving
to the next time point.
276 Seregina
Fig. 8.4: Heatmap and histogram of population partial correlations; 𝑇 = 200, 𝑝 = 50,
𝐾 =2
Fig. 8.5: Heatmap and histogram of sample partial correlations estimated using
GLASSO with no factors; 𝑇 = 200, 𝑝 = 50, 𝑞 = 2, 𝑞ˆ = 0
Factor GLASSO with 1 statistical factor; 𝑇 = 200, 𝑝 = 50, 𝐾 = 2, 𝐾ˆ = 1
Factor GLASSO with 2 statistical factors; 𝑇 = 200, 𝑝 = 50, 𝐾 = 2, 𝐾ˆ = 2
278 Seregina
Factor GLASSO with 3 statistical factors; 𝑇 = 200, 𝑝 = 50, 𝐾 = 2, 𝐾ˆ = 3
8.4 Graphical Models in the Context of Economics
In this section we review other applications of graphical models to economic problems

and revisit further extensions. We begin with a forecast combination exercise that
bears a close resemblance with portfolio allocation strategies reviewed in the previous
section. Then we deviate from the optimization problem of searching for the optimal
weights and review other important applications which rely on the use of the estimator
of precision matrix. We devote special attention to Vector Autoregressive (VAR)
models that emphasize the relationship between network estimation and recovery of
a sparse inverse covariance.
8.4.1 Forecast Combinations
Not surprisingly, the area of economic forecasting bears a close resemblance to

portfolio allocation exercise reviewed in the previous section. Suppose we have
𝑝 competing forecasts, b y𝑡 = ( 𝑦ˆ 1,𝑡 , . . . , 𝑦ˆ 𝑝,𝑡 ) ′, of the variable 𝑦 𝑡 , 𝑡 = 1, . . . ,𝑇. Let
′
y𝑡 = (𝑦 𝑡 , . . . , 𝑦 𝑡 ) . Define e𝑡 = y𝑡 −b y𝑡 = (𝑒 1𝑡 , . . . , 𝑒 𝑝𝑡 ) ′ to be a 𝑝 × 1 vector of forecast
errors. The forecast combination is defined as follows:
𝑦 𝑡𝑐 = w′b
b y𝑡
where w is a 𝑝 × 1 vector of weights. Define the mean-squared forecast error (MSFE)

to be a measure of risk MSFE(w, 𝚺) = w′ 𝚺w. As shown in Bates and Granger (1969),
the optimal forecast combination minimizes the variance of the combined forecast
error:
′
min MSFE = min E[w′e𝑡 e𝑡 w] = min w′ 𝚺w, s.t. w′ 𝜾 𝑝 = 1, (8.41)
w w w
where 𝜾 𝑝 is a 𝑝 × 1 vector of ones. The solution to (8.41) yields a 𝑝 × 1 vector of the

optimal forecast combination weights:
𝚯𝜾 𝑝
w= . (8.42)
𝜾′𝑝 𝚯𝜾 𝑝
If the true precision matrix is known, the equation (8.42) guarantees to yield the
optimal forecast combination. In reality, one has to estimate 𝚯. Hence, the out-of-
sample performance of the combined forecast is affected by the estimation error.
As pointed out by Smith and Wallis (2009), when the estimation uncertainty of
the weights is taken into account, there is no guarantee that the “optimal" forecast
combination will be better than the equal weights or even improve the individual
forecasts. Define 𝑎 = 𝜾′𝑝 𝚯𝜾 𝑝 /𝑝, and b
𝑎 = 𝜾′𝑝 𝚯𝜾
b 𝑝 /𝑝. We can write

MSFE(b w, 𝚺)
b 𝑎ˆ −1
|𝑎 − 𝑎| ˆ
− 1 = −1 − 1 =

,
| 𝑎|

MSFE(w, 𝚺) 𝑎 ˆ
and

( 𝚯−𝚯)𝜾 𝑝
b
1 ∥ 𝚯𝜾 𝑝 ∥ 1
𝑎 𝑝 + |𝑎 − b
𝑎| 𝑝
wb − w 1 ≤ .
|b
𝑎 |𝑎
Therefore, in order to control the estimation uncertainty in the MSFE and combination
weights, one needs to obtain a consistent estimator of the precision matrix 𝚯.
Lee and Seregina (2021a) apply the idea of the Factor Graphical LASSO described
in Subsection 8.3.4 to forecast combination for macroeconomics time-series. They
argue that the success of equal-weighted forecast combinations is partly due to the fact
that the forecasters use the same set of public information to make forecasts, hence,
they tend to make common mistakes. For example, they illustrate that in the European
Central Bank’s Survey of Professional forecasters of Euro-area real GDP growth, the
forecasters tend to jointly understate or overstate GDP growth. Therefore, the authors
stipulate that the forecast errors include common and idiosyncratic components,
which allows the forecast errors to move together due to the common error component.
Their paper provides a framework to learn from analyzing forecast errors: the authors
separate unique errors from the common errors to improve the accuracy of the
combined forecast.
280 Seregina
8.4.2 Vector Autoregressive Models
The need for graphical modeling and learning the network structure among a large set
of time series arises in economic problems that adhere to VAR framework. Examples
of such applications include macroeconomic policy making and forecasting, and
assessing connectivity among financial firms. Lin and Michailidis (2017) provide a
good overview that draws the links between VAR models and networks, which we
summarise below. The authors start with an observation that in many applications the
components of a system can be partitioned into interacting blocks. As an example,
Cushman and Zha (1997) examined the impact of monetary policy in a small open
economy. The economy is modeled as one block, whereas variables in foreign
economies as the other. Both blocks have their own autoregressive structure, and there
is unidirectional interdependence between the blocks: the foreign block influences
the small open economy, but not the other way around. Hence, there exists a linear
ordering amongst blocks. Another example provided by Lin and Michailidis (2017)
stems from the connection between the stock market and employment macroeconomic
variables (Farmer, 2015 that focuses on the impact through a wealth effect mechanism
of the former on the latter. In this case the underlying hypothesis of interest is that the
stock market influences employment, but not the other way around. An extension of
the standard VAR modeling introduces an additional exogenous block of variables
“X" that exhibits autoregressive dynamics – such extension is referred to as VAR-X
model. For instance, Pesaran, Schuermann and Weiner (2004) build a model to
study regional inter-dependencies where country specific macroeconomic indicators
evolve according to a VAR model, and they are influenced by key macroeconomic
variables from neighbouring countries/regions (an exogenous block). Abeysinghe
(2001) studies the direct and indirect impact of oil prices on the GDP growth of 12
Southeast and East Asian economies, while controlling for such exogenous variables
as the country’s consumption and investment expenditures along with its trade
balance.
Let us now follow Lin and Michailidis (2017) to formulate the aforementioned
VAR setup as a recursive linear dynamical system comprising two blocks of variables:
x𝑡 = Ax𝑡−1 + u𝑡 , (8.43)
z𝑡 = Bx𝑡−1 + Cz𝑡−1 + v𝑡 , (8.44)
where x𝑡 ∈ R 𝑝1 and z𝑡 ∈ R 𝑝2 are the variables in groups 1 and 2, respectively. Matrices

A and C capture the temporal intra-block dependence, while matrix B captures the
inter-block dependence. Note that the block of x𝑡 variables acts as an exogenous
effect to the evolution of the z𝑡 block. Furthermore, z𝑡 is Granger-caused by X𝑡 . Noise
processes {u𝑡 } and {v𝑡 } capture additional contemporaneous intra-block dependence
of x𝑡 and z𝑡 . In addition, the noise processes are assumed to follow zero mean
Gaussian distributions:
u𝑡 ∼ N (0, 𝚺𝑢 ) v𝑡 ∼ N (0, 𝚺 𝑣 ),
where 𝚺𝑢 and 𝚺 𝑣 are covariance matrices. The parameters of interest are transition
matrices A ∈ R 𝑝1 × 𝑝1 , B ∈ R 𝑝2 × 𝑝1 , C ∈ R 𝑝2 × 𝑝2 , and the covariance matrices 𝚺𝑢 and 𝚺 𝑣 .
Lin and Michailidis (2017) assume that the matrices A, C, 𝚯𝑢 ≡ 𝚺−1 𝑢 , and 𝚯𝑣 ≡ 𝚺 𝑣
−1
are sparse, whereas B can be either sparse or low-rank.

We now provide an overview of the estimation procedure used to obtain the ML
estimates of the aforementioned transition matrices and precision matrices, based on
Lin and Michailidis (2017). First, let us introduce some notations: the “response"
matrices from time 1 to T are defined as:
X 𝑇 = [𝑥 1 𝑥2 . . . 𝑥𝑇 ] ′ Z 𝑇 = [𝑧1 𝑧 2 . . . 𝑧𝑇 ] ′,
where {𝑥 0 , . . . , 𝑥𝑇 } and {𝑧 0 , . . . , 𝑧𝑇 } is centered time series data. Further, define the

“design" matrices from time 0 to T-1 as:
X = [𝑥 0 𝑥1 . . . 𝑥𝑇−1 ] ′ Z 𝑇 = [𝑧 0 𝑧1 . . . 𝑧𝑇−1 ] ′ .
The error matrices are denoted as U and V. The authors proceed by formulating
optimization problems using penalized log-likelihood functions to recover A and 𝚯𝑢 :
h
( A,
b 𝚯b 𝑢 ) = arg min tr 𝚯𝑢 (X 𝑇 − XA′) ′ (X 𝑇 − XA′)/𝑇 (8.45)
A,𝚯𝑢
i
− log|𝚯𝑢 | + 𝜆 𝐴 ∥A∥ 1 + 𝜌𝑢 ∥𝚯𝑢 ∥ 1,off ,
as well as B, C, and 𝚯𝑣 :
h
( B, b 𝑣 ) = arg min tr 𝚯𝑣 (Z 𝑇 − XB′ − ZC′) ′ (Z 𝑇 − XB′ − ZC′)/𝑇
b 𝚯
b C, (8.46)
B,C,𝚯𝑣
i
− log|𝚯𝑣 | + 𝜆 𝐵 R (B) + 𝜆𝐶 ∥C∥ 1 + 𝜌 𝑣 ∥𝚯𝑣 ∥ 1,off , ,
Í
where ∥𝚯𝑢 ∥ 1,off = 𝑖≠ 𝑗 𝜃 𝑢,𝑖 𝑗 , and 𝜆 𝐴, 𝜆 𝐵 , 𝜆𝐶 , 𝜌𝑢 , 𝜌 𝑣 are tuning parameters con-
trolling the regularization strength. The regularizer R (B) = ∥B∥ 1 if B is assumed to
Í𝑚𝑖𝑛( 𝑝2 , 𝑝1 )
be sparse, and R (B) = |||B||| ∗ = 𝑖=1 (singular values𝑖 of B) if B is assumed
to be low-rank, where |||·||| ∗ is a nuclear norm. To solve (8.45) and (8.46), Lin and
Michailidis (2017) develop two algorithms that iterate between estimating transition
matrices by minimizing the regularized sum of squared residuals keeping the precision
matrix fixed, and estimating precision matrices using GLASSO keeping the transition
matrices fixed. The finite sample error bounds for the obtained estimates are also
established.
In their empirical application, the authors extend the model of Farmer (2015):
they analyze the temporal dynamics of the log-returns of stocks with large market
capitalization and key macroeconomic variables. In the context of the aforementioned
notations, the x𝑡 block consists of the stock log-returns, whereas the Z𝑡 block consists
of the macroeconomic variables. The authors’ findings are consistent with the previous
empirical results documenting increased connectivity during the crisis periods and
282 Seregina
strong impact of the stock market on total employment arguing that the stock market
provides a plausible explanation for the great recession.
Basu, Li and Michailidis (2019) suggest that the sparsity assumption on (8.43)
may not be sufficient: fir instance, returns on assets tend to move together in a more
concerted manner during financial crisis periods. The authors proceed by studying
high-dimensional VAR models where the transition matrix A exhibits a more complex
structure: it is low rank and/or (group) sparse, which can be formulated as the
following optimization problem:
x𝑡 = Ax𝑡−1 + u𝑡 , u𝑡 ∼ N (0, 𝚺𝑢 )
A = L∗ + R∗ , rank(L∗ ) = 𝑟,
where L∗ is the low-rank component and R∗ is either sparse (S∗ ), or group-sparse

(G∗ ). The authors assume that the number of non-zero elements in a sparse case
(as defined by the cardinality of S∗ shown as an ℓ0 -norm: ∥·∥ 0 ) is ∥S∗ ∥ 0 = 𝑠, and
∥G∗ ∥ 2,0 = 𝑔 which denotes the number of nonzero groups in G∗ in the group sparse
case. The goal is to estimate L∗ and R∗ accurately based on 𝑇 ≪ 𝑝 2 . To overcome an
inherent identifiability issue in the estimation of sparse and low-rank components,
Basu, Li and Michailidis (2019) impose a well-known incoherence condition which
is sufficient for exact recovery of L∗ and R∗ by solving the following convex program:
( L, b = arg min 𝑙 (L, R)

b R) (8.47)
L∈𝛀,R
1 2
𝑙 (𝐿, 𝑅) ≡ X 𝑇 − X(L + R) 𝐹 + 𝜆 𝑁 ∥L∥ ∗ + 𝜇 𝑁 ∥R∥ ♢ ,
2
where 𝛀 = {L ∈ R 𝑝× 𝑝 : ∥L∥ max ≤ 𝛼/𝑝} (for sparse) or 𝛀 = {L ∈ R 𝑝× 𝑝 : ∥L∥ 2,max ≡
√
max 𝑘=1,...,𝐾 (L)𝐺𝑘 𝐹 ≤ 𝛽/ 𝐾 } (for group sparse); ∥·∥ ♢ represents ∥·∥ 1 or ∥·∥ 2,1
depending on sparsity or group sparsity of R. The parameters 𝛼 and 𝛽 control the
degree of non-identifiability of the matrices allowed in the model class. To solve the
optimization problem in (8.47), Basu, Li and Michailidis (2019) develop an iterative
algorithm based on the gradient descent, which they call “Fast Network Structure
Learning".
In their empirical application, the authors employ the proposed framework to
learn Granger causal networks of asset pricing data obtained from CRSP and WRDS.
They examine the network structure of realized volatilities of financial institutions
representing banks, primary broker/dealers and insurance companies. Two main
findings can be summarized as follows: (1) they document increased connectivity
pattern during the crisis periods; (2) significant sparsity of the estimated sparse
component provides a further scope for better examining specific firms that are key
drivers in the volatility network.
Another example of the use of graphical models for economic problems is to
measure systemic risk. For instance, network connectivity of large financial institutions
can be used to identify systematically important institutions based on the centrality
of their role in a network. Basu, Das, Michailidis and Purnanandam (2019) propose a
system-wide network measure that uses GLASSO to estimate connectivity among
many firms using a small sample size. The authors criticise a pairwise approach of
learning network structures drawing the analogy of the latter with the omitted variable
bias in standard regression models which leads to inconsistently estimated model
parameters. To overcome the aforementioned limitation of the pairwise approach
and correctly identify the interconnectedness structure of the system, Basu, Das et al.
(2019) fit a sparse high-dimensional VAR model and develop a testing framework for
Granger causal effects obtained from regularized estimation of large VAR models.
They consider the following model of stock returns of 𝑝 firms:
x𝑡 = Ax𝑡−1 + 𝜺 𝑡 , 𝜺 𝑡 ∼ N (0, 𝚺 𝜀 ), 𝚺 𝜀 = diag(𝜎12 , . . . , 𝜎𝑝2 ), 𝜎 2𝑗 > 0 ∀ 𝑗,
which is a simplified version of (8.43). The paper assumes that the financial network
is sparse: in other words, they require the true number of interconnections between
the firms to be very small. To recover partial correlations between stock returns, Basu,
Das et al. (2019) use a debiased Graphical LASSO as in (8.9).
In their empirical exercise, the authors use monthly returns for three financial
sectors: banks, primary broker/dealers and insurance companies. They focus on three
systemically important events: the Russian default and LCTM bankruptcy in late 1998,
the dot-com bubble accompanied with the growth of mortgage-backed securities
in 2002, and the global financial crisis of 2007-09. The paper finds connectivity,
measured by either the count of neighbours or distance between nodes, increases
before and during systemically important events. Furthermore, the approach developed
in the paper allows tracing the effect of a negative shock on a firm on the entire
network by tracing its effects through the direct linkages. Finally, using an extensive
simulation exercise, the authors show that debiased GLASSO outperforms competing
methods in terms of the estimation and detection accuracy.
Finally, non-stationarity is another distinctive feature of economic time series.
Basu and Rao (2022) take a further step in extending graphical models: they develop
a nonparametric framework for high-dimensional VARs based on GLASSO for
non-stationary multivariate time series. In addition to the concepts of conditional
dependence/independence commonly used in the graphical modelling literature, the
authors introduce the concepts of conditional stationarity/non-stationarity. The paper
demonstrates that the non-stationary graph structure can be learned from finite-length
time series in the Fourier domain. They conclude with numerical experiments showing
the feasibility of the proposed method.
8.5 Further Integration of Graphical Models with Machine

Learning
Given that many economic phenomena can be visualised as networks, graphical

models can serve as a useful tool to infer the structure of such networks. In this
chapter we aimed at reviewing recent advances in graphical modelling that have
made the latter more suitable for finance and economics applications. Tracing back
284 Seregina
historical development of commonly used econometric approaches, including non-

parametric and Bayesian ones, they started with a simplified environment, such as a
low-dimensional setup, and developed into more complex frameworks that deal with
high dimensions, missing observations, non-stationarity etc. As these developments
were growing, more parallels evolved suggesting that machine learning methods
can be viewed as complements rather than substitutes of the existing approaches.
Our stand is that the literature on graphical modelling is still getting integrated into
economic and finance applications, and a further development would be to establish
common grounds for this class of models with other machine learning methods. We
wrap up the chapter with a discussion on the latter.
To start off, instead of investing in all available stocks (as was implicitly assumed
in Section 8.3), several works focus on selecting and managing a subset of financial
instruments such as stocks, bonds, and other securities. Such technique is referred to as
a sparse portfolio. This stream of literature integrates graphical models with shrinkage
techniques (such as LASSO) and reinforcement learning. To illustrate, Seregina
(2021) proposes a framework for constructing a sparse portfolio in high dimensions
using nodewise regression. The author reformulates Markowitz portfolio allocation
exercise as a constrained regression problem, where the portfolio weights are shrunk
to zero. The paper finds that in contrast to non-sparse counterparts, sparse portfolios
are robust to recessions and can be used as hedging vehicles during such times.
Furthermore, they obtain the oracle bounds of sparse weight estimators and provide
guidance regarding their distribution. Soleymani and Paquet (2021) take a different
approach for constructing sparse portfolios: they develop a graph convolutional
reinforcement learning framework, DeepPocket, whose objective is to exploit the
time-varying interrelations between financial instruments. Their framework has three
ingredients: (1) a restricted stacked autoencoder (RSAE) for feature extraction and
dimensionality reduction. Concretely, the authors map twelve features including
opening, closing, low and high prices, and financial indicators to obtain a lower
dimensional representation of the data. They use the latter and apply (2) a graph
convolutional network (GCN) to acquire interrelations among financial instruments:
a GCN is a generalization of convolutional neural networks (CNN) to a data that has
a graph structure. The output of the GCN is passed to (3) a convolutional network for
each of the actor and the critic to enforce investment policy and estimate the return
on investment. The model is trained using historical data. The authors evaluate model
performance over three distinct investment periods including during the COVID-19
recession: they find superior performance of DeepPocket compared to market indices
and equally-weighted portfolio in terms of the return on investment.
GCNs provide an important step in merging the information contained in graphical
models with neural networks, which opens further ground to use graphical methods
in classification, link prediction, community detection, and graph embedding. Convo-
lution operator (which is defined as the integral of the product of the two functions
after one is reversed and shifted) is essential since it has proven efficient at extracting
complex features, and it represents the backbone of many deep learning models.
Graph kernels that use a “kernel trick" (Kutateladze, 2022) serve as an example of a
convolution filter used for GCNs.
References 285
Fellinghauer, Buhlmann, Ryffel, von Rhein and Reinhardt (2013) use random
forests (Breiman, Friedman, Olshen & Stone, 2017) in combination with nodewise
regression to extend graphical approaches to models with mixed-type data. They call
the proposed framework Graphical Random Forest (GRaFo). In order to determine
which edges should be included in the graphical model, the edges suggested by
the individual regressions need to be ranked such that a smaller rank indicates a
better candidate for inclusion. However, if variables are mixed-type, a global ranking
criterion is difficult to find. For instance, continuous and categorical response variables
are not directly comparable. To overcome this issue, the authors use random forests
for performing the individual nonlinear regressions and obtain the ranking scheme
from random forests’ variable importance measure. GRaFo demonstrates promising
performance on the two-health related data set for studying the interconnection of
functional health components, personal and environmental factors; and for identifying
which risk factors may be associated with adverse neurodevelopment after open-heart
surgery.
Finally, even though not yet commonly used in economic applications, a few
studies have combined graphical models with neural networks for classification tasks.
Ji and Yao (2021) develop a CNN-based model with Graphical LASSO (CNNGLasso)
to extract sparse topological features for brain disease classification. Their approach
can be summarized as a three-step procedure: (1) they develop a novel Graphical
LASSO model to reveal the sparse connectivity patterns of the brain network by
estimating the sparse inverse covariance matrices (SICs). In this model, the Cholesky
composition is performed on each SIC for ensuring its positive definite property, and
SICs are divided into several groups according to the classification task, which can
interpret the difference between patients and normal controls while maintaining the
inter-subject variability. (2) the filters of the convolutional layer are multiplied with
the estimated SICs in an element-wise manner, which aims at avoiding the redundant
features in the high level topological feature extraction. (3) the obtained sparse
topological features are used to classify patients with brain diseases from normal
controls. Ji and Yao (2021) use a CNN as a choice of a neural network, whereas for
economic application alternative models such as RNNs (Dixon & London, 2021)
and LSTMs (Zhang et al., 2019) were shown to demonstrate good performance for
time-series modelling and financial time-series prediction.
Acknowledgements The author would like to express her sincere gratitude to Chris Zhu
(ychris.zhu@gmail.com) who helped create a GitHub repository with toy examples of several
machine learning methods from the papers reviewed in this chapter. Please visit Seregina and Zhu
(2022) for further details.
References
Abeysinghe, T. (2001). Estimation of direct and indirect impact of oil price on

growth. Economics letters, 73(2), 147–153.
286 Seregina
Awoye, O. A. (2016). Markowitz minimum variance portfolio optimization using

new machine learning methods (Unpublished doctoral dissertation). University
College London.
Bai, J. (2003). Inferential theory for factor models of large dimensions. Econometrica,
71(1), 135–171. Retrieved from https://doi.org/10.1111/1468-0262.00392
Bai, J. & Ng, S. (2002). Determining the number of factors in approximate factor
models. Econometrica, 70(1), 191–221. Retrieved from https://doi.org/10.1111/
1468-0262.00273 doi: 10.1111/1468-0262.00273
Barigozzi, M., Brownlees, C. & Lugosi, G. (2018). Power-law partial correlation
network models. Electronic Journal of Statistics, 12(2), 2905–2929. Retrieved
from https://doi.org/10.1214/18-EJS1478
Basu, S., Das, S., Michailidis, G. & Purnanandam, A. (2019). A system-wide
approach to measure connectivity in the financial sector. Available at SSRN
2816137.
Basu, S., Li, X. & Michailidis, G. (2019). Low rank and structured modeling
of high-dimensional vector autoregressions. IEEE Transactions on Signal
Processing, 67(5), 1207-1222. doi: 10.1109/TSP.2018.2887401
Basu, S. & Rao, S. S. (2022). Graphical models for nonstationary time series. arXiv
Bates, J. M. & Granger, C. W. J. (1969). The combination of forecasts. Operations Re-
search, 20(4), 451–468. Retrieved from http://www.jstor.org/stable/3008764
Bishop, C. M. (2006). Pattern recognition and machine learning (information science
and statistics). Berlin, Heidelberg: Springer-Verlag.
Boyd, S., Parikh, N., Chu, E., Peleato, B. & Eckstein, J. (2011, January). Distributed
optimization and statistical learning via the alternating direction method
of multipliers. Found. Trends Mach. Learn., 3(1), 1–122. Retrieved from
http://dx.doi.org/10.1561/2200000016
Breiman, L., Friedman, J. H., Olshen, R. A. & Stone, C. J. (2017). Classification
and regression trees. Routledge.
Brownlees, C., Nualart, E. & Sun, Y. (2018). Realized networks. Journal of Applied
Econometrics, 33(7), 986-1006. Retrieved from https://onlinelibrary.wiley.com/
doi/abs/10.1002/jae.2642
Cai, T., Liu, W. & Luo, X. (2011). A constrained l1-minimization approach to sparse
precision matrix estimation. Journal of the American Statistical Association,
106(494), 594–607.
Cai, T. T., Hu, J., Li, Y. & Zheng, X. (2020). High-dimensional minimum variance
portfolio estimation based on high-frequency data. Journal of Econometrics,
214(2), 482-494.
Callot, L., Caner, M., Önder, A. O. & Ulasan, E. (2019). A nodewise regression
approach to estimating large portfolios. Journal of Business & Economic
Statistics, 0(0), 1-12. Retrieved from https://doi.org/10.1080/07350015.2019
.1683018
Chamberlain, G. & Rothschild, M. (1983). Arbitrage, factor structure, and mean-
variance analysis on large asset markets. Econometrica, 51(5), 1281–1304.
References 287
Clarke, R., de Silva, H. & Thorley, S. (2011). Minimum-variance portfolio compos-

ition. The Journal of Portfolio Management, 37(2), 31–45. Retrieved from
https://jpm.pm-research.com/content/37/2/31
Connor, G. & Korajczyk, R. A. (1988). Risk and return in an equilibrium APT:
Application of a new test methodology. Journal of Financial Economics, 21(2),
255–289. Retrieved from http://www.sciencedirect.com/science/article/pii/
0304405X88900621
Cullen, C. (1990). Matrices and linear transformations. Courier Corporation, 1990.
Retrieved from https://books.google.com/books?id=fqUTMxPsjt0C
Cushman, D. O. & Zha, T. (1997). Identifying monetary policy in a small open
economy under flexible exchange rates. Journal of Monetary economics, 39(3),
433–448.
Danaher, P., Wang, P. & Witten, D. M. (2014). The joint graphical LASSO for inverse
covariance estimation across multiple classes. Journal of the Royal Statistical
Society: Series B (Statistical Methodology), 76(2), 373-397. Retrieved from
https://rss.onlinelibrary.wiley.com/doi/abs/10.1111/rssb.12033
DeMiguel, V., Garlappi, L., Nogales, F. J. & Uppal, R. (2009). A generalized
approach to portfolio optimization: Improving performance by constraining
portfolio norms. Management Science, 55(5), 798–812. Retrieved from
http://www.jstor.org/stable/40539189
Dixon, M. & London, J. (2021). Financial forecasting with alpha-rnns: A time series
modeling approach. Frontiers in Applied Mathematics and Statistics, 6, 59.
Retrieved from https://www.frontiersin.org/article/10.3389/fams.2020.551138
Eves, H. (2012). Elementary matrix theory. Courier Corporation, 2012. Retrieved
from https://books.google.com/books?id=cMLCAgAAQBAJ
Fama, E. F. & French, K. R. (1993). Common risk factors in the returns on stocks
and bonds. Journal of Financial Economics, 33(1), 3–56. Retrieved from
http://www.sciencedirect.com/science/article/pii/0304405X93900235
Fama, E. F. & French, K. R. (2015). A five-factor asset pricing model. Journal of
Financial Economics, 116(1), 1–22. Retrieved from http://www.sciencedirect
.com/science/article/pii/S0304405X14002323
Fan, J., Liao, Y. & Mincheva, M. (2013). Large covariance estimation by thresholding
principal orthogonal complements. Journal of the Royal Statistical Society:
Series B, 75(4), 603–680.
Fan, J., Liu, H. & Wang, W. (2018, 08). Large covariance estimation through elliptical
factor models. The Annals of Statistics, 46(4), 1383–1414. Retrieved from
https://doi.org/10.1214/17-AOS1588
Fan, J., Zhang, J. & Yu, K. (2012). Vast portfolio selection with gross-exposure
constraints. Journal of the American Statistical Association, 107(498), 592-
606. Retrieved from https://doi.org/10.1080/01621459.2012.682825 (PMID:
23293404)
Farmer, R. E. (2015). The stock market crash really did cause the great recession.
Oxford Bulletin of Economics and Statistics, 77(5), 617–633.
Fellinghauer, B., Buhlmann, P., Ryffel, M., von Rhein, M. & Reinhardt, J. D.
(2013). Stable graphical model estimation with random forests for discrete,
288 Seregina
continuous, and mixed variables. Computational Statistics and Data Analysis,

64, 132-152. Retrieved from https://www.sciencedirect.com/science/article/
pii/S0167947313000789
Friedman, J., Hastie, T. & Tibshirani, R. (2007, 12). Sparse inverse covariance
estimation with the Graphical LASSO. Biostatistics, 9(3), 432-441. Retrieved
from https://doi.org/10.1093/biostatistics/kxm045
Goto, S. & Xu, Y. (2015). Improving mean variance optimization through sparse
hedging restrictions. Journal of Financial and Quantitative Analysis, 50(6),
1415–1441. doi: 10.1017/S0022109015000526
Green, R. C. & Hollifield, B. (1992). When will mean-variance efficient portfolios
be well diversified? The Journal of Finance, 47(5), 1785–1809. Retrieved
from http://www.jstor.org/stable/2328996
Hallac, D., Park, Y., Boyd, S. & Leskovec, J. (2017). Network inference via the
time-varying graphical LASSO. In Proceedings of the 23rd acm sigkdd
international conference on knowledge discovery and data mining (pp. 205–
213). New York, NY, USA: ACM. Retrieved from http://doi.acm.org/10.1145/
3097983.3098037
Hastie, T., Tibshirani, R. & Friedman, J. (2001). The elements of statistical learning.
New York, NY, USA: Springer New York Inc.
Jagannathan, R. & Ma, T. (2003). Risk reduction in large portfolios: Why imposing the
wrong constraints helps. The Journal of Finance, 58(4), 1651-1683. Retrieved
from https://onlinelibrary.wiley.com/doi/abs/10.1111/1540-6261.00580
Janková, J. & van de Geer, S. (2018). Chapter 14: Inference in high-dimensional
graphical models. In (p. 325 - 351). CRC Press.
Ji, J. & Yao, Y. (2021). Convolutional neural network with graphical LASSO to
extract sparse topological features for brain disease classification. IEEE/ACM
Transactions on Computational Biology and Bioinformatics, 18(6), 2327-2338.
doi: 10.1109/TCBB.2020.2989315
Johnstone, I. M. & Lu, A. Y. (2009). Sparse principal components analysis. arXiv
Koike, Y. (2020). De-biased graphical LASSO for high-frequency data. Entropy,
22(4), 456.
Kutateladze, V. (2022). The kernel trick for nonlinear factor modeling. Inter-
national Journal of Forecasting, 38(1), 165-177. Retrieved from https://
www.sciencedirect.com/science/article/pii/S0169207021000741
Ledoit, O. & Wolf, M. (2003). Improved estimation of the covariance matrix of stock
returns with an application to portfolio selection. Journal of Empirical Finance,
10(5), 603 - 621. Retrieved from http://www.sciencedirect.com/science/article/
pii/S0927539803000070
Ledoit, O. & Wolf, M. (2004a). Honey, I shrunk the sample covariance matrix.
The Journal of Portfolio Management, 30(4), 110–119. Retrieved from
https://jpm.iijournals.com/content/30/4/110
Ledoit, O. & Wolf, M. (2004b). A well-conditioned estimator for large-dimensional co-
variance matrices. Journal of Multivariate Analysis, 88(2), 365 - 411. Retrieved
from http://www.sciencedirect.com/science/article/pii/S0047259X03000964
References 289
doi: https://doi.org/10.1016/S0047-259X(03)00096-4
Lee, T.-H. & Seregina, E. (2021a). Learning from forecast errors: A new approach
to forecast combinations. arXiv:2011.02077.
Lee, T.-H. & Seregina, E. (2021b). Optimal portfolio using factor graphical LASSO.
arXiv:2011.00435.
Li, J. (2015). Sparse and stable portfolio selection with parameter uncertainty.
Journal of Business & Economic Statistics, 33(3), 381-392. Retrieved from
https://doi.org/10.1080/07350015.2014.954708
Lin, J. & Michailidis, G. (2017). Regularized estimation and testing for high-
dimensional multi-block vector-autoregressive models. Journal of Machine
Learning Research, 18(117), 1-49. Retrieved from http://jmlr.org/papers/v18/
17-055.html
Lu, J., Kolar, M. & Liu, H. (2015). Post-regularization inference for time-varying
nonparanormal graphical models. J. Mach. Learn. Res., 18, 203:1-203:78.
Markowitz, H. (1952). Portfolio selection*. The Journal of Finance, 7(1), 77-91.
Retrieved from https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1540-6261
.1952.tb01525.x
Meinshausen, N. & Bühlmann, P. (2006, 06). High-dimensional graphs and variable
selection with the LASSO. Ann. Statist., 34(3), 1436–1462. Retrieved from
https://doi.org/10.1214/009053606000000281
Millington, T. & Niranjan, M. (2017, 10). Robust portfolio risk minimization using
the graphical LASSO. In (p. 863-872). doi: 10.1007/978-3-319-70096-0_88
Pearl, J. (1995). Causal diagrams for empirical research. Biometrika, 82(4), 669–688.
Pesaran, M. H., Schuermann, T. & Weiner, S. M. (2004). Modeling regional
interdependencies using a global error-correcting macroeconometric model.
Journal of Business & Economic Statistics, 22(2), 129–162.
Pourahmadi, M. (2013). High-dimensional covariance estimation: With high-
dimensional data. John Wiley and Sons, 2013. Retrieved from https://
books.google.com/books?id=V3e5SxlumuMC
Ross, S. A. (1976). The arbitrage theory of capital asset pricing. Journal of Economic
Theory, 13(3), 341–360. Retrieved from http://www.sciencedirect.com/science/
article/pii/0022053176900466
Rothman, A. J., Bickel, P. J., Levina, E. & Zhu, J. (2008). Sparse permutation
invariant covariance estimation. Electron. J. Statist., 2, 494–515. Retrieved
from https://doi.org/10.1214/08-EJS176
Seregina, E. (2021). A basket half full: Sparse portfolios. arXiv:2011.04278.
Seregina, E. & Zhu, C. (2022). Chapter 8: GitHub Repository. https://github.com/
ekat92/Book-Project-Econometrics-and-MLhttps://github.com/ekat92/Book-
Project-Econometrics-and-ML.
Smith, J. & Wallis, K. F. (2009). A simple explanation of the forecast combination
puzzle. Oxford Bulletin of Economics and Statistics, 71(3), 331–355.
Soleymani, F. & Paquet, E. (2021, Nov). Deep graph convolutional reinforcement
learning for financial portfolio management – deeppocket. Expert Systems
290 Seregina
with Applications, 182, 115127. Retrieved from http://dx.doi.org/10.1016/

j.eswa.2021.115127
Stock, J. H. & Watson, M. W. (2002). Forecasting using principal components
from a large number of predictors. Journal of the American Statistical
Association, 97(460), 1167–1179. Retrieved from https://doi.org/10.1198/
016214502388618960
Tobin, J. (1958, 02). Liquidity Preference as Behaviour Towards Risk. The Review
of Economic Studies, 25(2), 65-86. Retrieved from https://doi.org/10.2307/
2296205
van de Geer, S., Buhlmann, P., Ritov, Y. & Dezeure, R. (2014, 06). On asymptotically
optimal confidence regions and tests for high-dimensional models. The
Annals of Statistics, 42(3), 1166–1202. Retrieved from https://doi.org/10.1214/
14-AOS1221
Verma, T. & Pearl, J. (1990). Causal networks: Semantics and expressiveness. In Uncer-
tainty in artificial intelligence (Vol. 9, p. 69 - 76). North-Holland. Retrieved from
http://www.sciencedirect.com/science/article/pii/B9780444886507500111
Witten, D. M. & Tibshirani, R. (2009). Covariance-regularized regression and
classification for high dimensional problems. Journal of the Royal Statistical
Society. Series B (Statistical Methodology), 71(3), 615–636. Retrieved from
Zhan, N., Sun, Y., Jakhar, A. & Liu, H. (2020). Graphical models for financial time
series and portfolio selection. In Proceedings of the first acm international
conference on ai in finance (pp. 1–6).
Zhang, X., Liang, X., Zhiyuli, A., Zhang, S., Xu, R. & Wu, B. (2019, jul). AT-
LSTM: An attention-based LSTM model for financial time series prediction.
IOP Conference Series: Materials Science and Engineering, 569(5), 052037.
Retrieved from https://doi.org/10.1088/1757-899x/569/5/052037
Zhou, S., Lafferty, J. & Wasserman, L. (2010, 1st Sep). Time varying undirected
graphs. Machine Learning, 80(2), 295–319. Retrieved from https://doi.org/
10.1007/s10994-010-5180-0
Chapter 9
Poverty, Inequality and Development Studies with
Machine Learning
Walter Sosa-Escudero, Maria Victoria Anauati and Wendy Brau
Abstract This chapter provides a hopefully complete ‘ecosystem’ of the literature on

the use of machine learning (ML) methods for poverty, inequality, and development
(PID) studies. It proposes a novel taxonomy to classify the contributions of ML
methods and new data sources used in this field. Contributions lie in two main
categories. The first is making available better measurements and forecasts of PID
indicators in terms of frequency, granularity, and coverage. The availability of more
granular measurements has been the most extensive contribution of ML to PID
studies. The second type of contribution involves the use of ML methods as well as
new data sources for causal inference. Promising ML methods for improving existent
causal inference techniques have been the main contribution in the theoretical arena,
whereas taking advantage of the increased availability of new data sources to build or
improve the outcome variable has been the main contribution in the empirical front.
These inputs would not have been possible without the improvement in computational
power.
9.1 Introduction
No aspect of empirical academic inquiry remains unaffected by the so-called ‘data

science revolution’, and development economics is not an exception. Nevertheless, the
Walter Sosa-Escudero B
Universidad de San Andres, CONICET and Centro de Estudios para el Desarrollo Humano (CEDH-
UdeSA) Buenos Aires, Argentina, e-mail: wsosa@udesa.edu.ar
Maria Victoria Anauati
Universidad de San Andres CONICET and CEDH-UdeSA, Buenos Aires, Argentina, e-mail:
vanauati@udesa.edu
Wendy Brau
Universidad de San Andres and CEDH-UdeSA, Buenos Aires, Argentina, e-mail: wbrau@udesa
.edu.ar
292 Sosa-Escudero at al.
combination of big data (tentatively defined as mostly observational data that arise
from interacting with interconnected devices) and machine learning (ML henceforth)
methods made its entrance at a moment when Economics was still embracing the
credibility revolution in empirical analysis (Angrist and Pischke (2010)): the adoption
of experimental or quasi-experimental datasets and statistical tools that allowed
researchers to identify causal effects cleanly. The Nobel Prizes awarded in a lapse of
only two years to Banerjee, Duflo and Kremer (2019), and to Angrist, Imbens and
Card (2021) are a clear accolade to this approach. In such a context, the promises of
big data/ML bring back memories of the correlation fallacies that the experimental
approach tried to avoid. Consequently, Economics is a relatively new comer to the data
science revolution that has already permeated almost every academic and professional
field.
Aside from these considerations, development economics –and its immediate
connection with policy– is a field that benefits enormously from detailed descriptions,
measurements and predictions in the complex, heterogeneous and multidimensional
contexts under which it operates. Additionally, the field has been particularly successful
in exploiting observational data to find causal channels, either through meticulous
institutional or historic analysis that isolate exogenous variation in such datasets (as
in quasi experimental studies) and/or by using tools specifically aimed at dealing with
endogeneities, such as difference-in-difference, instrumental variables or discontinuity
design strategies. The tension between these concerns and opportunities may explain
why, though relatively late, the use of ML techniques in development studies virtually
exploded in the last years.
This chapter provides a complete picture of the use of ML for poverty, inequality
development (PID) studies. The rate at which such studies have accumulated very
recently makes it impossible to offer an exhaustive panorama that is not prematurely
obsolete. Hence, the main goal of this chapter is to provide a useful taxonomy of the
contribution of ML methods in this field that helps understand the main advantages
and limitations of this approach.
Most of the chapter is devoted to two sections. The first one focuses on the use
of ML to provide better measurements and forecasts of poverty, inequality and
other development indicators. Monitoring and measurement is a crucial aspect of
development studies, and the availability of big data and ML provides an invaluable
opportunity to either improve existing indicators or illuminate aspects of social and
economic behavior that are difficult or impossible to reach with standard sources
like household surveys, census or administrative records. The chapter reviews such
improvements, in terms of combining standard data sources with non-traditional ones,
like satellite images, cell phone usage, social media and other digital fingerprints.
The use of such data sets and modern ML techniques led to dramatic improvements
in terms of more granular measurements (either temporal or geographically) or
being able to reach otherwise difficult regions like rural areas or urban slums. The
construction of indexes –as an application of non-supervised methods like modern
versions of principal component analysis (PCA)– and the use of regularization tools
to solve difficult missing data problems are also an important part of the use of ML in
development studies and are reviewed in Section 9.2. The number of relevant articles
on the subject is copious. This chapter focuses on a subset of them that help provide
a complete picture of the use of ML in development studies. The Electronic Online
Supplement contains a complete list of all articles reviewed in this chapter. Similarly,
Table 9.3 details the ML methods used in the articles reviewed.
The second part of this chapter focuses on the still emerging field of causal
inference with ML methods (see Chapter 3). Contributions come from the possibility
of exploiting ML to unveil heterogeneities in treatment effects, improve treatment
design, build better counterfactuals and to combine observational and experimental
datasets that further facilitate clean identification of causal effects. Table 9.1 describes
the main contributions of these two branches and highlights the key articles of each.
A final section explores advantages arising directly from the availability of more
computing power, in terms of faster algorithms or the use of computer based strategies
to either generate or interact with data, like bots or modern data visualization tools.
9.2 Measurement and Forecasting
ML help improve measurements and forecasts of poverty, inequality, and development

(PID) in three ways (see Table 9.1). First, ML tools can be used to combine
different data sources to improve data availability in terms time frequency and spatial
disaggregation (granularity) or extension (coverage). They are mainly supervised
ML algorithms –both for regression and classification– that are used for this purpose:
the response variable to be predicted is a PID indicator. Therefore, improving the
granularity or frequency of PID indicators implies predicting unobserved data points
using ML. Second, ML methods can be used to reduce data dimensionality, which
is useful to build indexes and characterize groups (using mainly unsupervised ML
algorithms), or to select a subset of relevant variables (through both supervised and
non-supervised methods) in order to design shorter and cheaper surveys. Finally, ML
can solve data problems like missing observations in surveys or the lack of panel data.
The lack of detailed and high quality data has been a timely deterrent of income
and wealth distribution studies. Household surveys, the most common source for
these topics, provide reliable measures of income and consumption but at a low
frequency or with a long delay, and at low levels of disaggregation. Census data,
which attempt to solve the concern of disaggregation, usually have poor information
on income or consumption. In addition, surveys and censuses are costly, and for many
relevant issues they are available at a low frequency, a problem particularly relevant
in many developing countries.
New non-traditional data sources, such as satellite images, digital fingerprints
from cell phone calls records, social media, or the ‘Internet of Things’, can be
fruitfully combined with traditional surveys, census and administrative data, with the
aim of improving poverty predictions in terms of their availability, frequency and
granularity. In this section, we first focus on the contributions in terms of improving
availability and time frequency of estimates (Section 9.2.1). Second, we focus on
spatial granularity (Section 9.2.2). In the Electronic Online Supplement, Table 9.1
Table 9.1: Taxonomy on ML contributions to PID studies
Category Contributions Key papers
Better measurements and Combining data sources to Elbers et al. (2003), Blumenstock
forecasts improve data availability, et al. (2015), Chi et al. (2022),
frequency, and granularity; Jean et al. (2016), Caruso et al.
dimensionality reduction; data (2015).
imputation.
Causal inference Heterogeneous Treatment Effects; Chowdhury et al. (2021),

optimal treatment assignment; Chernozhukov et al. (2018a,
handling high- dimensional data 2018b), Athey and Wager (2021),
and debiased ML; Banerjee et al. (2021a, 2021b),
machine-building counterfactuals; Belloni et al. (2017), Athey et al.
leveraging new data sources; (2020), Ratledge et al. (2021),
combining observational and Huang et al. (2015).
experimental data.
provides an exhaustive list of studies that aim at improving the availability, frequency
and granularity of poverty, inequality, and development indicators, along with their
main characteristics (the scope of the paper, the data sources, the contribution, the ML
methods used). Next, we review PID studies that aim at reducing data dimensionality
(Section 9.2.3) and, finally, we center on PID studies that use ML methods to solve
missing data problems (Section 9.2.4).
9.2.1 Combining Sources to Improve Data Availability
New data sources share several relevant characteristics. They are: (1) passively
collected data, (2) relatively easy to obtain at a significantly lower cost than data
from traditional sources, (3) updated frequently, and thus useful for generating nearly
real-time estimates of regional vulnerability and for providing early warning and
real-time monitoring of vulnerable populations. Therefore, one of the most immediate
applications of data science to PID studies is to provide a source of inexpensive
and more frequent estimations that complement or supplement national statistics.
Non-traditional data sources in general do not replace official statistics. On the
contrary, traditional ones often act as the ground truth data, i.e., the measure that is
known to be real or true, used as response variable to train and test the performance
of the predictive models.
Early work explored the potential offered by data from nighttime lights: satellite
photographs taken at night that capture light emitted from the Earth’s surface. Sutton
et al. (2007) is one of first studies to apply a spatial analytic approach to the patterns
in the nighttime imagery to proxy economic activity. They predict GDP at the state
level for China, India, Turkey, and the United States for the year 2000. Elvidge et
al. (2009) calculate a poverty index using population counts and the brightness of
nighttime lights for all countries in the world. Both studies use cross-sectional data to
predict poverty in a given year. Henderson et al. (2012) is the first study that measures
real income growth using panel data. They calculate a brightness score for each
pixel of a satellite image, and then aggregate these scores over a region to obtain an
indicator that can be used as a proxy for economic activity under the assumption that
lighting is a normal good. By examining cross-country GDP growth rates between
1992 and 2008, they develop a statistical framework that combines the growth in
this light measure for each country with estimates of GDP growth from the World
Development Indicators. Results show that the light-GDP elasticity lies between 0.28
and 0.32; this result is used to predict income growth for a set of countries with very
low capacity national statistical agencies.
Since then, the use of satellite images as a proxy for local economic activity has
increased significantly (see Donaldson & Storeygard, 2016, and Bennett & Smith,
2017, for a review). They lead to reliable poverty measures on a more frequent basis
than that in traditional data sources, and they complement official statistics when there
are measurement errors or when data are not available at disaggregated levels. They
are used to track the effectiveness of poverty-reduction efforts in specific localities
and to monitor poverty across time and space. For other examples, see Michalopoulos
and Papaioannou (2014), Hodler and Raschky (2014), and Kavanagh et al. (2016),
described in the Electronic Online Supplement, Table 9.1.
Despite the promise of satellite imagery as a proxy for economic output, Chen and
Nordhaus (2015) show that for time-series analysis estimations of economic growth
from multi-temporal nighttime imagery are not sufficiently accurate. In turn, Jean et
al. (2016) warned that nightlight data is less effective at distinguishing differences
in economic activity in areas at the bottom end of the income distribution, where
satellite images appear uniformly dark. They suggest extracting information from
daytime satellite imagery, which has a much higher resolution.
Satellite imagery are a type of remote sensing data, that is, data on the physical
characteristics of an area which are detected, monitored and collected by measuring
its radiation at a distance. Apart from satellite images, numerous studies resort
to the use of digital fingerprints left by cell phone transaction records, which are
increasingly ubiquitous even in very poor regions. Past history of mobile phone use
can be employed to infer socioeconomic status in the absence of official statistics.
Eagle et al. (2010) analyze the mobile and landline network of a large proportion of
the population in the UK coupled with the Multiple Deprivation Index, a composite
measure of relative prosperity of the community. Each residential landline number is
associated with the rank of the Multiple Deprivation Index of the exchange area in
which it is located. They map each census region to the telephone exchange area with
greatest spatial overlap. Results show that the diversity of individuals’ relationships is
strongly correlated with the economic development of communities. Soto et al. (2011)
and Frias-Martinez and Virseda (2012) extend the analysis to several countries in
Latin America. A key article on improving data availability by combining traditional
data with call detail records seems to be the widely cited work by Blumenstock et al.
(2015). They propose a methodology to predict poverty at the individual level based
on the intensity of cell phone usage.
Naturally, there are other data sources from the private sector which help predict
development indicators. An important article is Chetty et al. (2020a), which builds
a public database that tracks spending, employment, and other outcomes at a high
frequency (daily) and granular level (disaggregated by ZIP code, industry, income
group, and business size) that uses anonymized data from private companies. The
authors use this database to explore the heterogeneity of COVID-19’s impact on
the U.S. economy. Another example is Farrell et al. (2020), who use administrative
banking data in combination with zip code-level characteristics to provide an estimate
of gross family income using gradient boosting machines, understood as ensemble
of many classification or regression trees (see Chapter 2), in order to improve their
predictive performance. Boosting is a popular method to improve the accuracy or
predictions by locally retraining the model to handle errors from previous training
stages (see chapter 4 and Hastie et al., 2009 for more details on boosting).
Network data is also an important source of rich information for PID studies with
ML. UN Global (2016) explores the network structure of the international postal
system to produce indicators for countries’ socioeconomic profiles, such as GDP,
Human Development Index or poverty, analyzing 14 million records of dispatches
sent between 187 countries over 2010-2014. Hristova et al. (2016) measure the
position of each country in six different global networks (trade, postal, migration,
international flights, IP and digital communications) and build proxies for a number
of socioeconomic indicators, including GDP per capita and Human Development
Index.
Several studies explore how data from internet interactions and social media can
provide proxies for economic indicators at a fine temporal resolution. This type of data
stands out for its low cost of acquisition, wide geographical coverage and real-time
update. However, one disadvantage is that access to the web is usually limited in
low-income areas, which conditions its use for the prediction of socioeconomic
indicators in the areas where it is most needed. As many of the social media data
come in the form of text, natural language processing techniques (NLP) are necessary
for processing them. NLP is a subfield of ML that analyses human language in speech
and text (a classic reference is the book by Jurafsky & Martin, 2014). For example,
Quercia et al. (2012) explore Twitter users in London communities, and study the
relationship between sentiment expressed in tweets and community socioeconomic
well-being. To this end, a word count sentiment score is calculated by counting the
number of positive and negative words. There are different dictionaries annotating
the sentiment of words, i.e., whether they are positive or negative. The authors use a
commonly used dictionary called Linguistic Inquiry Word Count that annotates 2,300
English words. An alternative to using a dictionary would be manually annotating
the sentiment of a subset of tweets, and then use that subset to train an algorithm
that learns to predict the sentiment label from text features. Then, by averaging the
sentiment score of users in each community, they calculate the gross community
happiness. Lansley and Longley (2016) use an unsupervised learning algorithm to
classify geo-tagged tweets from Inner London into 20 distinctive topic groupings and
find that users’ socioeconomic characteristics can be inferred from their behaviors on
Twitter. Liu et al. (2016) analyze nearly 200 million users’ activities over 2009-2012
in the largest social network in China (Sina Microblog) and explore the relationship
between online activities and socioeconomic indices.
Another important branch of studies focuses on tracking labor market indicators.
Ettredge et al. (2005) find a significant association between the job-search vari-
ables and the official unemployment data for the U.S. Askitas and Zimmermann
(2009) use Google keyword searches to predict unemployment rates in Germany.
González-Fernández and González-Velasco (2018), use a similar methodology in
Spain. Antenucci et al. (2014) use data from Twitter to create indexes of job loss,
job search, and job posting using PCA. Llorente et al. (2015) also use data from
Twitter to infer city-level behavioral measures, and then uncover their relationship
with unemployment.
Finally, there several studies that combine multiple non-traditional data sources.
For instance, satellite imagery data provide information about physical properties of
the land, which are cost-effective but relatively coarse in urban areas. By contrast, call
records from mobile phones have high spatial resolution in cities though insufficient in
rural areas due to the sparsity of cellphone towers. Thus, recent studies show that their
combination can produce better predictions. Steele et al. (2017) is a relevant reference.
They use remote sensing data, call detail records and traditional survey-based data
from Bangladesh to provide the first systematic evaluation of the extent to which
different sources of input data can accurately estimate different measures of poverty,
namely the Wealth Index, Progress out of Poverty Index and reported household
income. They use hierarchical bayesian geostatistical models to construct highly
granular maps of poverty for these indicators. Chi et al. (2022) use data from satellites,
mobile phone networks, topographic maps, as well as aggregated and de-identified
connectivity data from Facebook to build micro-estimates of wealth and poverty at
2.4km resolution for low and middle-income countries. They first use deep learning
methods to convert the raw data to a set of quantitative features of each village. Deep
learning algorithms such as neural networks are ML algorithms organized in layers:
an input layer, an output layer, and hidden layers connecting the input and output
layers. The input layer takes in the initial raw data, the hidden layers processes the
data using nonlinear functions with certain parameters and outputs it to the next layer,
and the output layer is connected to the last hidden layer and presents the prediction
outcome (for a reference see Chapters 4 and 6 as well as the book by Goodfellow,
Bengio & Courville, 2016). Then, they use these features to train a supervised ML
model that predicts the relative and absolute wealth of each all populated 2.4km grid
cells (see Section 9.2.2 for more details on the methodology and results of this study.
See also Njuguna & McSharry, 2017 and Pokhriyal & Jacques, 2017 for further
examples).
Many data sources, such as phone call records, are often proprietary. Therefore, the
usefulness of the ML methods and their reproducibility depends on the accessibility
to these data. The work by Chetty et al. (2014), Chetty et al. (2018) or Blumenstock et
al. (2015) are examples of successful partnerships between governments, researchers
and the private sector. However, according to Lazer et al. (2020), access to data
from private companies is rarely available and when it is, the access is generally
established on an ad-hoc basis. On their articles on the contributions of big data to
development and social sciences, both Blumenstock (2018a) and Lazer et al. (2020)
call for fostering collaboration among scientists, development experts, governments,
civil society and the private sector for establishing clear guidelines for data-sharing.
Meanwhile, works that use publicly available data become particularly valuable. For
example, Jean et al. (2016) use satellite imagery that are publicly available and Rosati
et al. (2020) use only open-source data, many of which are collected via scraping
(see Section 9.2.2).
Moreover, these data often require dealing with privacy issues, as they are often
highly disaggregated and have sensitive personal information. Blumenstock (2018a)
states that a pitfall on the uses of big data for development is that there are few
data regulation laws and checks and balances to control access to sensitive data
in developing countries. However, Lazer et al. (2020) note that there are emerging
examples of methodologies and models that facilitate data analysis while preserving
privacy and keeping sensitive data secure. To mention a few, Blumenstock et al. (2015)
emphasize that they use an anonymized phone call records database and request for
informed consent to merge it with a follow-up phone survey database, which solicited
no personally identifying information. Chetty et al. (2020a) use anonymized data
from different private companies and took some to protect their confidentiality (such
as excluding outliers, reporting percentage changes relative to a baseline rather than
reporting levels of each series, or combining data from multiple firms).
9.2.2 More Granular Measurements
This section explores contributions of big data and ML in terms of data geolocation,
visualization techniques, and methods for data interpolation that facilitate access,
improve the granularity, and extended the coverage of development indicators to
locations where standard data are scarce or non-existent.
9.2.2.1 Data Visualization and High-Resolution Maps
One of the main contributions of ML to PID studies is the construction of high

resolution maps that help design more focused policies. The use of poverty maps has
become more widespread in the last two decades. A compilation of many of these
efforts can be found in (Chi et al., 2022). Other important development indicators in
areas such as health, education, or child-labor have also benefited from this approach,
as in Bosco et al. (2017), Graetz et al. (2018), or ILO-ECLAC (2018).
Bedi et al. (2007b) describes applications of poverty maps. They are a powerful
communication tool, as they summarize a large volume of data in a visual format that
is easy to understand while preserving spatial relationships. For instance, Chetty et
al. (2020) made available an interactive atlas of children outcomes such as earnings
distribution and incarceration rates by parental income, race and gender at the census
tract level in the United States, by linking census and federal income tax returns data.
Figure 9.1b shows a typical screenshot of their atlas. Soman et al. (2020) map an
index of access to street networks (as a proxy for slums) worldwide, and Chi et al.
(2022) map estimates of wealth and poverty around the world (see Table 9.1 in the
Electronic Online Supplement for more details).
Perhaps the main application of poverty maps is in program targeting, i.e.,
determining eligibility more precisely. Elbers et al. (2007) quantify the impact on
poverty of adopting a geographically-targeted budget using poverty maps for three
countries (Ecuador, Madagascar and Cambodia). Their simulations show that the
gains from increased granularity in targeting are important. For example, in Cambodia,
the poverty reduction that can be achieved using 54.5% of the budget with a uniform
transfer to each of the six provinces is the same as that achievable using only 30.8%
of the budget but targeting each of the 1594 communes. Finally, maps can also be
used for planning government policies at the sub-regional level, for analyzing the
relationship of poverty or other development indicators between neighboring areas,
for studying their geographic determinants or for evaluating the impact of policies
(Aiken, Bedoya, Coville & Blumenstock, 2020; Bedi, Coudouel & Simler, 2007a).
(a) Blumenstock et al. (2015) (b) Source: Chetty et al. (2020)
Fig. 9.1: Examples of visualizations of poverty and inequality maps
9.2.2.2 Interpolation
Spatial interpolation exploits the correlation between indices of population well-being

or development and geographical, socioeconomic or infrastructure features to predict
the value of indicators where survey data are not available (Bosco et al., 2017).
Researchers have developed several methods to use and combine both traditional and
non-traditional sources of data to estimate poverty or other measures of development
at a more granular level (such as the ELL method and others that use ML; see Table
9.1 in the Electronic Online Supplement).
The theoretical work of Elbers et al. (2003), coming from small area estimation
statistics and before ‘Data Science’ was even a term, is seminal in this area. Their ELL
method combines traditional survey and census data to improve granularity. Many
variables are usually available in surveys with higher temporal frequency but at an
aggregated or sparse spatial scale, and therefore are integrated with the finer scale
census data to infer the high resolution of a variable of interest. The ELL method
estimates the joint distribution of the variable of interest (avilable in the more frequent
surveys) and a vector of covariates, restricting the set of explanatory variables to
those that can be linked to observations in the census.
ML may improve the ELL method considerably, in two ways. Firstly, as supervised
ML techniques focus on predicting accurately out-of-sample and allow more flexibility,
they could improve the performance of linear imputation methods and of program
targeting when using poverty proxies treating them as a missing value that can be
predicted (Sohnesen & Stender, 2017; McBride & Nichols, 2018). Secondly, the ELL
method relies on the availability of census data, while ML methods often leverage
non-traditional data for poverty predictions. They are useful for preprocessing the data,
given that the new data available are often unstructured (i.e., they are not organized in
a predefined data format such as tables or graphs), unlabelled (i.e., each observation
is not linked with a particular value of the response variable), or high-dimensional
(e.g., image data come in the form of pixels arranged in at least two dimensions).
In particular, deep learning methods have been widely adopted to extract a subset
of numerical variables from this kind of data, as in Jean et al. (2016) or Chi et al.
(2022). In general, this subset of numerical variables are not interpretable, and are
only useful for prediction.
Most studies train supervised models to predict well-being indicators from
traditional data sources based on features from non-traditional ones. The work by
Blumenstock et al. (2015) is seminal in this approach. They take advantage of the
fact that cell phone data are available at an individual level to study the distribution
of wealth in Rwanda at a much finer level of disaggregation than official statistics
allow. The authors merge an anonymized cell phone call data set from 1.5 million
cell phone users and a follow-up phone socioeconomic survey of a geographically
stratified random sample of 856 individual subscribers. Then, they predict poverty and
wealth of individual subscribers and create regional estimates for 12,148 distinct cells,
instead of just 30 districts that the census allows (see Figure 9.1a for a visualization
of their wealth index). They start by estimating the first principal component of
several survey responses related to wealth to construct a wealth index. This wealth
index serves as the dependent variable they want to predict from the subscriber’s
historical patterns of phone use. Secondly, they use feature engineering to generate
the features of phone use. They employ a combinatorial method that automatically
generates thousands of metrics from the phone logs. Then, they train an elastic net
regularization model to predict wealth from the features on phone usage (Chapter 1
explains the elastic net as well as others shrinkage estimators). Finally, they validate
their estimations by comparing them with data collected by the Rwandan government.
Jean et al. (2016) is another seminal article. They use diurnal satellite imagery
and nighttime lights data to estimate cluster-level expenditures or assets in Nigeria,
Tanzania, Uganda, Malawi, and Rwanda. Clusters are roughly equivalent to villages
in rural areas or wards in urban areas. Using diurnal satellite imagery overcomes
the limitation of nightlights for tracking the livelihoods of the very poor due to the
fact that luminosity is low and shows little variation in areas with populations living
near and below the poverty line. In contrast, daytime satellite images can provide
more information, such as the roof material of houses or the distance to roads. Like
Blumenstock et al. (2015), their response variable is an asset index computed as
the first principal component of the Demographic and Health Surveys’ responses
to questions about asset ownership. They also use expenditure data from the World
Bank’s Living Standards Measurement Study (LSMS) surveys. The authors predict
these measures of poverty obtained from a traditional data source using features
extracted from a non-traditional data source: diurnal satellite imagery. However, these
images require careful preprocessing to extract relevant features for poverty prediction,
as they are unstructured and unlabelled. To overcome these challenges, Jean et al.
(2016) use Convolutional Neural Networks (CNNs) and a three-step transfer learning
approach. The three steps are the following:
1. Pretraining a CNN on ImageNet, a large dataset with labelled images. CNNs are
a particular type of neural network, widely used for image classification tasks,
that has at least one convolution layer. Briefly, each image can be thought of
as a matrix or array, where each pixel is represented with a numerical value.
Then, the convolution layer consists of performing matrix products between the
original matrix and other matrices called convolution filters, which are useful for
determining whether a local or low-level feature (e.g., certain shape or edges)
is present in an image. Therefore, in this first step the model learns to identify
low-level image features.
2. Fine-tuning the model for a more specific task: training it to estimate nighttime
light intensities, which proxy economic activity, from the input daytime satellite
imagery. Therefore, this second step is useful for extracting image features which
are relevant for poverty prediction, if some of the features that explain variation
in nightlights are also predictive of economic outcomes.
3. Training Ridge regression models to predict cluster-level expenditures and assets
using the features obtained in the previous steps.
One of the main contributions of the work by Jean et al. (2016) is that the satellite
images used are publicly available. In the same venue, Rosati et al. (2020) map health
vulnerability at the census block level in Argentina using only open-source data. They
create an index of health vulnerability using dimensionality reduction techniques).
Bosco et al. (2017) also take advantage of the availability of geolocated household
survey data. They apply bayesian learning methods and neural networks to predict
and map literacy, stunting and the use of modern contraceptive methods from a
combination of other demographic and health variables, and find that the accuracy of
these methods in producing high resolution maps disaggregated by gender is overall
high but varies substantially by country and by dependent variable.
At the end of the road, and probably reflecting the state of the art in the field,
is Chi et al. (2022). They obtain more than 19 million micro-estimates of wealth
(both absolute and relative measures) that cover the populated surface of all 135
low and middle-income countries at a 2.4km resolution. Wealth information from
Table 9.2: ML methods for data interpolation in three key papers
Paper Response variable Features Prediction
Blumenstock 1𝑠𝑡 principal Data engineering from phone calls Elastic net
et al. (2015) component from data Regularization
survey questions
Jean et al. 1𝑠𝑡 principal CNN and transfer learning from Ridge
(2016) component from satellite images regularization
survey questions
Chi et al. 1𝑠𝑡 principal CNN from satellite images + Gradient

(2022) component from features from phone calls and social boosting
survey questions media
traditional face-to-face Demographic and Health Surveys covering more than 1.3
million households in 56 different countries is the ground truth data. Once again, as
in the previous works, a relative wealth index is calculated by taking the first principal
component of 15 questions related to assets and housing characteristics. In turn, the
authors combine a variety of non-traditional data sources –satellites, mobile phone
networks, topographic maps and aggregated and de-identified connectivity data from
Facebook– to build features for each micro-region. In the case of satellite imagery,
the authors follow Jean et al. (2016) and use CNN to get 2048 features from each
image, and then keep their first 100 principal components. From these features they
train a gradient boosted regression tree to predict measurements of household wealth
collected through surveys, and find that including different data sources improves
the performance of their model, and that features related to mobile connectivity are
among the most predictive ones.
The algorithms of Chi et al. (2022) have also been implemented for aid-targeting
during the Covid-19 crisis by the governments of Nigeria and Togo. For the case
of Togo, Aiken et al. (2021) quantify that ML methods reduce errors of exclusion
by 4-21% relative to other geographic options considered by the government at the
time (such as making all individuals within the poorest prefectures or poorest cantons
eligible, or targeting informal workers). Likewise, Aiken et al. (2020) conclude that
supervised learning methods leveraging mobile phone data are useful for targeting
beneficiaries of an antipoverty program in Afghanistan. The errors of inclusion
and exclusion when identifying the phone-owning ultra-poor are similar to when
using survey-based measures of welfare, and methods combining mobile phone
data and survey data are more accurate than methods using any one of the data
sources. However, they highlight that its utility is limited by incomplete mobile phone
penetration among the program beneficiaries.
Table 9.2 outlines the methods used in the three key papers leveraging ML
techniques for interpolation in PID studies.
9.2.2.3 Extended Regional Coverage
Rural areas and informal settlements (slums) are typically underrepresented in official
statistics. Many studies using ML or non-traditional data sources have contributed to
extend data coverage to these areas.
Several of the interpolation studies mentioned before estimate development
indicators in rural areas (see the Electronic Online Supplement, Table 9.1 for a
detailed description of these studies). For example, Chi et al. (2022) show that their
model can differentiate variation in wealth within rural areas, and Aiken et al. (2021)
specifically designed their model to determine eligibility in a rural assistance program.
The model in Engstrom et al. (2017), using day-time and night-time satellite imagery
is even more accurate in rural areas than in urban ones. Watmough et al. (2019) use
remote sensor satellite data to predict rural poverty at the household level.
The detection of high poverty areas such as slums (i.e., urban poverty) has been
widely explored. It has its origins in spatial statistics and geographical information
systems literature, and the body of work has increased since the availability of high-
resolution remote sensing data, satellite and street view images and georeferenced
data from crowd-sourced maps, as well as new ML methods to process them. Kuffer
et al. (2016) and Mahabir et al. (2018) review studies using remote sensing data for
slum mapping.
A variety of methods have been employed to detect slums. For instance, image
texture analysis extracts features based on the shape, size and tonal variation within an
image. The most frequently used method is object-based image analysis (commonly
known as OBIA), a set of techniques to segment the image into meaningful objects
by grouping adjacent pixels. Texture analysis might serve as input for OBIA as in
Kohli et al. (2012) and Kohli et al. (2016), and OBIA can also be combined with
georeferenced data (GEOBIA techniques). More recently, ML techniques have been
added to the toolkit for slum mapping, and they have been highly accurate (Kuffer
et al., 2016; Mahabir et al., 2018). Supervised ML algorithms use a subset of data
labelled as slums to learn which combination of data features are relevant to identify
slums, in order to minimize the error when predicting if there are slums in unlabeled
areas. Roughly, two different supervised ML approaches have been used. The first one
is to use a two-step strategy (Soman et al., 2020): preprocessing the data to extract an
ex-ante and ad-hoc set of features, and then training a model to classify areas into
slums or not. Different techniques can be combined for feature extraction, such as
texture analysis or GEOBIA for images, and different algorithms can be used for the
classification tasks. Some of the works using this approach are Baylé (2016), Graesser
et al. (2012), Wurm et al. (2017), Owen and Wong (2013), Huang et al. (2015),
Schmitt et al. (2018), Dahmani et al. (2014) and Khelifa and Mimoun (2012). The
second and more recent approach is using deep learning techniques, where feature
extraction takes place automatically and jointly with classification, instead of being
defined ad-hoc and ex-ante. For example, Maiya and Babu (2018) and Wurm et al.
(2017) use CNNs.
9.2.2.4 Extrapolation
The evidence on the possibility of extrapolating the results of a model trained in a

certain location to a different one is mixed. For example, Bosco et al. (2017) find
a large variability in the accuracy of the models across countries in their study.
This means that not all the geolocated socioeconomic variables are equally useful
for predicting education and health indicators in every country, so a model that
works well in one country cannot be expected to work well in another. In this
respect, Blumenstock (2018b) find that models trained using mobile phone data from
Afghanistan are quite inaccurate to predict the wealth in Rwanda and vice-versa. The
multi-data supervised models of Chi et al. (2022) are also generalizable between
countries and they find that ML models trained in one country are more accurate
to predict wealth in other countries when applied to neighboring countries and to
countries with similar observable characteristics.
Regarding slum detection, given that the morphological structure of slums varies in
different parts of the world and at different growth stages, models often need to be re-
trained for different contexts (Taubenböck, Kraff & Wurm, 2018; Soman et al., 2020).
As a result, slum mapping tends to be oriented to specific areas (Mahabir et al., 2018),
although the possibility of generalizing results varies across methods. According
to Kuffer et al. (2016), texture-based methods are among the most robust across
cities and imagery. Soman et al. (2020) develop a methodology based on topological
features from open source map data with the explicit purpose of generalizing results
across the world.
In sum, there are two kinds of factors that hinder extrapolation. First, the het-
erogeneity of the relationship between observable characteristics and development
indicators across different locations. Second, there may be noise inherent to the data
collection process, such as satellite images taken at different times of the day or in
different seasons, although the increased collection frequency of satellite images
may help to tackle it. However, domain generalization techniques, which have been
increasingly applied in other fields (Dullerud et al., 2021; Zhuo & Tan, 2021; Kim,
Kim, Kim, Kim & Kim, 2019; Gulrajani & Lopez-Paz, 2020), are still under-explored
in poverty and development studies. One example is the work by Wald et al. (2021),
that introduces methods to improve multi-domain calibration by training or modifying
an existing model that achieve better performance on unseen domains, and apply their
methods to the dataset collected by Yeh et al. (2020).
9.2.3 Dimensionality Reduction
Much of the literature points towards the multidimensional nature of welfare (Sen,
1985), which translates almost directly into that of poverty or deprivation. Even
when there is an agreement on the multidimensionality of well-being, there remains
the problem of deciding how many dimensions are relevant. This problem of
dimensionality reduction is important not only for a more accurate prediction of
poverty but also for a more precise identification of which variables are relevant in
order to design shorter surveys with lower non-response rates and lower costs.
Non-supervised ML techniques, such as clustering and factor analysis, can contrib-
ute towards that direction. Traditionally, research has addressed the multidimensional-
ity of welfare by first reducing the dimensionality of the original welfare space using
factor methods, and finally proceeding to identify the poor based on this reduced set
of variables. For instance, Gasparini et al. (2013) apply factor analysis to 12 variables
in the Gallup World Poll for Latin American countries, concluding that three factors
are necessary for representing welfare (related to income, subjective well-being and
basic needs). In turn, Luzzi et al. (2008) apply factor analysis to 32 variables in the
Swiss Household Panel, concluding that four factors suffice (summarizing financial,
health, neighborhood and social exclusion conditions). Then, they use these factors
for cluster analysis in order to identify the poor.
Instead, Caruso et al. (2015) propose a novel methodology that first identifies the
poor and then explores the dimensionality of welfare. They first identify the poor by
applying clustering methods on a rather large set of attributes and then reduce the
dimension of the original welfare space by finding the smallest set of attributes that
can reproduce as accurately as possible the poor/non-poor classification obtained
in the first stage, based on a ‘blinding’ variable selection method as in Fraiman
et al. (2008). The reduced set of variables identified in the second stage is a strict
subset of the variables originally in the welfare space, and hence readily interpretable.
Therefore, their methodology overcomes one of the limitations of PCA and factor
analysis: they result in an index that is by construction a linear combination of all the
original features. To solve this problem Gasparini et al. (2013) and Luzzi et al. (2008)
use rotation techniques. However, rotations cannot guarantee that enough variables
have a zero loading in each component to render them interpretable. In turn, Merola
and Baulch (2019) suggest using sparse PCA to obtain more interpretable asset
indexes from household survey data in Vietnam and Laos. Sparse PCA techniques
embed regularization methods into PCA so as to obtain principal components with
sparse loadings, that is, each principal component is a combination of only a subset
of the original variables.
Edo et al. (2021) aim at identifying the middle class. They propose a method for
building multidimensional well-being quantiles from an unidimensional well-being
index obtained with PCA. Then, they reduce the dimensionality of welfare using the
‘blinding’ variable selection method of Fraiman et al. (2008). Others use supervised
ML methods to select a subset of relevant variables. Thoplan (2014) apply random
forests to identify the key variables that predict poverty in Mauritius, and Okiabera
(2020) use this same algorithm to identify key determinants of poverty in Kenya.
Mohamud and Gerek (2019) use the wrapper feature selector in order to find a set of
features that allows classifying a household into four possible poverty levels.
Another application is poverty targeting, which generally implies ranking or classi-
fying households from poorest to wealthiest and selecting the program beneficiaries.
Proxy Means Tests are common tools for this task. They consist of selecting from
a large set of potential observables a subset of household characteristics that can
account for a substantial amount of the variation in the dependent variable. For that,
stepwise regressions are in general used, and the best performing tool is selected
following the criteria of the best in-sample performance. Once the Proxy Means
Tests has been estimated from a sample, the tool can be applied to the subpopulation
selected for intervention to rank or classify households according to Proxy Means
Tests score. This process involves the implementation of a household survey in the
targeted subpopulation so as to assign values for each of the household characteristics
identified during the tool development. ML methods –in particular, ensemble methods–
have been shown to outperform Proxy Mean Tests. McBride and Nichols (2018) show
that regression forests and quantile regression forests algorithms can substantially
improve the out-of-sample performance of Proxy Mean Tests. These methods have
the advantage of selecting the variables that offer the greatest predictive accuracy
without the need to resort to stepwise regression and/or running multiple model
specifications.
9.2.4 Data Imputation
In addition to improving the temporal and spatial frequency of estimates, ML

techniques have been exploited to solve other missing data problems. Rosati (2017)
compares the performance of an ensemble of LASSO regression models against the
traditional ‘hot-deck’ imputation method for missing data in Argentina’s Permanent
Household Survey. Chapter 1 in this book discusses the LASSO in detail.
ML methods can also be used to compensate for the lack of panel data, a very
promising area of research that brings together econometrics, ML and PID studies.
The exploitation of panel data is at the heart of the contributions of econometrics, but
they are often not available, or suffer from non-random attrition problems. Adding
to the literature on synthetic panels for welfare dynamics (Dang, Lanjouw, Luoto &
McKenzie, 2014, which also builds on Elbers et al., 2003), Lucchetti (2018) uses
LASSO to estimate economic mobility from cross-sectional data. Later on, Lucchetti
et al. (2018) propose to combine LASSO with predictive mean matching (LASSO-
PMM). Although this methodology does not substitute panel data, Lucchetti et al.
(2018)’s findings are sufficiently encouraging to suggest that estimating economic
mobility using LASSO-PMM may approximate actual welfare indicators in settings
where cross-sections are routinely collected, but where panel data are unavailable.
Another example is Feigenbaum (2016) who uses ML models and text comparison
to link individuals across datasets that lack clean identifiers and which are rife with
measurement and transcription issues in order to achieve a better understanding of
intergenerational mobility. The methodology is applied to match children from the
1915 Iowa State Census to their adult-selves in the 1940 Federal Census.
Finally, Athey et al. (2021) leverage the computer science and statistics literature
on matrix completion for imputing the missing elements in a matrix to develop a
method for constructing credible counterfactuals in panel data models. In the same
vein, Doudchenko and Imbens (2016) propose a more flexible version of the synthetic
control method using elastic net. Both methods are described in Section 9.3.4, as
their main objective is improving causal inference (see for example, Ratledge et al.,
2021; Clay, Egedesø, Hansen, Jensen & Calkins, 2020; Kim & Koh, 2022).
9.2.5 Methods
The previous discussion is organized by the type of application of ML methods

(source combination, granularity and dimensionality). An alternative route is to
describe the contributions organized by the type of ML method they use. Table 9.3
adopts this approach and organizes the reviewed literature in Section 9.2 using in
the standard classification of ML tools, in terms of ‘supervised versus unsupervised
learning’. To save space, the Table identifies only one paper that, in our opinion, best
exemplifies the use of each technique. In the Electronic Online Supplement, Table
9.2 provides a complete list of the reviewed articles, classified by method.
9.3 Causal Inference
An empirical driving force of development studies in the last two decades is the
possibility of identifying clean causal channels through which policies affect outcomes.
In the last years, the flexible nature of ML methods has been exploited in many
dimensions that helped improved standard causal tools. Chapter 3 in this book presents
a detailed discussion of ML methods for estimating treatment effects. This section
describes the implementation and contribution of such methods in PID studies, in six
areas: estimation of heterogeneous effects, optimal design treatment, dealing with
high-dimensional data, the construction of counterfactuals, the ability to construct
otherwise unavailable outcomes and treatments, and the possibility of using ML to
combine observational and experimental data.
9.3.1 Heterogeneous Treatment Effects
The ‘flexible’ nature of ML methods is convenient to discover and estimate hetero-

geneous treatment effects (HTE), i.e., treatment effects that vary among different
population groups, otherwise difficult to capture in the rigid and mostly linear standard
econometric specifications. Nevertheless, a crucial concern in the causal inference
literature is the ability to perform valid inference. That is, the possibility of not only
provide point estimations or predictions but to gauge its sampling variability, in
terms of confidence intervals, standard errors or p-values that facilitate comparisons
or the evaluation of relevant hypothesis. The ‘multiple hypothesis testing problem’
is particularly relevant when researchers search iteratively (automatically or not)
for treatment effect heterogeneity over a large number of covariates. Consequently,
Table 9.3: ML methods for improving PID measurements and forecasts
Method Papers
Supervised learning
Trees and ensembles Decision and regression Aiken et al. (2020)

trees
Boosting Chi et al. (2022)
Random forest Aiken et al. (2020)
Nonlinear regression Generalized additive models Burstein et al. (2018)
methods
Gaussian process regression Pokhriyal and Jacques (2017)
Lowess regression Chetty et al. (2018)
K nearest neighbors Yeh et al. (2020)
Naive Bayes Venerandi et al. (2015)
Discriminant analysis Robinson et al. (2007)
Support vector machines Glaeser et al. (2018)
Regularization and LASSO Lucchetti (2018)
feature selection
Ridge regression Jean et al. (2016)
Elastic net Blumenstock et al. (2015)
Wrapper feature selector Afzal et al. (2015)
Correlation feature selector Gevaert et al. (2016)
Other spatial regression methods Burstein et al. (2018)
Deep learning Neural networks Jean et al. (2016)
Unsupervised learning
Factor analysis Gasparini et al. (2013)

PCA (including its derivations, eg., sparse PCA) Blumenstock et al. (2015),
Clustering methods (e.g., k-means) M. Burke et al. (2016)
Processing new data
Natural Language Processing Sheehan et al. (2019)

Other authomatized feature extraction (not deep learning) Blumenstock et al. Blumenstock
et al. (2015)
Network analysis Eagle et al. (2010)
the implementation of ML methods for causal analysis must deal not only with the
standard non-linearities favored by descriptive-predictive analysis, but also satisfy
the inferential requirements of impact evaluation of policies.
Instead of estimating the population marginal average treatment effect (ATE),
𝐴𝑇 𝐸 ≡ 𝐸 [𝑌𝑖 (𝑇𝑖 = 1) −𝑌𝑖 (𝑇𝑖 = 0)]
the growing literature on HTE has proposed several parametric, semi-parametric,

and non-parametric approaches to estimate the conditional average treatment effect
(CATE):
𝐶 𝐴𝑇 𝐸 ≡ 𝐸 [𝑌𝑖 (𝑇𝑖 = 1) −𝑌𝑖 (𝑇𝑖 = 0)|𝑋𝑖 = 𝑥]
A relevant part of the literature on HTE builds on regression tree methods, based
on the seminal paper by Athey and Imbens (2016) (Chapter 3 of this book explains
HTE methods in detail). Trees are a data-driven approach that finds a partition of the
covariate space that groups observations with similar outcomes. When the partition
groups observations with different outcomes, trees are usually called ‘decision trees’.
Instead, Athey and Imbens (2016) propose methods for building ‘causal trees’: trees
that partition the data into subgroups that differ by the magnitude of their treatment
effects. To determine the splits, they propose using a different objective function that
rewards increases in the variance of treatment effects across leaves and penalizes
splits that increase within-leave variance. Causal trees allow for valid inference
for the estimated causal effects in randomized experiments and in observational
studies satisfying unconfoundedness. To ensure valid estimates, the authors propose
a sample-splitting approach, which they call ‘honesty’. It consists of using one subset
of the data to estimate the model parameters (i.e., the tree structure), and a different
subset (the ‘estimation sample’) to estimate the average treatment effect in each leaf.
Thus, the asymptotic properties of treatment effect estimates within leaves are the
same as if the tree partition had been exogenously given. Finally, similar to standard
regression trees, prunning proceeds by cross-validation, but in this case, the criterion
for evaluating the performance of the tree in held-out data is based on treatment effect
heterogeneity instead of predictive accuracy.
The causal tree method has several advantages. It is easy to explain, and, in the
case of a randomized experiment, it is convenient to interpret, as the estimate in
each leaf is simply the sample average treatment effect. As is the case of decision
trees, an important disadvantage is their high variance. Wager and Athey (2018)
extend the standard causal tree strategy to random forests, which they refer to as the
causal forests method. Essentially, a causal forest is the average of a large number
of causal trees, where trees differ from one another due to resampling. The authors
establish asymptotic normality results for the estimates of treatment effects under the
unconfoundedness assumption, allowing for valid statistical inference. They show
that causal forest estimates are consistent for the true treatment effect, and have an
asymptotically gaussian and centered sampling distribution if each individual tree in
the forest is estimated using honesty as previously defined, under some more subtle
assumptions regarding the size of the subsample used to grow each tree. In addition,
this method outperforms other nonparametric estimation methods of HTE such as
nearest neighbors or kernels as it is useful for mitigating the ‘curse of dimensionality’

in high-dimensional cases. Forests can be thought of as a nearest neighbors method
with an adaptive neighborhood metric, where closeness between observations is
defined with respect to a decision tree, and the closest points fall in the same leaf.
Therefore, there is a data-driven approach to determine which dimensions of the
covariate space are important to consider when selecting nearest neighbors, and
which can be discarded.
Athey et al. (2019) generalize causal forests to estimate HTE in an instrumental
variables framework. Let 𝑥 be a specific value of the space of covariates, the goal is
to estimate treatment effects 𝜏 that vary with 𝑥. To estimate the function 𝜏(𝑥), the
authors use forest-based algorithms to define similarity weights 𝛼𝑖 (𝑥) that measure
the relevance of the i-th observation for estimating the treatment effect at 𝑥. The
weights are defined as the number of times that the 𝑖-th observation ended up in
the same terminal leaf than 𝑥 in a causal forest. Then, a local generalized method
of moments is used to estimate treatment effects for a particular value of 𝑥, where
observations are weighted.
In turn, Chernozhukov et al. (2018a) propose an approach that, instead of using a
specific ML tool, e.g., the tree-based algorithm, applies generic techniques to explore
heterogeneous effects in randomized experiments. The method focuses on providing
valid inference on certain key features of CATE: linear predictors of the heterogeneous
effects, average effects sorted by impact groups, and average characteristics of most
and least impacted units. The empirical strategy relies on building a proxy predictor
of the CATE and then developing valid inference on the key features of the CATE
based on this proxy predictor. The method starts by splitting the data into an auxiliary
and a main sample. Using the auxiliary sample, a proxy predictor for the CATE is
constructuted with any ML method (e.g., elastic net, random forests, neural networks,
etc.). Then, the main sample and the proxy predictor are used for estimating the key
features of the CATE. There are other approaches to estimate HTE, and the literature
is fast-growing (see Chapter 3).
Regarding applications in PID studies, Chowdhury et al. (2021) use two of the
above ML approaches to investigate the heterogeneity in the effects of an RCT of an
antipoverty intervention, based on the ‘ultra-poor graduation model’, in Bangladesh.
This intervention model is composed of a sequence of supports including a grant of
productive assets, hands-on coaching for 12-24 months, life-skills training, short-term
consumption support, and access to financial services. The goal is that the transferred
assets help to develop micro-enterprises, while all the other components are related
to protecting the enterprise and/or increasing productivity. The authors explore the
trade-off between immediate reduction in poverty, measured by consumption, and
building assets for longer-term gains among the beneficiaries of the program. In order
to handle heterogeneous impacts, they use the causal forest estimator of Athey et al.
(2019) and the generic ML approach proposed by Chernozhukov et al. (2018a), and
focus on two outcomes, household wealth and expenditure. They find a large degree
of heterogeneity in treatment effects, especially on asset accumulation.
Chernozhukov et al. (2018a) illustrate their approach with an application to a
RCT aimed at evaluating the effect of nudges on demand for immunization in India.
The intervention cross-randomized three main nudges (monetary incentives, sending

SMS reminders, and seeding ambassadors). Other relevant contributions that apply
Chernozhukov et al. (2018a)’s methods to PID studies include Mullally et al. (2021)
who study the heterogeneous impact of a program of livestock transfers in Guatemala;
and Christiansen and Weeks (2020), who analyze the heterogeneous impact of
increased access to microcredit using data from three other studies in Morocco,
Mongolia and Bosnia and Herzegovina. Both papers find that, for some outcomes,
insignificant average effects might mask heterogeneous impacts for different groups;
and they both use elastic net and random forest to create the proxy predictor. Deryugina
et al. (2019) use the generic ML method to explore the HTE of air pollution on
mortality, health care use, and medical costs. They find that life expectancy as well as
its determinants (e.g., advanced age, presence of serious chronic conditions, and high
medical spending) vary systematically with pollution vulnerability, and individuals
identified as most vulnerable to pollution have significantly lower life expectancies
than those identified as least vulnerable.
Other studies applied the generalized random forest method of Athey et al. (2019)
on PID topics. For instance, Carter et al. (2019) employed it to understand the
source of the heterogeneity in the impacts of a rural business development program
in Nicaragua on three outcomes: income, investment and per-capita household
consumption expenditures. First, the authors find that the marginal impact varies
across the conditional distribution of the outcome variables using conditional quantile
regression methods. Then, they use a generalized random forest to identify the
observable characteristics that predict which households are likely to benefit from
the program. Daoud and Johansson (2019) use generalized random forest to estimate
the heterogeneity of the impacts of International Monetary Fund programs on child
poverty. Farbmacher et al. (2021) also use this methodology to estimate how the
cognitive effects of poverty vary among groups with different characteristics.
Finally, other authors have applied causal random forest to study HTE of labor
market programs. For instance, Davis and Heller (2017, 2020) use this method to
estimate the HTE of two randomized controlled trials of a youth summer jobs program.
They find that the subgroup that improved employment after the program is younger,
more engaged in school, more Hispanic, more female, and less likely to have an
arrest record. Knaus et al. (2020) examine the HTE of job search programmes for
unemployed workers concluding that unemployed persons with fewer employment
opportunities profit more from participating in these programs (see also Bertrand,
Crépon, Marguerie & Premand, 2017; Strittmatter, 2019). Finally, J. Burke et al.
(2019) use causal random forests but to explore the heterogeneous effect of credit
builder loan program on borrowers, providers, and credit market information, finding
significant heterogeneity, most starkly with respect to baseline installment credit
activity.
9.3.2 Optimal Treatment Assignment
One of the motivations for understanding treatment effect heterogeneity is informing

the decision on who to treat, or optimally assigning each individual to a specific
treatment. The problem of treatment assignments is present in a number of situations,
for example when the government has to decide which financial-aid package (if any)
should be given out to which college students, who will benefit most from receiving
poverty aid or which subgroup of people enrolled in a program will benefit the most
from it. Ideally, individuals should be assigned to the treatment associated with the
most beneficial outcome. Identifying the needs of specific sub-groups within the target
population can improve the cost effectiveness of any intervention. A growing literature
has shown the power of ML methods to address this problem in both observational
and randomized controlled trials. Most of these studies are still concerned with the
statistical properties of their proposed methods (Kitagawa & Tetenov, 2018; Athey &
Wager, 2021), hence applied studies naturally still lag behind.
When the objective is to learn a rule or policy that maps individual’s observable
characteristics to one of the available treatments, some authors leverage on ML
methods for estimating the optimal assignment rule. Athey and Wager (2021) develop
a framework to learn an optimal policy rules that not only focuses on experimental
samples, but also allows for observational data. Specifically, the authors propose
using a new family of algorithms for choosing who to treat by minimizing the loss
from failing to use the (infeasible) ideal policy, referred to as ‘the regret’ of the policy.
The optimization problem can be thought as a classification ML task with a different
loss function to minimize. Thus, it can be solved with off-the-shelf classification tools,
such as decision trees, support vector machines or recursive partitioning, among other
methods. In particular, the algorithm proposed by the authors starts by computing
doubly robust estimators of the ATE and then selecting the assignment rule that solves
the optimization problem using decision trees, where the three depth is determined
by cross-validation. The authors illustrate their method by identifying the enrollees
of a program who are most likely to benefit from the intervention. The intervention is
the GAIN program, a welfare-to-work program that provides participants with a mix
of educational resources and job search assistance in California (USA).
Zhou et al. (2018) further study the problem of learning treatment assignment
policies in the case of multiple treatment choices using observational data. They
develop an algorithm based on three steps. First, they estimate the ATE via doubly
robust estimators. Second, they use a K-fold algorithmic structure similar to cross-
validation, where the data is divided into folds to estimate models using all data
except for one fold. But instead of using the K-fold structure for selecting the
hyperparameters or tuning models, they use it to estimate a score for each observation
and each treatment arm. Third, they solve the policy optimization problem: selecting
a policy that maximizes an objective function constructed with the outputs of the first
and second steps.
The articles described above focus on a static setting where a decision-maker has
just one observation for each subject and decides how to treat her. In contrast, other
problems of interest may involve a dynamic component whereby the decision-maker
has to decide based on time-varying covariates, for instance when to recommend

mothers to stop breastfeeding to maximize infants’ health or when to turn off
ventilators for intensive care patients to maximize health outcomes. Nie et al. (2021)
study this problem and develop a new type of doubly robust estimator for learning
such dynamic treatment rules using observational data under the assumption of
sequential ignorability. Sequential ignorability means that any confounders that affect
making a treatment choice at certain moment have already been measured by that
moment.
Finally, another application using a different approach than those described above
is Björkegren et al. (2020). They develop a ML method to infer the preferences
that are consistent with observed allocation decisions. They applied this method to
PROGRESA, a large anti-poverty program that provides cash transfers to eligible
households in Mexico. The method starts by estimating the HTE of the intervention
using the causal forests ML method. Then, they consider an allocation based on a
score or ranking, in terms of school attendance, child health and consumption. Next,
ordinal logit is used to identify the preferences consistent with the ranking between
households. Results show that the program prioritizes indigenous households, poor
households and households with children. They also evaluate the counterfactual
allocations that should have occurred had the policymaker placed higher priority on
certain types of impacts (e.g., health vs. education) or certain types of households
(e.g., lower-income or indigenous).
9.3.3 Handling High-Dimensional Data and Debiased ML
Most supervised ML methods purposedly bias estimates with the aim of improving
prediction. Notably, this is the case of regularization methods. Sparse regression
methods such as Ridge, LASSO or elastic net (discussed in Chapter 1) include a
penalty term (a ‘penalized regression adjustment’) that shrinks estimations in order
to avoid overfitting, resulting in a regularization bias.
The standard econometric solution to deal with biases is to include the necessary
controls (Ahrens et al., 2021; Athey et al., 2018). In a high dimensional context this a
delicate issue since the set of potential controls may be greater than the number of
observations. The same problem can arise at the first stage of instrumental variables,
as the set of instruments might be high-dimensional as well. High-dimensionality
poses a problem for two reasons. First, if the number of covariates is greater than
the number of observations, traditional estimation is not possible. But even if the
number of potential covariates is smaller than the number of observations, including
a large set of covariables may result in overfitting. ML methods for regularization
help to handle high-dimensional data and are precisely designed to avoid overfitting
and to allow for estimation. However, as they are designed for prediction, they bias
the estimates with the aim of reducing the prediction error and they can lead to
incorrect inference about model parameters (Chernozhukov et al., 2018b; Belloni,
Chernozhukov & Hansen, 2014a).
Belloni et al. (2012, 2014a, 2014b) propose post double selection (PDS) methods
using LASSO to address this problem. The key assumption is that the ‘true model’ for
explaining the variation in the dependent variable is approximately sparse with respect
to the available variables: either not all regressors belong in the model, or some of
the correspondent coefficients are well-approximated by zero. Therefore, the effect of
confounding factors can be controlled for up to a small approximation error (Belloni
et al., 2014b). Under this assumption, LASSO can be used to select variables that are
both relevant for the dependent variable (with the treatment variable not being subject
to selection) and to the treatment variable (Ahrens et al., 2021). Then, the authors
derive the conditions and procedures for valid inference after model selection using
PDS. Belloni et al. (2017) extend those conditions to provide inferential procedures
for parameters in program evaluation (also valid in approximately sparse models)
to a general moment-condition framework and to a wide variety of traditional and
ML methods. Chapter 3 of this book discusses in detail the literature on Double or
Debiased ML and its extensions, such as Chernozhukov et al. (2018b) who propose
an strategy applicable to many ML methods for removing the regularization bias
and the risk of overfitting. Their general idea is predicting both the outcome and the
value of treatment based on the other covariates with any ML model, obtaining the
residuals of both predictions, and then regressing both residuals.
Apart from avoiding omitted variable bias, another reason for including control
covariates is improving the efficiency of estimates, such as in randomized controlled
experiments (RCT). Therefore, regularization methods may be useful in the case of
RCTs with a high dimensional set of potential controls. Wager et al. (2016) show
that the estimates obtained by any type of regularization method with an intercept
yield unbiased estimates of the ATE in RCTs and propose a procedure for building
confidence intervals for the ATE. Other strains of work tackle specific issues of high
dimensionality in RCTs. The strategies proposed by Chernozhukov et al. (2018a) (see
Section 9.3.1) are valid in high dimensional settings. Banerjee et al. (2021b) develop a
new technique for handling high-dimensional treatment combinations. The technique
allows them to determine, among a set of 75 unique treatment combinations, which
of them are effective and which one is the most effective. This solves a trade-off that
researchers face when implementing an RCT. On the one side, they could choose
ex-ante and ad-hoc a subset of possible treatments, with the risk of letting an effective
treatment combination aside. On the other side, they could implement every possible
treatment combination, but treat only a few individuals with each of them, and then
lack the statistical power to test whether its effects are significant. Their procedure
implements many treatment combinations, but then has two steps. The first one uses
the post-LASSO procedure of Belloni and Chernozhukov (2013) to pool similar
treatments and prune ineffective treatments. The second step estimates the effect of
the most effective one, correcting for an upward bias as suggested by Andrews et al.
(2019).
Some studies apply PDS LASSO in PID studies. For instance, Banerjee et al.
(2021a) use it to select control variables and to estimate the impact on poverty of
the switch from in-kind food assistance to vouchers to purchase food on the market
in Indonesia. Martey and Armah (2020) use it to study the effect of international
migration on household expenditure, working, production hours and poverty of the

left-behind household members in Ghana; and Dutt and Tsetlin (2021) to argue
that poverty strongly impacts development outcomes. Others use the procedure for
sensitivity analysis and robustness checks, such as Churchill and Sabia (2019), who
study the effect of minimum wages on low-skilled immigrants’ well-being, and Heß et
al. (2021), who study of the impact of a development program in Gambia on economic
interactions within rural villages. Skoufias and Vinha (2020) use the debiased ML
method by Chernozhukov et al. (2018b) to study the relationships between child
stature, mother’s years of education, and indicators of early childhood development.
To illustrate their methods, Belloni et al. (2017) estimate the average and quantile
effects of eligibility and participation in the 401(k) tax-exemption Plan in the US
on assets, finding that there is a larger impact at high quantiles. Finally, Banerjee
et al. (2021b) apply their technique for multiple RCT treatments to a large-scale
experiment in India aimed at determining which among multiple possible policies
has the largest impact in increasing the number of immunizations and which is
the most cost-effective, as well as at quantifying their effects. The interventions
cross-randomized three main nudges (monetary incentives, sending SMS reminders,
and seeding ambassadors) to promote immunization, resulting in 75 unique policy
combinations. The authors found the policy with the largest impact (information
hubs, SMS reminders, incentives that increase with each immunization) and the most
cost-effective (information hubs, SMS reminders, no incentives).
9.3.4 Machine-Building Counterfactuals
ML methods can also be used to construct credible counterfactuals for the treated
units in order to estimate causal effects. That is, to impute the potential outcomes
for the treated units had not they been treated. Athey et al. (2021) and Doudchenko
and Imbens (2016) propose methods for building credible counterfactuals that rely
on regularization methods frequently used in ML. In general terms, these ML-based
methods help to overcome multiple data and causal inference challenges, usually
present in observational studies, that cannot be addressed with conventional methods.
For instance, as Ratledge et al. (2021) suggest, they contribute to tackle the problem
of different evolution of outcomes between targeted and untargeted populations and
to deal with threats to identification such as non-parallel trends in pre-treatment
outcomes, which make methods such as difference in differences unreliable.
Athey et al. (2021) leverage on the computer science and statistics literature
on matrix completion. Matrix completion methods are imputation techniques for
guessing at the missing values in a matrix. In the case of causal inference, the
matrix is that of potential control outcomes, where values are missing for the treated
unit-periods. They propose an estimator for the missing potential control outcomes
based on the observed control outcomes of the untreated unit-periods. As in the matrix
completion literature, missing elements are imputed assuming that the complete
matrix is the sum of a low rank matrix plus noise. Briefly, they model the complete
data matrix 𝑌 as 𝑌 = 𝐿 ∗ + 𝜖 where 𝐸 [𝜖 |𝐿 ∗ ] = 0 and 𝜖 is a measurement error. 𝐿 ∗ is

the low-rank matrix to be estimated via regularization, by minimizing the sum of
squares and adding a penalty term that consists of the nuclear norm of 𝐿 multiplied
by a constant 𝜆 to be chosen via cross-validation.
In turn, Doudchenko and Imbens (2016) propose a modified version of the
synthetic control estimator that uses elastic net. Again, the goal is to impute the
unobserved control outcomes for the treated unit, in order to estimate the causal effect.
The synthetic control method constructs a set of weights such that covariates and
pre-treatment outcomes of the treated unit are approximately matched by a weighted
average of control units. To do so, it imposes some restrictions: the weights should be
nonnegative and sum to one. That allows to define the weighs even if the number
of control units exceeds the number of control outcomes. The authors propose a
more general estimator that allows the weights to be nonnegative, do not restrict their
sum and allow for a permanent additive difference between the treated unit and the
controls. When the number of control units exceeds the number of pre-treatment
periods, they propose using an elastic net penalty term for obtaining the weights.
Ratledge et al. (2021) apply both of the above methods to estimate the causal
impact of electricity access on wealth asset in Uganda. The ML-based causal
inference approaches help them to overcome two challenges. First, the fact that
they deal with observational data, as the electrification expansion was not part of a
randomized experiment. They have georreferenced data on the multi-year expansion
in a grid throughout the country, so they can use the observed outcome of the
control unit-periods to impute the potential outcome of the treated unit-periods,
had not the electricity distribution taken place. Second, the ML-based methods
help to tackle the problem of different evolution of outcomes between targeted and
untargeted populations. The authors argue that they are more robust to some threats
to identification such as non-parallel trends in pre-treatment outcomes. Kim and
Koh (2022) apply matrix completion methods as a robustness check to study the
effect of the access to health insurace coverage in subjective well-being in the United
States. Finally, Clay et al. (2020) apply synthetic control with elastic net to study
the immediate and long-run effects of a community-based health intervention at the
beginning of the twentieth century in controlling tuberculosis and reducing mortality.
9.3.5 New Data Sources for Outcomes and Treatments
Many PID studies take advantage of the increased availability of new data sources
for causal identification. New data can serve to approximate outcome variables,
to identify treated units and to look for sources of exogenous variability or key
control variables. In the Electronic Online Supplement, Table 9.3 lists these studies,
classifying the non-traditional data source use in three non-exclusive categories:
(i) outcome construction, (ii) treatment or control variable construction, and (iii)
exogenous variability. It also describes the type of data source used, the evaluation
method and, when applicable, the ML method employed, among other features of
each paper.
Several aspects can be highlighted from Table 9.3 in the Electronic Online
Supplement. First, ML methods mainly intervene in an intermediate step for data
prepossessing (e.g. making use of satellite image data). However, in just a few cases
ML methods also play a role in estimating the causal effect, for example, in Ratledge
et al. (2021). Second, the most common contribution of new data sources to causal
inference studies is to build or improve the outcome variable. In many cases, existing
data provide only a partial or aggregate picture of the outcome, which makes it
difficult to measure changes in the outcome over time that can be correctly attributed
to the treatment. New sources of data, as well as ML methods, can contribute to solve
this problem.
In the Electronic Online Supplement, Table 9.3 shows that most of the causal
inference studies that rely on ML to improve the outcome measure use satellite data.
A key example is the work of Huang et al. (2021), which show that relying solely on
satellite data is enough to assess the impact of an anti-poverty RCT on household
welfare. Specifically, they measure housing quality among treatment and control
households by combining high-resolution daytime imagery and state-of-the-art deep
learning models. Then, using difference-in-differences they estimate the program
effects on housing quality (see also Bunte et al., 2017, and Villa, 2016, for similar
studies in Liberia and Colombia respectively). In turn, Ratledge et al. (2021) stand out
for using daytime satellite imagery, as well as ML-based causal inference methods,
for estimating the causal impact of rural electrification expansion on welfare. In
particular, they use CNNs to build a wealth index based on daytime satellite imagery.
Then, to estimate the ATE, they apply matrix completion and synthetic controls
with elastic net. Both methods have the advantage of being more robust against the
possibility of non-parallel trends in pre-treatment outcomes (see Section 9.3.4).
Ratledge et al. (2021) note two challenges to be addressed when the outcome
variable in an impact evaluation is measured using ML predictions based on new data
sources. Firstly, models that predict the outcome variable of interest may include the
intervention of interest as covariate itself. For instance, a satellite-based model used
to predict poverty will be unreliable for estimating the causal effect of a new road
construction on poverty, if the ML model considers whether a location is poor based
on the road. To address this first challenge, the authors predict poverty excluding
variables related to the intervention from the set of features. Secondly, for certain
outcomes such as economic well-being, predictions based on ML models often have
lower variance than the ground true variable to be predicted, and they over-predict
for poorer individuals and under-predict for wealthier ones. Hence, if the intervention
targets a specific part of the outcome distribution, as is the case of poverty programs,
bias in outcome measurement could bias estimates of treatment effects. To address
this second challenge, the authors predict poverty adding a term to the mean squared
error loss function that penalizes bias in each quintile of the wealth distribution.
Another relevant example is Hodler and Raschky (2014), who study the phe-
nomenon of regional favoritism by exploring whether subnational administrative
regions are more developed when they are the birth region of the current political
leader. They use satellite data on nighttime light intensity to approximate the outcome
measure, i.e., economic activity at the subnational level. Specifically, they build a panel
data set with 38,427 subnational regions in 126 countries and annual observations
from 1992 to 2009, and use traditional econometric techniques to estimate the effect
of birthplaces of political leaders on economic development of those areas. After
including several controls, region fixed effects and country-year dummy variables,
they find that being the birth region of political leaders results in a more intense
nighttime light, providing evidence for regional favoritism. In turn, Alix-Garcia et
al. (2013) use satellite images to study the effect of poverty-alleviation programs on
environmental degradation. In particular, they assess the impact of income transfers
on deforestation, taking advantage of both the threshold in the eligibility rule for
Oportunidades, a conditional cash transfer program in Mexico, and random variation
in the pilot phase of the program. They use satellite images to measure deforestation at
the locality level, which is the outcome variable. By applying regression discontinuity,
they find that additional income raises consumption of land-intensive goods and
increases deforestation.
Other studies use geo-referenced data from new sources, such as innovative systems
of records, to build the outcome variable. For instance, Chioda et al. (2016) take
advantage of CompStat, a software used by law enforcement to map and visualize
crime, to build crime rates at the school neighborhood level and explore the effects
on crime of reductions in poverty and inequality associated with conditional cash
transfers. Using traditional econometric techniques, such as instrumental variables
and difference-in-differences approach, they find a robust and significant negative
effect of cash transfers on crime resulting from lower poverty and inequality. Several
studies use geo-referenced data obtained from geographic information systems (GIS),
which connects data to a map, integrating location data with all types of descriptive
information (see the Electronic Online Supplement, Table 9.3 for detailed examples).
Articles using GIS data are part of a more extensive literature that intersects with
spatial econometrics. In the Electronic Online Supplement, Table 9.3 includes some
relevant causal inference studies that use this type of data in PID studies.
To a lesser extent, the growing availability of high-quality data has also been used
for identifying treated units. For example, with the aim of assessing the impact of a
cyclone on household expenditure, Warr and Aung (2019) use satellite images, in the
weeks immediately following that event, to identify the treated and control households
living in affected and not-affected areas, respectively. To build the counterfactual,
they use a statistical model to predict real expenditures at the household level, based
on survey data and controlling for region and year fixed effects and a large set
of covariates. The estimated impact of the cyclone is computed as the difference
between the observed outcome, in which the cyclone happened, and the simulated,
counterfactual outcome (see the Electronic Online Supplement, Table 9.3 , for other
examples).
Finally, a group of articles has used the new source of data as the source of
exogenous variability. For instance, this is the case of Faber and Gaubert (2019),
who assess the impact of tourism on long-term economic development. They use
a novel database containing municipality-level information as well as geographic
information systems (GIS) database including remote sensing satellite data. They
exploit geological, oceanographic, and archaeological variation in ex ante local tourism
attractiveness across the Mexican coastline as a source of exogenous variability. For
that, they build variables based on the GIS and satellite data. Using a reduced-form
regression approach, they find that tourism attractiveness has strong and significant
positive effects on municipality total employment, population, local GDP, and wages
relative to less touristic regions (see the Electronic Online Supplement, Table 9.3, for
other examples).
9.3.6 Combining Observational and Experimental Data
The big data revolution facilitates researchers the access to larger, more detailed and
representative data sets. The largely observational nature of the big data phenomenon
clashes with RCT’s, where exogenous variation is guaranteed by design. Hence, it
is a considerable challenge to combine both sources to leverage the advantages of
each: size, availability and, hopefully, external validity of observational data sets, and
clean causal identification of RCT’s.
Athey et al. (2020) focus on the particular case where both the experimental
and observational data contain the same information (a secondary –e.g., short term–
outcome, pre-treatment variables, and individual treatment assignments) except for the
primary (often long term) outcome of interest that is only available in the observational
data. The method relies on the assumptions that the RCT has both internal and external
validity and that the differences in the results between the RCT and the observational
study are due to endogenous selection into the treatment or lack of internal validity
in the observational study. In addition, they introduce a novel assumption, named
‘latent unconfoundedness’, which states that the unobserved confounders that affect
treatment assignment and the secondary outcome in the observational study are
the same unobserved confounders that affect treatment assignment and the primary
outcome. This assumption allows linking the biases in treatment-control differences in
the secondary outcome (which is estimated in the experimental data) to the biases in
treatment-control comparisons in the primary outcome (which the experimental data
are silent about). Under these assumptions, they propose three different approaches
to estimate the ATE: (1) imputing the missing primary outcome in the experimental
sample, (2) weighting the units in the observational sample to remove biases, and
(3) control function methods. They apply these methods to estimate the effect of
class size on eight grade test scores in New York without conducting an experiment.
They combine data from the New York school system (the ‘observational sample’)
and from Project STAR (the ‘experimental sample’, a RCT conducted before). As
described before, both samples include the same information except for test scores
for the eight grade (the primary outcomes), which is only in the observational
sample since the Project STAR data includes only test scores for the third grade (the
secondary outcome). Both samples include pre-treatment variables (gender, whether
the student gets a free lunch, and ethnicity). The authors arrive at the same results
using both control function and imputation methods and conclude that the biases in
the observational study are substantial, but results are more credible when they apply
their approach to adjust the results based on the experimental data.
9.4 Computing Power and Tools
The increase in computational power has been essential in the contributions of new
ML methods and the use of new data. To begin with, improvements in data storage
and processing capabilities made it possible to work with massively large data sets.
One example is Christensen et al. (2020), who needed large processing capacities
to use air pollution data at high levels of disaggregation. In turn, the quality and
resolution of satellite images, which are widely used in PID studies (see the Electronic
Online Supplement, Table 9.1), has improved over the years thanks to the launch of
new satellites (Bennett & Smith, 2017).
Also, computational power allows for more flexible models. The complexity of an
algorithm measures the amount of computer time it would take to complete given
an input of size 𝑛 (where 𝑛 is the size in units of bits needed to represent the input).
It is commonly estimated by the number of elementary operations performed for
each unit of input –for example, an algorithm with complexity 𝑂 (𝑛) has a linear time
complexity and an algorithm with complexity 𝑂 (𝑛2 ) has a quadratic time complexity.
Improvements in computing processing capacity can reduce the time required to use
an algorithm of a given complexity for processing a larger scale of data, or the time
required to use algorithms of greater complexity. Thompson et al. (2020) analyze
how deep learning performance depends on computational power in the domains of
image classification and object detection. Deep learning is particularly dependent
on computing power, due to their overparameterization and the large amount of
training data used to improve performance. Circa 2010, deep learning models were
ported to the GPU (Graphics Processing Unit) of the computer instead of in the CPU
(Central Processing Unit), accelerating notably their processing: initially yielding
a 5 to 15 times speed up, which was up to 35 times by 2012. The 2012 ImageNet
competition of AlexNet, a CNN run in a GPU, was a milestone and a turning point in
the widespread use of neural networks. Many of the papers reviewed in PID studies
with ML use deep learning, typically, CNN; the oldest dates from 2016 (see the
Electronic, Online Supplement, Table 9.1). For instance, Head et al. (2017) use
CNNs to process satellite images to measure poverty in Sub-Saharan Africa. For
large countries, such as Nigeria, they use GPU computing to train the CNN; this task
took several days. Maiya and Babu (2018) also train a CNN using the GPU to map
slums from satellite images.
In addition, a variety of new computational tools became easily accessible for PID
studies. First, the assembly of bots to conduct experiments immediately and on a
large scale via web platforms, as in Christensen et al. (2021). The authors developed
a software bot at the National Center for Supercomputing Applications that sent
fictitious queries to property managers on an online rental housing platform in order
to test the existence of discrimination patterns. The process automation allowed them
to scale data collection –they examined more than 25 thousand interactions between
property managers and fictitious renters across the 50 largest cities in the United
States– and to monitor discrimination at a low cost. The bot targeted listings that
had been listed the day before on the platform, and sent one inquiry each of the
three following day, using fictitious identities drawn in random sequence from a set
of 18 names that elicited cognitive associations with one of three ethnic categories
(African American, Hispanic/Latin, and White). Then, the bot followed the listing
for 21 days registering whether the property was rented or not. Similarly, Carol et al.
(2019) tested for the presence of discrimination in the German carpooling market by
programming a bot that sent requests to drivers from fictitious profiles. Experimental
designs using bots are also common in health and educational interventions (see, for
example, Maeda et al., 2020; Agarwal et al., 2021).
Second, the use of interactive tools for data visualization has been increasing in
PID studies, specially for mapping wealth at a granular scale (see Section 9.2.2).
Table 9.1 in the Electronic Online Supplement lists many of the available links for
viewing the visualization and, often, for data downloading. Many of them can be
accessed for free, but not all. An illustrative example of a free data visualization and
downloading tool are opportunityatlas.org from Chetty et al. (2020) (see Section
9.2.2 and a screenshot in Figure 9.1b) and tracktherecovery.org from Chetty et al.
(2020a) (see Section 9.2.1).
Third, open source developments and crowd-sourced data systems make software
and data available more accessible. Most of the techniques in the reviewed studies are
typical ML techniques for which there are open source Python and R libraries. For
example, the Inter-American Development Bank’s Code for Development initiative
aims at improving access to applications, algorithms and data among other tools.
As for crowd-sourced data, one example is OpenStreetMap, often used for mapping
wealth at a granular level such as in Soman et al. (2020) (see others in Table 9.1 of
the Electronic Online Supplement).
Finally, the combination of software development and IT operations to provide
continuous delivery and deployment of trained models might be a perhaps still
underexploited toolkit in PID studies. That is, instead of training the model just
once, re-train it on a frequent basis, automatically with new data. There are three
distinct cases of model updating: periodical training and offline predictions, periodical
training and online (on-demand) predictions, and online training and predictions
(fine-tuning of model predictions while receiving new data as feedback). The second
and third cases are more common for applications that interact with users and thus
need real-time predictions. Instead, the first case could be sometimes applicable to
PID studies, as PID indicators predictions might require frequent updating for two
main reasons: adding more data points might improve the model’s performance, and
keeping the model updated with world changes, in order to deliver relevant predictions.
There are different tools that facilitate model re-training commonly used in the ML
industry. For example, Apache Airflow for data orchestration, that is, an automated
process that can take data from multiple locations and manages the workflow of
processing it to obtain the desired features, training a ML model, uploading it to
some application or platform, re-running it; MLflow to keep track of experiments

and model versions; Great Expectations to test the new data in order to ensure some
basic data quality standards. One example of application is the Government of the
City of San Diego in the United States. They started using Apache Airflow for the
project StreetsSD, a map that provides an up-to-date status of paving projects. The
coordination of tasks also allows the Government to send automatic notifications
based on metrics and thresholds, and generate automated reports to share with
different Government departments.
This chapter provides a hopefully complete ‘ecosystem’ of the fast growing literature
on the use of ML methods for PID studies. The suggested taxonomy classifies
relevant contributions into two main categories: (1) studies aimed at providing better
measurements and forecasts of PID indicators, (2) studies using ML methods and new
data sources for answering causal questions. In general, ML methods have proved to
be effective tools for predictions and pattern recognition; consequently, studies in
the first category are more abundant. Instead, causal analysis requires to go beyond
the point estimates or predictions that dominate the practice of ML, demanding
inferential tools that help measure uncertainty and evaluate relevant hypothesis for
policy implementation, and are, quite expectedly, of a more complex nature, which
explains their relative scarcity compared to studies in the first group.
Regarding contributions in measurements and forecasts of PID indicators, ML can
improve them in three ways.
First, many studies combine different data sources, especially non-traditional
ones, and rely on the flexibility of ML methods to improve the prediction of a PID
outcome in terms of its time frequency, granularity or coverage. Granularity has
been the most extensive contribution of ML to PID studies: Figure 9.2 shows that
72 of the reviewed studies contribute in terms of granularity, whereas 29 in terms
of frequency. In relation to non-traditional data, Table 9.1 in the Electronic Online
Supplement shows that satellite images are the main source of data in PID research
with ML. Finally, in terms of specific ML techniques, we find that tree-based methods
are the most popular (Figure 9.3). Very likely, their popularity is related to their
intuitive nature that facilitates communication with non-technical actors (policy
makers, practitioners) and to their flexibility, that help discover non-linearities that
are difficult, when not impossible, to handle ex-ante with the mostly parametric
nature of standard econometric tools. Quite interestingly, OLS and its variants is
the method most frequently used as a benchmark to compare results. In turn, deep
learning methods dominates the use of satellite images (Figure 9.4).
Another important use of ML in PID studies is for data dimensionality, to help
understand factors behind multidimensional notions like poverty or inequality to
design shorter and less costly surveys, to construct indexes, and to find latent
socioeconomic groups among individuals. Finally, ML is a promising alternative do
deal with classic data problems, such as handling missing data or compensating for
the lack of panel data.
Fig. 9.2: Contributions of ML to PID studies in measurement and forecasting. Source:

reviewed articles, described in Table 9.1 in the Electronic Online Supplement.
In turn, there are two main contributions of ML to causal inference in PID studies.
The first one involves adapting ML techniques for causal inference, while the second
one takes advantage of the increased availability of new data sources for impact
evaluation analysis. The first line of work is still mainly theoretical and has contributed
with many promising methods for improving existent causal inference techniques,
such as methods for estimating heterogeneous treatment effects or handling high
dimensional data, among others. As expected, relatively few studies have applied these
methods to address causal inference questions in PID studies, although the number is
growing fast. Instead, the number of studies that use ML to take advantage of new data
sources for causal analysis is substantially larger, to build new outcome variables in
impact evaluation studies, to look for alternative sources of exogenous variability, or
to optimally combine observational and experimental data. Most commonly, the new
data are used for building or improving the outcome variable, and satellite images
are the most frequent source used for this purpose. For example, there is evidence
that it is possible to assess the impact of anti-poverty RCTs on household welfare
relying solely on satellite data, without the need of conducting baseline and follow
up surveys. In addition, the reviewed literature suggests that in empirical PID studies
involving causal inference, ML methods mainly intervene in an intermediate step for
data processing, although in a few cases they also play a role in estimating the causal
effect.
Fig. 9.3: ML methods for PID studies improving measurement and forecasting. Source:
reviewed articles, described in Table 9.1 in the Electronic Online Supplement.
Fig. 9.4: ML methods in PID studies improving measurement and forecasting, by

data source. Source: reviewed articles, described in Table 9.1 in the Electronic Online
Supplement.
References 325
Several clear patterns emerge from the review. First, new data sources (such as
mobile phone calls data or satellite images) do not replace traditional data (i.e.,
surveys and census), but rather complement them. Indeed, traditional data sets are a
key part of the process. Many PID studies use traditional data as the ground truth
to train a model that predicts a PID outcome based on features built from the new
data, and are hence necessary to evaluate the perfomance of ML methods applied
to alternative sources. They are an essential benchmark as they are usually less
biased and more representative of the population. Traditional data sources are also
needed in studies that use dimensionality reduction techniques to infer the structure
of the underlying reality. In turn, in causal inference studies the new data sources
complement traditional data to improve the source of exogeneity or to build the
outcome or control variables, allowing for improved estimates.
Finally, many –if not all– of the contributions described throughout this chapter
would not have been possible without the improvement in computational power.
One of the most important advances is deep learning, which is extremely useful for
processing satellite images, likely the main alternative data source.
The literature using ML in PID studies is growing fast, and most reviewed studies
date from 2015 onwards. However, the framework we provide is hopefully useful to
locate each new study in terms of its contribution relative to traditional econometric
techniques, and traditional data sources.
Acknowledgements We thank the members of the Centro de Estudios para el Desarrollo Humano
(CEDH) at Universidad de San Andrés for extensive discussions. Joaquín Torré provided useful
insights for the section on computing power. Mariana Santi provided excellent research assistance.
We specially thank the editors, Felix Chan and László Mátyás, for their comments on an earlier
version, that helped improve this chapter considerably. All errors and omissions are our responsibility.
References
Afzal, M., Hersh, J. & Newhouse, D. (2015). Building a better model: Variable
selection to predict poverty in Pakistan and Sri Lanka. World Bank Research
Working Paper.
Agarwal, D., Agastya, A., Chaudhury, M., Dube, T., Jha, B., Khare, P. & Raghu, N.
(2021). Measuring effectiveness of chatbot to improve attitudes towards gender
issues in underserved adolescent children in India (Tech. Rep.). Cambridge,
MA: Harvard Kennedy School.
Ahrens, A., Aitken, C. & Schaffer, M. E. (2021). Using machine learning methods to
support causal inference in econometrics. In Behavioral predictive modeling
in economics (pp. 23–52). Springer.
Aiken, E., Bedoya, G., Coville, A. & Blumenstock, J. E. (2020). Targeting development
aid with machine learning and mobile phone data: Evidence from an Anti-
poverty intervention in Afghanistan. In Proceedings of the 3rd acm sigcas
conference on computing and sustainable societies (pp. 310–311).
Aiken, E., Bellue, S., Karlan, D., Udry, C. R. & Blumenstock, J. (2021). Machine
learning and mobile phone data can improve the targeting of humanitarian
assistance (Tech. Rep.). Cambridge, MA: National Bureau of Economic
Research.
Alix-Garcia, J., McIntosh, C., Sims, K. R. & Welch, J. R. (2013). The ecological foot-
print of poverty alleviation: Evidence from Mexico’s Oportunidades program.
Review of Economics and Statistics, 95(2), 417–435.
Andrews, I., Kitagawa, T. & McCloskey, A. (2019). Inference on winners (Tech.
Rep.). Cambridge, MA: National Bureau of Economic Research.
Angrist, J. D. & Pischke, J.-S. (2010). The credibility revolution in empirical
economics: How better research design is taking the con out of econometrics.
Journal of economic perspectives, 24(2), 3–30.
Antenucci, D., Cafarella, M., Levenstein, M., Ré, C. & Shapiro, M. D. (2014). Using
social media to measure labor market flows (Tech. Rep.). Cambridge, MA:
National Bureau of Economic Research.
Askitas, N. & Zimmermann, K. F. (2009). Google econometrics and unemployment
forecasting. Applied Economics Quarterly.
Athey, S., Bayati, M., Doudchenko, N., Imbens, G. & Khosravi, K. (2021). Matrix
completion methods for causal panel data models. Journal of the American
Statistical Association, 1–15.
Athey, S., Chetty, R. & Imbens, G. (2020). Combining experimental and observational
data to estimate treatment effects on long term outcomes. arXiv preprint
arXiv:2006.09676.
Athey, S., Imbens, G. W. & Wager, S. (2018). Approximate residual balancing:
Debiased inference of average treatment effects in high dimensions. Journal
of the Royal Statistical Society: Series B (Statistical Methodology), 80(4),
597–623.
Athey, S., Tibshirani, J. & Wager, S. (2019). Generalized random forests. The Annals
of Statistics, 47(2), 1148–1178.
Athey, S. & Wager, S. (2021). Policy learning with observational data. Econometrica,
89(1), 133–161.
Banerjee, A., Chandrasekhar, A. G., Dalpath, S., Duflo, E., Floretta, J., Jackson,
M. O., . . . Shrestha, M. (2021b). Selecting the most effective nudge: Evidence
from a large-scale experiment on immunization (Tech. Rep.). Cambridge, MA:
National Bureau of Economic Research.
Banerjee, A., Hanna, R., Olken, B. A., Satriawan, E. & Sumarto, S. (2021a). Food vs.
food stamps: Evidence from an at-scale experiment in Indonesia (Tech. Rep.).
Cambridge, MA: National Bureau of Economic Research.
Baylé, F. (2016). Detección de villas y asentamientos informales en el partido
de La Matanza mediante teledetección y sistemas de información geográfica
(Unpublished doctoral dissertation). Universidad de Buenos Aires. Facultad de
Ciencias Exactas y Naturales.
References 327
Bedi, T., Coudouel, A. & Simler, K. (2007a). More than a pretty picture: Using poverty
maps to design better policies and interventions. World Bank Publications.
Bedi, T., Coudouel, A. & Simler, K. (2007b). Maps for policy making: Beyond the
obvious targeting applications. In More than a pretty picture: Using poverty
maps to design better policies and interventions (pp. 3–22). World Bank
Publications.
Belloni, A., Chen, D., Chernozhukov, V. & Hansen, C. (2012). Sparse models
and methods for optimal instruments with an application to eminent domain.
Econometrica, 80(6), 2369–2429.
Belloni, A. & Chernozhukov, V. (2013). Least squares after model selection in
high-dimensional sparse models. Bernoulli, 19(2), 521–547.
Belloni, A., Chernozhukov, V., Fernández-Val, I. & Hansen, C. (2017). Program
evaluation and causal inference with high-dimensional data. Econometrica,
85(1), 233–298.
Belloni, A., Chernozhukov, V. & Hansen, C. (2014a). High-dimensional methods and
inference on structural and treatment effects. Journal of Economic Perspectives,
28(2), 29–50.
Studies, 81(2), 608–650.
Bennett, M. M. & Smith, L. C. (2017). Advances in using multitemporal night-
time lights satellite imagery to detect, estimate, and monitor socioeconomic
dynamics. Remote Sensing of Environment, 192, 176–197.
Bertrand, M., Crépon, B., Marguerie, A. & Premand, P. (2017). Contemporaneous and
post-program impacts of a Public Works Program (Tech. Rep.). Washington,
DC: World Bank.
Björkegren, D., Blumenstock, J. E. & Knight, S. (2020). Manipulation-proof machine
learning. arXiv preprint arXiv:2004.03865.
Blumenstock, J. (2018a). Don’t forget people in the use of big data for development.
Nature Publishing Group.
Blumenstock, J. (2018b). Estimating economic characteristics with phone data. In
Aea papers and proceedings (Vol. 108, pp. 72–76).
Blumenstock, J., Cadamuro, G. & On, R. (2015). Predicting poverty and wealth from
mobile phone metadata. Science, 350(6264), 1073–1076.
Bosco, C., Alegana, V., Bird, T., Pezzulo, C., Bengtsson, L., Sorichetta, A., . . . Tatem,
A. J. (2017). Exploring the high-resolution mapping of gender-disaggregated
development indicators. Journal of The Royal Society Interface, 14(129),
20160825.
Bunte, J. B., Desai, H., Gbala, K., Parks, B. & Runfola, D. M. (2017). Natural
resource sector fdi and growth in post-conflict settings: Subnational evidence
from Liberia. Williamsburg (VA): AidData.
Burke, J., Jamison, J., Karlan, D., Mihaly, K. & Zinman, J. (2019). Credit building or
credit crumbling? A credit builder loan’s effects on consumer behavior, credit
scores and their predictive power (Tech. Rep.). Cambridge, MA8: National
Bureau of Economic Research.
Burke, M., Heft-Neal, S. & Bendavid, E. (2016). Sources of variation in under-5

mortality across Sub-Saharan Africa: A spatial analysis. The Lancet Global
Health, 4(12), e936–e945.
Burstein, R., Cameron, E., Casey, D. C., Deshpande, A., Fullman, N., Gething, P. W.,
. . . Hay, S. I. (2018). Mapping child growth failure in Africa between 2000
and 2015. Nature, 555(7694), 41–47.
Carol, S., Eich, D., Keller, M., Steiner, F. & Storz, K. (2019). Who can ride along?
Discrimination in a German carpooling market. Population, space and place,
25(8), e2249.
Carter, M. R., Tjernström, E. & Toledo, P. (2019). Heterogeneous impact dynamics of
a rural business development program in Nicaragua. Journal of Development
Economics, 138, 77–98.
Caruso, G., Sosa-Escudero, W. & Svarc, M. (2015). Deprivation and the dimen-
sionality of welfare: A variable-selection cluster-analysis approach. Review of
Income and Wealth, 61(4), 702–722.
Chen, X. & Nordhaus, W. (2015). A test of the new viirs lights data set: Population
and economic output in Africa. Remote Sensing, 7(4), 4937–4947.
& Robins, J. (2018b). Double/debiased machine learning for treatment and
structural parameters. Oxford University Press Oxford, UK.
Chernozhukov, V., Demirer, M., Duflo, E. & Fernandez-Val, I. (2018a). Generic
machine learning inference on heterogeneous treatment effects in random-
ized experiments, with an application to immunization in India (Tech. Rep.).
Washington: National Bureau of Economic Research.
Chetty, R., Friedman, J. N., Hendren, N., Jones, M. R. & Porter, S. R. (2018). The
Opportunity Atlas: Mapping the childhood roots of social mobility (Tech. Rep.).
Chetty, R., Friedman, J. N., Hendren, N., Stepner, M. & Team, T. O. I. (2020a). How
did covid-19 and stabilization policies affect spending and employment? A new
real-time economic tracker based on private sector data. National Bureau of
Economic Research Cambridge, MA.
Chetty, R., Hendren, N., Jones, M. R. & Porter, S. R. (2020). Race and economic
opportunity in the united states: An intergenerational perspective. The Quarterly
Journal of Economics, 135(2), 711–783.
Chetty, R., Hendren, N., Kline, P. & Saez, E. (2014). Where is the land of opportunity?
The geography of intergenerational mobility in the united states. The Quarterly
Journal of Economics, 129(4), 1553–1623.
Chi, G., Fang, H., Chatterjee, S. & Blumenstock, J. E. (2022). Microestimates of
wealth for all low-and middle-income countries. Proceedings of the National
Academy of Sciences, 119(3).
Chioda, L., De Mello, J. M. & Soares, R. R. (2016). Spillovers from conditional
cash transfer programs: Bolsa Família and crime in urban Brazil. Economics
of Education Review, 54, 306–320.
Chowdhury, R., Ceballos-Sierra, F. & Sulaiman, M. (2021). Grow the pie or have it?
using machine learning for impact heterogeneity in the Ultra-poor Graduation
References 329
Model (Tech. Rep. No. WPS - 170). Berkeley, CA: Center for Effective Global
Action, University of California, Berkeley.
Christensen, P., Sarmiento-Barbieri, I. & Timmins, C. (2020). Housing discrimination
and the toxics exposure gap in the United States: Evidence from the rental
market. The Review of Economics and Statistics, 1–37.
Christensen, P., Sarmiento-Barbieri, I. & Timmins, C. (2021). Racial discrimina-
tion and housing outcomes in the United States rental market (Tech. Rep.).
Christiansen, T. & Weeks, M. (2020). Distributional aspects of microcredit expansions
(Tech. Rep.). Cambridge, England: Faculty of Economics, University of
Cambridge.
Churchill, B. F. & Sabia, J. J. (2019). The effects of minimum wages on low-skilled
immigrants’ wages, employment, and poverty. Industrial Relations: A Journal
of Economy and Society, 58(2), 275–314.
Clay, K., Egedesø, P. J., Hansen, C. W., Jensen, P. S. & Calkins, A. (2020).
Controlling tuberculosis? Evidence from the first community-wide health
experiment. Journal of Development Economics, 146, 102510.
Dahmani, R., Fora, A. A. & Sbihi, A. (2014). Extracting slums from high-resolution
satellite images. Int. J. Eng. Res. Dev, 10, 1–10.
Dang, H.-A., Lanjouw, P., Luoto, J. & McKenzie, D. (2014). Using repeated cross-
sections to explore movements into and out of poverty. Journal of Development
Economics, 107, 112–128.
Daoud, A. & Johansson, F. (2019). Estimating treatment heterogeneity of International
Monetary Fund programs on child poverty with generalized random forest.
Davis, J. & Heller, S. (2017). Using causal forests to predict treatment heterogeneity:
An application to summer jobs. American Economic Review, 107(5), 546–50.
Davis, J. & Heller, S. (2020). Rethinking the benefits of youth employment programs:
The heterogeneous effects of summer jobs. Review of Economics and Statistics,
102(4), 664–677.
Deryugina, T., Heutel, G., Miller, N. H., Molitor, D. & Reif, J. (2019). The mortality
and medical costs of air pollution: Evidence from changes in wind direction.
American Economic Review, 109(12), 4178–4219.
Donaldson, D. & Storeygard, A. (2016). The view from above: Applications of
satellite data in economics. Journal of Economic Perspectives, 30(4), 171–98.
Doudchenko, N. & Imbens, G. W. (2016). Balancing, regression, difference-in-
differences and synthetic control methods: A synthesis (Tech. Rep.). Cambridge,
MA: National Bureau of Economic Research.
Dullerud, N., Zhang, H., Seyyed-Kalantari, L., Morris, Q., Joshi, S. & Ghassemi,
M. (2021). An empirical framework for domain generalization in clinical
settings. In Proceedings of the conference on health, inference, and learning
(pp. 279–290).
Dutt, P. & Tsetlin, I. (2021). Income distribution and economic development: Insights
from machine learning. Economics & Politics, 33(1), 1–36.
Eagle, N., Macy, M. & Claxton, R. (2010). Network diversity and economic
development. Science, 328(5981), 1029–1031.
Edo, M., Escudero, W. S. & Svarc, M. (2021). A multidimensional approach

to measuring the middle class. The Journal of Economic Inequality, 19(1),
139–162.
Elbers, C., Fujii, T., Lanjouw, P., Özler, B. & Yin, W. (2007). Poverty alleviation
through geographic targeting: How much does disaggregation help? Journal
of Development Economics, 83(1), 198–213.
Elbers, C., Lanjouw, J. O. & Lanjouw, P. (2003). Micro-level estimation of poverty
and inequality. Econometrica, 71(1), 355–364.
Electronic Online Supplement. (2022). Electronic Online Supplement of this Volume.
https://sn.pub/0ObVSo.
Elvidge, C. D., Sutton, P. C., Ghosh, T., Tuttle, B. T., Baugh, K. E., Bhaduri, B. &
Bright, E. (2009). A global poverty map derived from satellite data. Computers
& Geosciences, 35(8), 1652–1660.
Engstrom, R., Hersh, J. S. & Newhouse, D. L. (2017). Poverty from space: Using
high-resolution satellite imagery for estimating economic well-being. World
Bank Policy Research Working Paper(8284).
Ettredge, M., Gerdes, J. & Karuga, G. (2005). Using web-based search data to predict
macroeconomic statistics. Communications of the ACM, 48(11), 87–92.
Faber, B. & Gaubert, C. (2019). Tourism and economic development: Evidence from
Mexico’s coastline. American Economic Review, 109(6), 2245–93.
Farbmacher, H., Kögel, H. & Spindler, M. (2021). Heterogeneous effects of poverty
on attention. Labour Economics, 71, 102028.
Farrell, D., Greig, F. & Deadman, E. (2020). Estimating family income from
administrative banking data: A machine learning approach. In Aea papers and
proceedings (Vol. 110, pp. 36–41).
Feigenbaum, J. J. (2016). A machine learning approach to census record linking.
Retrieved March, 28, 2016.
Fraiman, R., Justel, A. & Svarc, M. (2008). Selection of variables for cluster analysis
and classification rules. Journal of the American Statistical Association,
103(483), 1294–1303.
Frias-Martinez, V. & Virseda, J. (2012). On the relationship between socio-economic
factors and cell phone usage. In Proceedings of the fifth international conference
on information and communication technologies and development (pp. 76–84).
Gasparini, L., Sosa-Escudero, W., Marchionni, M. & Olivieri, S. (2013). Multidi-
mensional poverty in Latin America and the Caribbean: New evidence from
the Gallup World Poll. The Journal of Economic Inequality, 11(2), 195–214.
Gevaert, C., Persello, C., Sliuzas, R. & Vosselman, G. (2016). Classification of
informal settlements through the integration of 2d and 3d features extracted
from UAV data. ISPRS annals of the photogrammetry, remote sensing and
spatial information sciences, 3, 317.
Glaeser, E. L., Kominers, S. D., Luca, M. & Naik, N. (2018). Big data and big cities:
The promises and limitations of improved measures of urban life. Economic
Inquiry, 56(1), 114–137.
González-Fernández, M. & González-Velasco, C. (2018). Can google econometrics
predict unemployment? Evidence from Spain. Economics Letters, 170, 42–45.
References 331
Goodfellow, I., Bengio, Y. & Courville, A. (2016). Deep Learning. MIT Press.
(http://www.deeplearningbook.org)
Graesser, J., Cheriyadat, A., Vatsavai, R. R., Chandola, V., Long, J. & Bright, E.
(2012). Image based characterization of formal and informal neighborhoods
in an urban landscape. IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, 5(4), 1164–1176.
Graetz, N., Friedman, J., Friedman, J., Osgood-Zimmerman, A., Burstein, R., Biehl,
C., Molly H.and Shields, . . . Hay, S. I. (2018). Mapping local variation in
educational attainment across Africa. Nature, 555(7694), 48–53.
Gulrajani, I. & Lopez-Paz, D. (2020). In search of lost domain generalization. arXiv
Hastie, T., Tibshirani, R., Friedman, J. H. & Friedman, J. H. (2009). The elements of
statistical learning: data mining, inference, and prediction (Vol. 2). Springer.
Head, A., Manguin, M., Tran, N. & Blumenstock, J. E. (2017). Can human
development be measured with satellite imagery? In Ictd (pp. 8–1).
Henderson, J. V., Storeygard, A. & Weil, D. N. (2012). Measuring economic growth
from outer space. American economic review, 102(2), 994–1028.
Heß, S., Jaimovich, D. & Schündeln, M. (2021). Development projects and economic
networks: Lessons from rural Gambia. The Review of Economic Studies, 88(3),
1347–1384.
Hodler, R. & Raschky, P. A. (2014). Regional favoritism. The Quarterly Journal of
Economics, 129(2), 995–1033.
Hristova, D., Rutherford, A., Anson, J., Luengo-Oroz, M. & Mascolo, C. (2016).
The international postal network and other global flows as proxies for national
wellbeing. PloS one, 11(6), e0155976.
Huang, L. Y., Hsiang, S. & Gonzalez-Navarro, M. (2021). Using satellite imagery
and deep learning to evaluate the impact of anti-poverty programs. arXiv
Huang, X., Liu, H. & Zhang, L. (2015). Spatiotemporal detection and analysis of
urban villages in mega city regions of China using high-resolution remotely
sensed imagery. IEEE Transactions on Geoscience and Remote Sensing, 53(7),
3639–3657.
ILO-ECLAC. (2018). Child labour risk identification model. Methodology to design
preventive strategies at local level. https://dds.cepal.org/redesoc/publicacion
?id=4886.
Jean, N., Burke, M., Xie, M., Davis, W. M., Lobell, D. B. & Ermon, S. (2016).
Combining satellite imagery and machine learning to predict poverty. Science,
353(6301), 790–794.
Jurafsky, D. & Martin, J. H. (2014). Speech and language processing. US: Prentice
Hall.
Kavanagh, L., Lee, D. & Pryce, G. (2016). Is poverty decentralizing? quantifying
uncertainty in the decentralization of urban poverty. Annals of the American
Association of Geographers, 106(6), 1286–1298.
Khelifa, D. & Mimoun, M. (2012). Object-based image analysis and data mining
for building ontology of informal urban settlements. In Image and signal
processing for remote sensing xviii (Vol. 8537, p. 85371I).

Kim, S., Kim, H., Kim, B., Kim, K. & Kim, J. (2019). Learning not to learn:
Training deep neural networks with biased data. In Proceedings of the ieee/cvf
conference on computer vision and pattern recognition (pp. 9012–9020).
Kim, S. & Koh, K. (2022). Health insurance and subjective well-being: Evidence
from two healthcare reforms in the United States. Health Economics, 31(1),
233–249.
Kitagawa, T. & Tetenov, A. (2018). Who should be treated? empirical welfare
maximization methods for treatment choice. Econometrica, 86(2), 591–616.
Knaus, M. C., Lechner, M. & Strittmatter, A. (2020). Heterogeneous employment
effects of job search programmes: A machine learning approach. Journal of
Human Resources, 0718–9615R1.
Kohli, D., Sliuzas, R., Kerle, N. & Stein, A. (2012). An ontology of slums for
image-based classification. Computers, Environment and Urban Systems, 36(2),
154–163.
Kohli, D., Sliuzas, R. & Stein, A. (2016). Urban slum detection using texture and
spatial metrics derived from satellite imagery. Journal of spatial science, 61(2),
405–426.
Kuffer, M., Pfeffer, K. & Sliuzas, R. (2016). Slums from space—15 years of slum
mapping using remote sensing. Remote Sensing, 8(6), 455.
Lansley, G. & Longley, P. A. (2016). The geography of Twitter topics in London.
Computers, Environment and Urban Systems, 58, 85–96.
Lazer, D. M., Pentland, A., Watts, D. J., Aral, S., Athey, S., Contractor, N., . . .
Wagner, C. (2020). Computational social science: Obstacles and opportunities.
Science, 369(6507), 1060–1062.
Liu, J.-H., Wang, J., Shao, J. & Zhou, T. (2016). Online social activity reflects
economic status. Physica A: Statistical Mechanics and its Applications, 457,
581–589.
Llorente, A., Garcia-Herranz, M., Cebrian, M. & Moro, E. (2015). Social media
fingerprints of unemployment. PloS one, 10(5), e0128692.
Lucchetti, L. (2018). What can we (machine) learn about welfare dynamics from
cross-sectional data? World Bank Policy Research Working Paper(8545).
Lucchetti, L., Corral, P., Ham, A. & Garriga, S. (2018). Lassoing welfare dynamics
with cross-sectional data (Tech. Rep. No. 8545). Washington, DC: World
Bank.
Luzzi, G. F., Flückiger, Y. & Weber, S. (2008). A cluster analysis of multidimensional
poverty in Switzerland. In Quantitative approaches to multidimensional poverty
measurement (pp. 63–79). Springer.
Maeda, E., Miyata, A., Boivin, J., Nomura, K., Kumazawa, Y., Shirasawa, H., . . .
Terada, Y. (2020). Promoting fertility awareness and preconception health
using a chatbot: A randomized controlled trial. Reproductive BioMedicine
Online, 41(6), 1133–1143.
Mahabir, R., Croitoru, A., Crooks, A. T., Agouris, P. & Stefanidis, A. (2018). A
critical review of high and very high-resolution remote sensing approaches for
detecting and mapping slums: Trends, challenges and emerging opportunities.
References 333
Urban Science, 2(1), 8.

Maiya, S. R. & Babu, S. C. (2018). Slum segmentation and change detection: A deep
learning approach. arXiv preprint arXiv:1811.07896.
Martey, E. & Armah, R. (2020). Welfare effect of international migration on the
left-behind in Ghana: Evidence from machine learning. Migration Studies.
McBride, L. & Nichols, A. (2018). Retooling poverty targeting using out-of-sample
validation and machine learning. The World Bank Economic Review, 32(3),
531–550.
Merola, G. M. & Baulch, B. (2019). Using sparse categorical principal components
to estimate asset indices: New methods with an application to rural Southeast
Asia. Review of Development Economics, 23(2), 640–662.
Michalopoulos, S. & Papaioannou, E. (2014). National institutions and subnational
development in Africa. The Quarterly journal of economics, 129(1), 151–213.
Mohamud, J. H. & Gerek, O. N. (2019). Poverty level characterization via fea-
ture selection and machine learning. In 2019 27th signal processing and
communications applications conference (siu) (pp. 1–4).
Mullally, C., Rivas, M. & McArthur, T. (2021). Using machine learning to estimate the
heterogeneous effects of livestock transfers. American Journal of Agricultural
Economics, 103(3), 1058–1081.
Nie, X., Brunskill, E. & Wager, S. (2021). Learning when-to-treat policies. Journal
of the American Statistical Association, 116(533), 392–409.
Njuguna, C. & McSharry, P. (2017). Constructing spatiotemporal poverty indices
from big data. Journal of Business Research, 70, 318–327.
Okiabera, J. O. (2020). Using random forest to identify key determinants of poverty
in kenya. (Unpublished doctoral dissertation). University of Nairobi.
Owen, K. K. & Wong, D. W. (2013). An approach to differentiate informal settlements
using spectral, texture, geomorphology and road accessibility metrics. Applied
Geography, 38, 107–118.
Pokhriyal, N. & Jacques, D. C. (2017). Combining disparate data sources for improved
poverty prediction and mapping. Proceedings of the National Academy of
Sciences, 114(46), E9783–E9792.
Quercia, D., Ellis, J., Capra, L. & Crowcroft, J. (2012). Tracking "gross community
happiness" from tweets. In Proceedings of the acm 2012 conference on
computer supported cooperative work (pp. 965–968).
Ratledge, N., Cadamuro, G., De la Cuesta, B., Stigler, M. & Burke, M. (2021). Using
satellite imagery and machine learning to estimate the livelihood impact of
electricity access (Tech. Rep.). Cambridge, MA: National Bureau of Economic
Research.
Robinson, T., Emwanu, T. & Rogers, D. (2007). Environmental approaches to poverty
mapping: An example from Uganda. Information development, 23(2-3),
205–215.
Rosati, G. (2017). Construcción de un modelo de imputación para variables de
ingreso con valores perdidos a partir de ensamble learning: Aplicación en la
encuesta permanente de hogares (EPH). SaberEs, 9(1), 91–111.
Rosati, G., Olego, T. A. & Vazquez Brust, H. A. (2020). Building a sanitary vulner-
ability map from open source data in Argentina (2010-2018). International

Journal for Equity in Health, 19(1), 1–16.
Schmitt, A., Sieg, T., Wurm, M. & Taubenböck, H. (2018). Investigation on the
separability of slums by multi-aspect Terrasar-x dual-co-polarized high resolu-
tion spotlight images based on the multi-scale evaluation of local distributions.
International journal of applied earth observation and geoinformation, 64,
181–198.
Sen, A. (1985). Commodities and Capabilities. Oxford University Press.
Sheehan, E., Meng, C., Tan, M., Uzkent, B., Jean, N., Burke, M., . . . Ermon, S.
(2019). Predicting economic development using geolocated wikipedia articles.
In Proceedings of the 25th acm sigkdd international conference on knowledge
discovery & data mining (pp. 2698–2706).
Skoufias, E. & Vinha, K. (2020). Child stature, maternal education, and early
childhood development (Tech. Rep. Nos. Policy Research Working Paper, No.
9396). Washington, DC: World Bank.
Sohnesen, T. P. & Stender, N. (2017). Is random forest a superior methodology for
predicting poverty? an empirical assessment. Poverty & Public Policy, 9(1),
118–133.
Soman, S., Beukes, A., Nederhood, C., Marchio, N. & Bettencourt, L. (2020). World-
wide detection of informal settlements via topological analysis of crowdsourced
digital maps. ISPRS International Journal of Geo-Information, 9(11), 685.
Soto, V., Frias-Martinez, V., Virseda, J. & Frias-Martinez, E. (2011). Prediction of
socioeconomic levels using cell phone records. In International conference on
user modeling, adaptation, and personalization (pp. 377–388).
Steele, J. E., Sundsøy, P. R., Pezzulo, C., Alegana, V. A., Bird, T. J., Blumenstock, J.,
. . . Bengtsson, L. (2017). Mapping poverty using mobile phone and satellite
data. Journal of The Royal Society Interface, 14(127), 20160690.
Strittmatter, A. (2019). Heterogeneous earnings effects of the job corps by gender: A
translated quantile approach. Labour Economics, 61, 101760.
Sutton, P. C., Elvidge, C. D. & Ghosh, T. (2007). Estimation of gross domestic
product at sub-national scales using nighttime satellite imagery. International
Journal of Ecological Economics & Statistics, 8(S07), 5–21.
Taubenböck, H., Kraff, N. J. & Wurm, M. (2018). The morphology of the Arrival
City, A global categorization based on literature surveys and remotely sensed
data. Applied Geography, 92, 150–167.
Thompson, N. C., Greenewald, K., Lee, K. & Manso, G. F. (2020). The computational
limits of deep learning. arXiv preprint arXiv:2007.05558.
Thoplan, R. (2014). Random forests for poverty classification. International Journal
of Sciences: Basic and Applied Research (IJSBAR), North America, 17.
UN Global. (2016). Building proxy indicators of national wellbeing with postal
data. Project Series, no. 22, https://www.unglobalpulse.org/document/building
-proxy-indicators-of-national-wellbeing-with-postal-data/.
Venerandi, A., Quattrone, G., Capra, L., Quercia, D. & Saez-Trumper, D. (2015).
Measuring urban deprivation from user generated content. In Proceedings of
the 18th acm conference on computer supported cooperative work & social
References 335
computing (pp. 254–264).

Villa, J. M. (2016). Social transfers and growth: Evidence from luminosity data.
Economic Development and Cultural Change, 65(1), 39–61.
113(523), 1228–1242.
Wager, S., Du, W., Taylor, J. & Tibshirani, R. J. (2016). High-dimensional regression
adjustments in randomized experiments. Proceedings of the National Academy
of Sciences, 113(45), 12673–12678.
Wald, Y., Feder, A., Greenfeld, D. & Shalit, U. (2021). On calibration and out-of-
domain generalization. arXiv preprint arXiv:2102.10395.
Warr, P. & Aung, L. L. (2019). Poverty and inequality impact of a natural disaster:
Myanmar’s 2008 cyclone Nargis. World Development, 122, 446–461.
Watmough, G. R., Marcinko, C. L., Sullivan, C., Tschirhart, K., Mutuo, P. K., Palm,
C. A. & Svenning, J.-C. (2019). Socioecologically informed use of remote
sensing data to predict rural household poverty. Proceedings of the National
Academy of Sciences, 116(4), 1213–1218.
Wurm, M., Taubenböck, H., Weigand, M. & Schmitt, A. (2017). Slum mapping in
polarimetric SAR data using spatial features. Remote sensing of environment,
194, 190–204.
Yeh, C., Perez, A., Driscoll, A., Azzari, G., Tang, Z., Lobell, D., . . . Burke, M. (2020).
Using publicly available satellite imagery and deep learning to understand
economic well-being in Africa. Nature communications, 11(1), 1–11.
Zhou, Z., Athey, S. & Wager, S. (2018). Offline multi-action policy learning:
Generalization and optimization. arXiv preprint arXiv:1810.04778.
Zhuo, J.-Y. & Tan, Z.-M. (2021). Physics-augmented deep learning to improve
tropical cyclone intensity and size estimation from satellite imagery. Monthly
Weather Review, 149(7), 2097–2113.
Chapter 10
Machine Learning for Asset Pricing
Jantje Sönksen
Abstract This chapter reviews the growing literature that describes machine learning
applications in the field of asset pricing. In doing so, it focuses on the additional
benefits that machine learning – in addition to, or in combination with, standard
econometric approaches – can bring to the table. This issue is of particular importance
because in recent years, improved data availability and increased computational
facilities have had huge effects on finance literature. For example, machine learning
techniques inform analyses of conditional factor models; they have been applied
to identify the stochastic discount factor and purposefully to test and evaluate
existing asset pricing models. Beyond those pertinent applications, machine learning
techniques also lend themselves to prediction problems in the domain of empirical
asset pricing.
10.1 Introduction
Research in the domain of empirical asset pricing traditionally has embraced the
fruitful interaction of finance and the development of econometric/statistical models,
pushing the envelope of both economic theory and empirical methodology. This
effort is personified by Lars P. Hansen, who received the Nobel Prize in Economics
for his contributions to asset pricing, and who also developed empirical methods that
form the pillars of modern econometric analysis: the generalized method of moments
(GMM, Hansen, 1982), along with its variant, the simulated method of moments
(SMM, Duffie & Singleton, 1993). Both GMM and SMM are particularly useful for
empirical asset pricing and closely connected to it, but these methods also have many
more applications (Hall, 2005).
Asset pricing is concerned with answering the question of why some assets pay
higher average returns than others. Two influential monographs, by Cochrane (2005)
Jantje Sönksen B
Eberhard Karls University, Tübingen, Germany, e-mail: jantje.soenksen@uni-tuebingen.de
338 Sönksen
and Singleton (2006), provide comprehensive synopses of the state of empirical asset
pricing research in the mid-2000s, in which they emphasize the close interactions
of asset pricing theory/financial economics, econometric modeling, and method
development. Notably, neither book offers a discussion of machine learning methods.
Evidently, even though machine learning models have been employed in finance
research since the 1990s – and particularly in efforts to forecast financial time series –
their connection with theory-based asset pricing had not yet been adequately worked
out at that time. Applications of machine learning methods appeared more like
transfers, such that researchers leveraged models that had proven successful for
forecasting exercises in environments with more favorable signal-to-noise ratios
(e.g., artificial neural networks) to financial time series data. Because they did not
seek to establish clear connections with financial economic theory, early adoptions
of machine learning in finance may well be characterized as measurement without
theory.
In turn, machine learning methods did not become part of the toolkit used in theory-
based empirical finance/asset pricing. However, in more recent years, a literature
sparked that consistently connects theory-based empirical asset pricing with machine
learning techniques, resulting in contributions, published in leading finance journals,
that substantially augment the standard econometric toolbox available for empirical
finance.
Superficially, this surge of machine learning applications for asset pricing might
seem due mostly to increased data availability (i.e., big data in finance), together with
increasingly powerful computational resources. But with this chapter, I identify and
argue for deeper reasons. The recent success of machine learning stems from its
unique ability to help address and alleviate some long-standing challenges associated
with empirical asset pricing, which standard econometric methods have had
problems dealing with. Therefore, with this chapter, I outline which aspects of
empirical asset pricing benefit particularly from machine learning, then detail how
machine learning can complement standard econometric approaches.1
Identification of the Stochastic Discount Factor and Machine Learning

I adopt an approach that mimics the one pursued by Cochrane (2005) and Singleton
(2006). Both authors take the basic asset pricing equation of financial economics as a
theoretical starting point for their discussion of empirical methodologies. Accordingly,
I consider the following version of the basic asset pricing equation:
𝑖
𝑃𝑡𝑖 = E𝑡 [𝑋𝑡+1 𝑚 𝑡+1 (𝜹)], (10.1)
where E𝑡 denotes the expected value conditional on time 𝑡 information, and is the 𝑃𝑡𝑖
𝑖 offered by asset 𝑖, which is the sum of the price of the
price of a future payoff 𝑋𝑡+1
asset and cash payments in 𝑡 + 1.2 Next, 𝑚 𝑡+1 is the stochastic discount factor (SDF)
1 Surveys on the question how machine learning techniques can be applied in the field of empirical
asset pricing are also provided by Weigand (2019), Giglio, Kelly and Xiu (2022), and Nagel (2021).
𝑖 amounts to 𝑃 𝑖 + 𝐷 𝑖 , where 𝐷 𝑖
2 For stocks, 𝑋𝑡+1 𝑡+1 𝑡+1 𝑡+1 denotes the dividend in 𝑡 + 1. Chapter 1 in
Cochrane (2005) provides a useful introduction to standard asset pricing notation and concepts.
10 Machine Learning for Asset Pricing 339
also referred to as the pricing kernel. The SDF is the central element in empirical
asset pricing, because the same discount factor must be able to price all assets of a
certain class (e.g., there is an 𝑚 𝑡+1 that fulfills Equation (10.1) for all stocks and their
individual payoffs 𝑋𝑡+1 𝑖 ). In preference-based asset pricing, the SDF represents the
marginal rate of substitution between consumption in different periods, and the vector
𝜹 contains parameters associated with investors’ risk aversion or time preference.
Imposing less economic structure, the fundamental theorem of financial economics
states that in the absence of arbitrage, a positive SDF exists.
In traditional empirical asset pricing, one finds stylized structural models that
imply a specific SDF, possibly nonlinear, the parameters of which can be estimated
with either GMM or SMM. The power utility SDF used by Hansen and Singleton
(1982) is a canonical example. Ad hoc specifications of the SDF also appear in the
empirical asset pricing literature, for which Cochrane (2005) emphasizes the need to
remain mindful of the economic meaning of the SDF associated with investor utility.
Because they can provide a negotiating balance between these opposites –
theory-based versus ad hoc specifications – machine learning methods have proven
useful in identifying and recovering the SDF. I review and explain these types of
contributions in Section 10.2.
Selection of Test Assets and Machine Learning

Two types of payoffs are most important for empirical asset pricing. The first is the
𝑖 = 𝑋 𝑖 /𝑃 𝑖 , for which Equation 10.1 becomes:
gross return of asset 𝑖, 𝑅𝑡+1 𝑡+1 𝑡
𝑖
E𝑡 [𝑅𝑡+1 𝑚 𝑡+1 (𝜹)] = 1. (10.2)
The price of any gross return thus is 1. Considering a riskless payoff, Equation (10.2)
yields an expression for the risk-free rate, given by3
𝑓 1
𝑅𝑡+1 = .
E𝑡 [𝑚 𝑡+1 (𝜹)]
The second type of payoff that is important for empirical asset pricing is the excess
return of asset 𝑖, defined as the return of asset 𝑖 in excess of a reference return,
𝑒,𝑖
𝑅𝑡+1 𝑖 − 𝑅 𝑏 . The chosen reference return is often the risk-free rate 𝑅 𝑓 . It
= 𝑅𝑡+1 𝑡+1 𝑡+1
follows that the price of any excess return is 0:
𝑒,𝑖
E𝑡 [𝑅𝑡+1 𝑚 𝑡+1 (𝜹)] = 0. (10.3)
Equation (10.3) also can be conveniently rewritten such that

𝑒,𝑖 𝑖 𝑓 𝑓 𝑖
E𝑡 (𝑅𝑡+1 ) = E𝑡 (𝑅𝑡+1 ) − 𝑅𝑡+1 = −𝑅𝑡+1 · cov𝑡 (𝑚 𝑡+1 , 𝑅𝑡+1 ), (10.4)
𝑓
3 Note that a riskless payoff implies that 𝑋𝑡+1 is known with certainty at some point in time 𝑡. Thus,
𝑓
the return 𝑅𝑡+1 is part of the information set in 𝑡 and can be drawn out of the E𝑡 [·] operator.
340 Sönksen
which yields an expression for the risk-premium associated with an asset 𝑖. The sign
and size of the premium, reflected in the conditional expected excess return on asset
𝑖, are determined by the conditional covariance of the asset return and the SDF on the
right-hand side of Equation (10.4). By inserting 𝑅𝑡+1 𝑚 , the return of a market index,
for 𝑅𝑡+1 , Equation (10.4) gives an expression for the market equity premium.
𝑖
Equation (10.3) and its reformulation in Equation (10.4) thus represent a corner-
stone of empirical asset pricing. Because Equation (10.3) is a conditional moment
constraint, it provides the natural starting point for the application of moment-based
estimation techniques. In particular, using instrumental variables 𝑧 𝑡 (part of the
econometrician’s and investor’s information set, observed at time 𝑡), it is possible to
generate unconditional moment conditions,
𝑒,𝑖
E𝑡 [𝑅𝑡+1 𝑧 𝑡 𝑚 𝑡+1 (𝜹)] = 0. (10.5)
Cochrane (2005) notes that Equation (10.5) can be conceived of as the conditioned
down version (using the law of total expectation) of the basic asset pricing equation
𝑒,𝑖
for a payoff 𝑅𝑡+1 𝑧 𝑡 of an asset class, which Cochrane refers to as managed portfolios.
By choosing a set of test assets and their excess returns, as well as a set of instruments,
the modeler can create a set of moment conditions that form the basis for GMM or
SMM estimations.
In this second application, standard econometric analysis for empirical asset
pricing again benefits from machine learning. It is not obvious which test assets and
which instruments to select to generate managed portfolios. Among the plethora of
test assets to choose from, and myriad potential instrumental variables, Cochrane
(2005) simply recommends choosing meaningful test assets and instruments, which
begs the underlying question. The selection of test assets and instruments serves two
connected purposes: to ensure that the parameters of the SDF can be efficiently
estimated, such that the moment conditions are informative about those parameters,
and to challenge the asset pricing model (i.e., the SDF model) with meaningful test
assets. In Section 10.3, I explain how machine learning can provide useful assistance
for these tasks.
Conditional Linear Factor Models and Machine Learning

Linear factor models have traditionally been of paramount importance in empirical
asset pricing; they imply an SDF that can be written as a linear function of 𝐾 risk
factors:
𝑚 𝑡+1 = 𝛿0 + 𝛿1 𝑓1,𝑡+1 + 𝛿2 𝑓2,𝑡+1 + · · · + 𝛿 𝐾 𝑓𝐾 ,𝑡+1 . (10.6)

The GMM approach works for asset pricing models with linear and nonlinear SDF
alike. However, linear factor models also can lend themselves to linear regression-
based analysis (see Chapter 12, Cochrane, 2005), provided that the test asset payoffs
are excess returns. The risk factors 𝑓 𝑘 can be excess returns themselves, and in many
empirical asset pricing models, they are. Empirical implementations of the capital
asset pricing model (CAPM) use the excess return of a wealth portfolio proxy as a
single factor. Another well-known example is the Fama-French three-factor model,

which in addition uses the risk factors HML (value; high-minus-low book-to-market)
and SMB (size; small-minus-big), both of which are constructed as excess returns
of long-short portfolios. Using the basic asset pricing equation as it applies to
an excess return and a linear factor model in which the factors are excess returns
provides an alternative way to write Equation (10.3), using 𝑧 𝑡 = 1 in an expected
return-beta-lambda representation:
E[𝑅𝑡𝑒,𝑖 ] = 𝛽1 𝜆1 + 𝛽2 𝜆2 + · · · + 𝛽𝐾 𝜆 𝐾 , (10.7)
such that the 𝛽 𝑘 for 𝑘 = 1, . . . , 𝐾 are linear projection coefficients that result from a
population regression of 𝑅 𝑒,𝑖 on the 𝐾 factors, and 𝜆 𝑘 denotes the expected excess
return of the 𝑘’th risk factor.
Although they are empirically popular, linear factor models induce both theoretical
and methodological problems, and again, machine learning has proven useful for
addressing them. In particular, a notable methodological issue arises because the
search for linear factor model specifications has created a veritable factor zoo,
according to Harvey, Liu and Zhu (2015) and Feng, Giglio and Xiu (2020). Cochrane
(2005) calls for discipline in selecting factors and emphasizes the need to specify
their connections to investor preferences or the predictability of components of the
SDF, yet their choice often is ad hoc. Harvey et al. (2015) identify 316 risk factors
introduced in asset pricing literature since the mid-1960s, when the CAPM was first
proposed (Sharpe, 1964, Lintner, 1965, Mossin, 1966), many of which are (strongly)
correlated, such that it rarely is clear whether a candidate factor contains genuinely
new information within the vast factor zoo.4 Another methodological concern involves
time-varying parameters in the SDF. In many situations, it may be argued that the
parameters in Equation (10.6) should be time-dependent, and therefore,
𝑚 𝑡+1 = 𝛿0,𝑡 + 𝛿1,𝑡 𝑓1,𝑡+1 + 𝛿2,𝑡 𝑓2,𝑡+1 + · · · + 𝛿 𝐾 ,𝑡 𝑓𝐾 ,𝑡+1 . (10.8)

The CAPM is a prominent example in which the market excess return is the only
factor, such that 𝐾 = 1 in Equation (10.8). However, a theory-consistent derivation
of the CAPM’s SDF implies that indeed the parameters in the linear SDF must be
time-dependent. Regardless of the rationale for including time-varying parameters
of the SDF, conditioning down the conditional moment restriction in Equation
(10.3) by using the law of total expectations is not possible, so an expected return
𝛽-𝜆 representation equivalent to Equation (10.7) must have time-varying 𝛽 and 𝜆.
Time-varying 𝛽 also emerge if the test assets are stocks for firms with changing
business models. Jegadeesh, Noh, Pukthuanthong, Roll and Wang (2019) and Ang,
Liu and Schwarz (2020) argue that an aggregation of stocks into portfolios according
to firm characteristics – which arguably alleviates the non-stationarity of the return
series – may not be innocuous.
4 Harvey et al. (2015) restrict their analysis to factors published in a leading finance journal or
proposed in outstanding working papers. There are many other factors, not accounted for by Harvey
et al., that have been suggested.
342 Sönksen
As a result, neither GMM nor regression-based analysis are directly applicable. In

the particular case of the CAPM, the Hansen-Richard critique states that the CAPM
is not testable in the first place (Hansen & Richard, 1987). To provide a solution to
the problem, Cochrane (1996) proposes scaling factors by using affine functions of
time 𝑡 variables for the parameters in Equation (10.6); this approach has been applied
successfully by Lettau and Ludvigson (2001). Yet to some extent, these notions
represent ad hoc and partial solutions; prior theory does not establish clearly which
functional forms should be used.
Dealing with the challenges posed by linear factor models thus constitutes the
third area in which machine learning proves useful for empirical asset pricing. In
Section 10.4, I outline how machine learning can ‘tame the factor zoo’ and address
the problem of time-varying parameters in conditional factor asset pricing models.
Predictability of Asset Returns and Machine Learning

Cochrane (2005) emphasizes that the basic asset pricing Equation (10.1) does not
rule out the predictability of asset returns. In fact, the reformulated basic asset pricing
equation for an excess return in (10.4) not only provides an expression for the risk
premium associated with asset 𝑖 but also establishes the optimal prediction of the
𝑒,𝑖
excess return 𝑅𝑡+1 , provided the loss function is the mean squared error (MSE) of the
forecast. The conditional expected value is the MSE-optimal forecast. Accordingly,
to derive theory-consistent predictions, Equation (10.4) represents a natural starting
point. Conceiving the conditional covariance cov𝑡 (𝑚 𝑡+1 , 𝑅𝑡+1 𝑖 ) as a function of time
𝑡 variables, one could consider flexible functional forms and models to provide
MSE-optimal excess return predictions. This domain represents a natural setting for
the application of machine learning methods.
Exploiting the return predictability that is theoretically possible has always been an
active research area, with obvious practical interest; machine learning offers intriguing
possibilities along these lines. In the present context with its low signal to noise ratio,
diligence is required in creating training and validation schemes when those highly
flexible multidimensional statistical methods are employed. Moreover, limitations
implied by theory-based empirical assessments of risk premia using information
contained in option data should be considered. I present this fourth area of asset
pricing using machine learning by giving best practice examples in Section 10.5 of
this chapter.
In summary, and in line with the outline provided in the preceding sections, Section
10.2 contains an explanation of how machine learning techniques can help identify
the SDF. Section 10.3 elaborates on the use of machine learning techniques to test and
evaluate asset pricing models. Then Section 10.4 reports on applications of machine
learning to estimate linear factor models. In Section 10.5, I explain how machine
learning techniques can be applied to prediction problems pertaining to empirical
asset pricing. Finally, I offer some concluding remarks in Section 10.6.
10.2 How Machine Learning Techniques Can Help Identify

Stochastic Discount Factors
Chen, Pelger and Zhu (2021) aim to identify an SDF that is able to approximate risk
premia at the stock level. Their study is particularly interesting, in that they (i) propose
a loss function based on financial economic theory; (ii) use three distinct types of
neural networks, each of which is tailored to a particular aspect of the identification
strategy; and (iii) relate that strategy to the GMM framework. Furthermore, Chen et
al. argue that whilst machine learning methods are suitable for dealing with the vast
set of conditioning information an SDF might depend on (e.g., firm characteristics,
the business cycle) and generally can model complex interactions and dependencies,
their application to empirical asset pricing may be hampered by the very low signal-
to-noise ratio of individual stock returns. To counteract this problem and to provide
additional guidance in the training of machine learning models, Chen et al. propose
imposing economic structure on them through the no-arbitrage condition. Instead of
relying on the loss functions conventionally used in machine learning applications
(e.g., MSE minimization), they revert to the basic asset pricing Equation (10.3). This
theory-guided choice of the loss function represents a key novelty and contribution
of their study. Two further contributions refer to the way in which Chen et al. extract
hidden states of the economy from a panel of macroeconomic variables and to the
data-driven technique that is used to construct managed portfolios (as I describe in
more detail subsequently).
Regarding the SDF, Chen et al. (2021) assume a linear functional form, using
excess returns as factors and accounting for time-varying factor weights:
𝑁
∑︁
𝑚 𝑡+1 = 1 − 𝑒,𝑖
𝜔𝑡 ,𝑖 𝑅𝑡+1 = 1 − 𝝎𝑡′ R𝑡+1
𝑒
, (10.9)
𝑖=1
where 𝑁 denotes the number of excess returns. Note that the SDF depicted in Equation
(10.9) is a special case of the SDF representation in Equation (10.8), in which the
factors are excess returns. When using solely excess returns as test assets, the mean of
the SDF is not identified and can be set to any value. Chen et al. (2021) choose 𝛿0,𝑡 = 1,
which has the interesting analytical implication that 𝝎𝑡 = E𝑡 [R𝑡+1 𝑒 R𝑒 ′ ] −1 E [R𝑒 ].
𝑡+1 𝑡 𝑡+1
The weights in Equation (10.9) are both time-varying and asset-specific. They
are functions of firm-level characteristics (I𝑡 ,𝑖 ) and information contained in mac-
roeconomic time series (I𝑡 ). By including not only stock-level information but also
economic time series in the information set, the SDF can capture the state of the
economy (e.g., business cycle, crises periods). Chen et al. (2021) rely on more
than 170 individual macroeconomic time series, many of which are strongly cor-
related. From this macroeconomic panel, they back-out a small number of hidden
variables that capture the fundamental dynamics of the economy using a recurrent
long-short-term-memory network (LSTM), such that h𝑡 = ℎ(I𝑡 ) and 𝜔𝑡 ,𝑖 = 𝜔(h𝑡 , I𝑡 ,𝑖 ).
As suggested by its name, this type of neural network is particularly well suited to
detecting both short- and long-term dependencies in the macroeconomic panel and
344 Sönksen
thus for capturing business cycle dynamics. To account for nonlinearities and possibly
complex interactions between I𝑡 and h𝑡 , 𝜔𝑡 ,𝑖 can be modeled using a feedforward
neural network. For an introduction to neural networks, see Chapter 4.
To identify the SDF weights 𝝎, Chen et al. (2021) rely on managed portfolios. The
idea behind using managed portfolios is that linear combinations of excess returns are
excess returns themselves. Thus, instead of exclusively relying on the excess returns
𝑒,𝑖
of individual stocks 𝑅𝑡+1 for estimation purposes, it is possible to use information
available at time 𝑡 to come up with portfolio weights g𝑡 and new test assets R𝑡+1 𝑒 g .
𝑡
Historically, managed portfolios have been built in an ad hoc fashion, such as by using
time 𝑡 information on the price-dividend ratio. Choosing a data-driven alternative,
Chen et al. (2021) construct each element of g𝑡 ,𝑖 as a nonlinear function of I𝑡 ,𝑖 and a
𝑔
set of hidden macroeconomic state variables h𝑡 = ℎ𝑔 (I𝑡 ), where the hidden states in
𝑔
h𝑡 and h𝑡 may differ. However, just as with the SDF specification, Chen et al. infer
ℎ𝑔 (·) with an LSTM, and they model g𝑡 using a feedforward neural network.
Putting all of these elements together – no-arbitrage condition, managed portfolios,
SDF and portfolio weights derived from firm-level and macroeconomic information –
leads to the moment constraints:
" 𝑁
! #
∑︁
𝑒,𝑖 𝑒
E 1− 𝜔𝑡 ,𝑖 𝑅𝑡+1 R𝑡+1 g𝑡 = 0,
𝑖=1
𝑔
which serve as a basis for identifying the SDF. However, 𝑔(h𝑡 , I𝑡 ,𝑖 ) would allow
for constructing infinitely many different managed portfolios. To decide which
portfolios are most helpful for pinning down the SDF, Chen et al. build on Hansen
and Jagannathan’s (1997) finding that the SDF proxy that minimizes the largest
possible pricing error is the one that is closest to an admissible true SDF in least
square distance. Therefore, Chen et al. set up the empirical loss function:
! 2
𝑁 𝑁
𝑔 1 ∑︁ 𝑇𝑖 1 ∑︁ ∑︁
𝑒,𝑖 𝑒,𝑖

𝐿 (𝝎| ĝ, h𝑡 , h𝑡 , I𝑡 ,𝑖 ) = 1− 𝜔𝑡 ,𝑖 𝑅𝑡+1 𝑅𝑡+1 ĝ𝑡 ,𝑖

𝑁 𝑖=1 𝑇 𝑇𝑖 𝑡 ∈𝑇 𝑖=1

𝑖
and formulate a minimax optimization problem in the spirit of a generative adversarial

network (GAN):
𝑔 𝑔
𝝎,
ˆ ĥ𝑡 , ĝ, ĥ𝑡 = min max 𝐿(𝝎| ĝ, h𝑡 , h𝑡 , I𝑡 ,𝑖 ).
𝑔 𝝎,h𝑡 g,h𝑡
Within this framework, Chen et al. (2021) pit two pairs of neural networks against
each other: The LSTM that extracts h𝑡 from macroeconomic time series and the
feedforward network that generates 𝝎 from h𝑡 together with the asset-specific
characteristics I𝑡 ,𝑖 are trained to minimize the loss function, whereas the other LSTM
and feedforward network serve as adversaries, aiming to construct managed portfolios
that are particularly hard to price, such that they maximize pricing errors.
For their empirical analysis, Chen et al. (2021) use monthly excess returns on
𝑁 = 10, 000 stocks for the period from January 1967 to December 2016 and thereby
establish five main findings. First, performance of highly flexible machine learning
models in empirical asset pricing can be improved by imposing economic structure,
such as in the form of a no-arbitrage constraint. Second, interactions between stock-
level characteristics matter. Third, the choice of test assets is important. This finding
is in line with the literature presented in Section 10.3. Fourth, macroeconomic states
matter and it is important to consider the full time series of macroeconomic variables
instead of reduced-form information, such as first differences or the last observation.
Fifth, the performance of the flexible neural network approach proposed by Chen et
al. (2021) might be improved, if it were combined with multifactor models of the
IPCA-type, as discussed in Section 10.4.2.
An alternative approach to identifying SDFs is brought forward by Korsaye, Quaini
and Trojani (2019) who enforce financial economic theory on the SDF by minimizing
various concepts of SDF dispersion whilst imposing constraints on pricing errors.
The approach by Korsaye et al. is inspired by the works of Hansen and Jagannathan
(1991) on identifying the admissible minimum variance SDF. This SDF constitutes a
lower bound on the variance of any admissible SDF and thus upper-bounds the Sharpe
ratio that can be attained using linear portfolios of traded securities. The authors
argue that the constraints which they impose on the pricing errors can be justified,
for example, by market frictions and they name the thus resulting model-free SDFs
minimum dispersion smart SDFs (S-SDFs). Korsaye et al. consider different measures
of dispersion, as well as multiple economically motivated penalties on the pricing
errors. Furthermore, they develop the econometric theory required for estimation and
inference of S-SDFs and propose a data-driven method for constructing minimum
variance SDFs.
10.3 How Machine Learning Techniques Can Test/Evaluate Asset

Pricing Models
The idea that the choice of test assets is important when evaluating and testing asset
pricing models – either by standard econometric techniques or incorporating machine
learning methods – is not novel. For example, Lewellen, Nagel and Shanken (2010)
criticize a common test of asset pricing models that uses book-to-market sorted
portfolios as test assets. In a similar vein, Ahn, Conrad and Dittmar (2009) warn that
constructing portfolios based on characteristics that are known to be correlated with
returns might introduce a data-snooping bias. They suggest an alternative strategy,
which forms base assets by grouping securities that are more strongly correlated, and
compare the inferences drawn from this set of basis assets with those drawn from
other benchmark portfolios.
In more recent, machine learning-inspired literature on empirical asset pricing,
Chen et al. (2021) reiterate the importance of selecting test assets carefully. In another
approach, with a stronger focus on model evaluation, Bryzgalova, Pelger and Zhu
(2021a) argue that test assets conventionally used in empirical asset pricing studies
do not only fail to provide enough of a challenge to the models under consideration
346 Sönksen
but also contribute to the growing factor zoo. To circumvent this issue, they propose
so-called asset pricing trees (AP-Trees) that can construct easy-to-interpret and
hard-to-price test assets. An overview of other tree-based approaches is available in
Chapter 2 of this book.
Bryzgalova, Pelger and Zhu (2021a) consider a linear SDF, spanned by 𝐽 managed
portfolios, constructed from 𝑁 excess returns:
𝐽
∑︁ 𝑁
∑︁
𝑒,man 𝑒,man 𝑒,𝑖
𝑚 𝑡+1 = 1 − 𝜔𝑡 , 𝑗 R𝑡+1, 𝑗 with R𝑡+1, 𝑗 = 𝑓 (𝐶𝑡 ,𝑖 )𝑅𝑡+1 , (10.10)
𝑗=1 𝑖=1
where 𝑓 (𝐶𝑡 ,𝑖 ) denotes a nonlinear function of the time 𝑡 stock characteristics 𝐶𝑡 ,𝑖 .

The representation in Equation (10.10) resembles the linear SDF used by Chen et
al. (2021), as described in Section 10.2 of this chapter. However, the SDF weights they
applied were also functions of macroeconomic time series. Additionally, Bryzgalova,
Pelger and Zhu approach Equation (10.10) from the perspective of trying to find the
tangency portfolio, that is, the portfolio on the mean-variance frontier that exhibits
the highest Sharpe ratio:
E[𝑅 𝑒 ] 𝜎(𝑚)
𝑆𝑅 = with |𝑆𝑅| ≤ ,
𝜎(𝑅 𝑒 ) E(𝑚)
where 𝜎(𝑚) denotes the unconditional standard deviation of the SDF. The mean-
variance frontier is the boundary of the mean-variance region, which contains
all accessible combinations of means and variances of the assets’ excess returns.
Put differently, for any given variance of 𝑅 𝑒 , the mean-variance frontier answers
the question of which mean excess returns are accessible. Importantly, all excess
returns that lie exactly on the mean-variance frontier are perfectly correlated with
all other excess returns on the boundary and also perfectly correlated with the
(𝑚)
SDF. Furthermore, 𝜎 E(𝑚) is the slope of the mean-variance frontier at the tangency
portfolio and upper-bounds the Sharpe ratio of any asset.5 In this sense, Bryzgalova,
Pelger and Zhu (2021a) attempt to span the SDF by forming managed portfolios
Í 𝑒,man
through 𝑓 (𝐶𝑡 ,𝑖 ) and selecting weights 𝝎, such that 𝐽𝑗=1 𝜔𝑡 ,𝑖 R𝑡+1, 𝑗 approximates the
tangency portfolio. This theoretical concept serves as a foundation for their pruning
strategy. Bryzgalova, Pelger and Zhu (2021a) construct the managed portfolios in
Equation (10.10) using trees, which group stocks into portfolios on the basis of
their characteristics. Trees lend themselves readily to this purpose, because they are
reminiscent of the stock characteristic-based double- and triple-sorts that have served
portfolio construction purposes for decades. Applying the tree-based approach, a
large number of portfolios could be constructed.
To arrive at a sensible number of interpretable portfolios that help recover the SDF,
Bryzgalova, Pelger and Zhu (2021a) also introduce a novel pruning strategy that aims
at maximizing the Sharpe ratio. To apply this technique, the authors compute the
variance-covariance matrix of the cross-section of the candidate managed portfolios
5 The upper bound is derived in Appendix 1.

constructed with AP-Trees, 𝚺, ˆ then assemble the corresponding portfolio means in

𝝁ˆ . Furthermore, they rephrase the problem of maximizing the Sharpe ratio of the
tangency portfolio proxy as a variance minimization problem, subject to a minimum
mean excess return of the portfolio proxy. An 𝐿 1 -penalty imposed on portfolio weights
helps select those portfolios relevant to spanning the SDF. Additionally, the sample
average portfolio returns are shrunken to their cross-sectional average value by an
𝐿 2 -penalty, accounting for the fact that extreme values are likely due to over- or
underfitting (see Chapter 1 of this book for an introduction to 𝐿 1 and 𝐿 2 penalties).
The optimization problem thus results in
1 ′ˆ 1
min 𝝎 𝚺𝝎 + 𝜆1 ||𝝎|| 1 + 𝜆2 ||𝝎|| 22 ,
𝝎 2 2
𝑠.𝑡. 𝝎 ′1 = 1 (10.11)
𝝎 ′ 𝝁ˆ ≥ 𝜇0 ,
where 1 denotes a vector of ones, and 𝜆1 , 𝜆2 , and 𝜇0 are hyperparameters. Apart

from the inclusion of financial economic theory, a key difference between the pruning
technique outlined in Equation (10.11) and standard techniques used in related
literature (see Chapter 2 for more details on tree-based methods and pruning) is that
in the former case, the question of whether to collapse children nodes cannot be
answered solely on the basis of these nodes (e.g., by comparing the Sharpe ratios of
the children nodes to that of the parent node). Instead, the decision depends on the
𝑒,man
Sharpe ratio of 𝝎R𝑡+1 , so all other nodes must be taken into account too.6
Noting some parallels between the studies by Chen et al. (2021) and Bryzgalova,
Pelger and Zhu (2021a) – both adapt highly nonlinear machine learning approaches
with financial economic theory to identify the SDF, both highlight the importance
of selecting test assets that help span the SDF, and both propose strategies for the
construction of such managed portfolios – the relevant differences in their proposed
strategies and the focus of their studies are insightful as well. In particular, in the
neural network-driven approach by Chen et al. (2021), constructing hard-to-price test
assets is critical, because they adopt Hansen and Jagannathan’s (1997) assertion that
minimizing pricing errors for these assets corresponds to generating an SDF proxy
that is close to a true admissible SDF. The managed portfolios constructed by the
GAN are complex, nonlinear functions of the underlying macroeconomic time series
and stock-specific characteristics, such that they are not interpretable in a meaningful
way. In contrast, the portfolios obtained from the AP-Trees proposed by Bryzgalova,
Pelger and Zhu (2021a) can be straightforwardly interpreted as a grouping of stocks
that share certain characteristics. The thus identified portfolios have value of their
own accord and can be used as test assets in other studies.
Bryzgalova, Pelger and Zhu (2021a) evaluate the ability of their framework to
recover the SDF and construct hard-to-price test assets from monthly data between
January 1964 and December 2016. They consider 10 firm-specific characteristics
and arrive at three pertinent conclusions. First, a comparison with conventionally
6 Further methodological details are available from Bryzgalova, Pelger and Zhu (2021b).
348 Sönksen
sorted portfolios shows that the managed portfolios constructed from AP-Trees
yield substantially higher Sharpe ratios and thus recover the SDF more successfully.
Second, the way AP-Trees capture interactions between characteristics is particularly
important. Third, according to robustness checks, imposing economic structure on
the trees by maximizing the Sharpe ratio is crucial to out-of-sample performance.
Giglio, Xiu and Zhang (2021) are also concerned with the selection of test assets
for the purpose of estimating and testing asset pricing models. They point out that
the identification of factor risk premia hinges critically on the test assets under
consideration and that a factor may be labeled weak because only a few of the
test assets are exposed to it, making standard estimation and inference incorrect.
Therefore, their novel proposal for selecting assets from a wider universe of test assets
and estimating the risk premium of a factor of interest, as well as the entire SDF,
explicitly accounts for weak factors and assets with highly correlated risk exposures.
The procedure consists of two steps that are conducted iteratively: Given a particular
factor of interest, the first step selects those test assets that exhibit the largest (absolute)
correlation with that factor. These test assets are then used in principal component
analysis (PCA) for the construction of a latent factor. Then, a linear projection is used
to make both the test asset returns and the factor orthogonal to the latent factor before
returning to the selection of the most correlated test assets. Giglio, Xiu and Zhang
refer to their proposed methodology as supervised principal component analysis
(SPCA) and argue that the iterative focus on those test assets that exhibit the highest
(absolute) correlation with the (residual) factor of interest ensures that weak factors
are also captured. In establishing the asymptotic properties of the SPCA estimator,
and comparing its limiting behavior with that of other recently proposed estimators
(e.g., Ridge, LASSO, and partial least squares), Giglio, Xiu and Zhang report that the
SPCA outperforms its competitors in the presence of weak factors not only in theory,
but also in finite samples.
10.4 How Machine Learning Techniques Can Estimate Linear

Factor Models
This section illustrates how machine learning techniques can alleviate the issues
faced in the context of linear factor models outlined in the introduction: dealing with
time-varying parameters and the selection of factors. In Section 10.4.1, I present a
two-step estimation approach proposed by Gagliardini, Ossola and Scaillet (2016).
Their initial methodology was not strictly machine learning (the boundaries with
‘standard’ econometric method development are blurred anyway), but in later variants,
the identification strategy included more elements of machine learning. In Section
10.4.2, I present the instrumented principal components analysis (IPCA) proposed by
Kelly, Pruitt and Su (2019), which modifies standard PCA, a basic machine learning
method. A workhorse method, IPCA can effectively analyze conditional linear factor
models. Section 10.4.3 presents an extension of IPCA that adopts more machine
learning aspects, namely, Gu, Kelly and Xiu’s (2021) autoencoder approach. In
Section 10.4.4, I outline the regularized Bayesian approach introduced by Kozak,

Nagel and Santosh (2018). Finally, Section 10.4.5 lists some recent contributions that
address the selection of factors, including the problem of weak factors. Appendix 2
contains an overview of different PCA-related approaches.
10.4.1 Gagliardini, Ossola, and Scaillet’s (2016) Econometric Two-Pass

Approach for Assessing Linear Factor Models
Gagliardini et al. (2016) propose a novel econometric methodology to infer time-

varying equity risk premia from a large unbalanced panel of individual stock returns
under conditional linear asset pricing models. Their weighted two-pass cross-sectional
estimator incorporates conditioning information through instruments – some of which
are common to all assets and others are asset-specific – and its consistency and
asymptotic normality are derived under simultaneously increasing cross-sectional and
time-series dimensions. In their empirical analysis, Gagliardini et al. (2016) consider
monthly return data between July 1964 and December 2009 for about 10,000 stocks.
They find that risk premia are large and volatile during economic crises and appear to
follow the macroeconomic cycle.
A competing approach comes from Raponi, Robotti and Zaffaroni (2019), who
consider a large cross-section, but – in contrast with Gagliardini et al. (2016) – use a
small and fixed number of time-series observations. Gagliardini, Ossola and Scaillet
(2019) build on these developments, focusing on omitted factors. They propose
a diagnostic criterion for approximate factor structure in large panel data sets. A
misspecified set of observable factors will turn risk premia estimates obtained by
means of 2-pass regression worthless. Under correct specification, however, errors
will be weakly cross-sectionally correlated and this serves as the foundation of the
newly proposed criterion, which checks for observable factors whether the errors
are weakly cross-sectionally correlated or share one or more unobservable common
factor(s). This approach to determine the number of omitted common factors can
also be applied in a time-varying context.
Bakalli, Guerrier and Scaillet (2021) still work with a two-pass approach and aim
at time-varying factor loadings, but this time, they make use of machine learning
methods (𝐿 1 -penalties) to ensure sparsity in the first step. In doing so, they address a
potential weakness of the method put forward by Gagliardini et al. (2016), which
refers to the large number of parameters required to model time-varying factor
exposures and risk premia. Bakalli et al. thus develop a penalized two-pass regression
with time-varying factor loadings. In the first pass, a group LASSO is applied to
target the time-invariant counterpart of the time-varying models thereby maintaining
compatibility with the no arbitrage restrictions. The second pass delivers risk premia
estimates to predict equity excess returns. Bakalli et al. derive the consistency for
their estimator and exhibit its good out-of-sample performance using a simulation
study and monthly return data on about 7,000 stocks in the period from July 1963 to
December 2019.
350 Sönksen
10.4.2 Kelly, Pruitt, and Su’s (2019) Instrumented Principal

Components Analysis
Kelly et al. (2019) propose a modeling approach for the cross-section of returns. Their
IPCA method is motivated by the idea that stock characteristics might line up with
average returns, because they serve as proxies for loadings on common (but latent)
risk factors. The methodology allows for latent factors and time-varying loadings by
introducing observable characteristics that instrument for the unobservable dynamic
loadings. To account for this idea, the authors model excess returns as:
𝑒,𝑖
𝑅𝑡+1 = 𝛼𝑖,𝑡 + 𝜷𝑖,𝑡 f𝑡+1 + 𝜀𝑖,𝑡+1
′
where 𝛼𝑖,𝑡 = z𝑖,𝑡 𝚪 𝛼 + 𝜈 𝛼,𝑖,𝑡 (10.12)
′
and 𝛽𝑖,𝑡 = z𝑖,𝑡 𝚪 𝛽 + 𝝂 𝛽,𝑖,𝑡 ,
where f𝑡+1 denotes a vector of latent risk factors, and z𝑖,𝑡 is a vector of characteristics
that is specific to stock 𝑖 and time 𝑡. 𝚪 𝛼 and 𝚪 𝛽 are time-independent matrices.
The time dependency of 𝛼𝑖,𝑡 and 𝜷𝑖,𝑡 reflects the ways that the stock characteristics
themselves may change over time. The key idea of Kelly et al. (2019) is to estimate
f𝑡+1 , 𝚪 𝛼 , and 𝚪 𝛽 jointly via:
𝑇−1
! −1 𝑇−1
!
′
∑︁ h i′
⊗ ˆf̃𝑡+1ˆf̃𝑡+1 Z𝑡 ⊗ ˆf̃𝑡+1
∑︁
vec 𝚪ˆ = Z𝑡′ Z𝑡 ′ ′ 𝑒
R𝑡+1
𝑡=1 𝑡=1
′ ′
−1 ′
and ˆf̃ = 𝚪ˆ 𝛽 Z𝑡 Z𝑡 𝚪ˆ 𝛽 𝚪ˆ 𝛽 Z𝑡′ (R𝑡+1
𝑒
− Z𝑡 𝚪ˆ 𝛼 ),
𝑡+1
where 𝚪ˆ = 𝚪ˆ 𝛼 , 𝚪ˆ 𝛽

and ˆf̃ = [1, f̂ ′ ] ′ .
𝑡+1 𝑡+1
This specification allows for variation in returns to be either attributed to factor

exposure (by means of 𝚪 𝛽 ) or to an anomaly intercept (through 𝚪 𝛼 ).7 The IPCA
framework also supports tests regarding the characteristics under consideration, such
as whether a particular characteristic captures differences in average returns that are
not associated with factor exposure and thus with compensation for systematic risk.
With IPCA, it is possible to assess the statistical significance of one characteristic
while controlling for all others. In addressing the problem of the growing factor zoo,
Kelly et al. (2019) argue that such tests can reveal the genuine informational content of
a newly discovered characteristic, given the plethora of existing competitors. For both
types of tests, the authors rely on bootstrap inference. They outline the methodological
theory behind IPCA in Kelly, Pruitt and Su (2020).
In their empirical analysis, Kelly et al. (2019) use a data set provided by Freyberger,
Neuhierl and Weber (2020) that contains 36 characteristics of more than 12,000
stocks in the period between July 1962 and May 2014. Allowing for five latent factors,
the authors find that IPCA outperforms the Fama-French five-factor model. They also
7 Note that a unique identification requires additional restrictions, which Kelly et al. (2019) impose
through the orthogonality constraint 𝚪′𝛼 𝚪 𝛽 = 0.
find that only 10 of the 36 characteristics are statistically significant at the 1% level
and that applying IPCA to just these 10 characteristics barely affects model fit.
10.4.3 Gu, Kelly, and Xiu’s (2021) Autoencoder
Kelly et al. (2019) make use of stock covariates to allow for time-varying factor
exposures. However, they impose that covariates affect factor exposures in a linear
fashion. Accounting for interactions between the covariates or higher orders is
generally possible, but it would have to be implemented manually (i.e., by computing
the product of two covariates and adding it as an additional instrument).
Gu et al. (2021) offer a generalization of IPCA that still is applicable at the
stock level and accounts for time-varying parameters but that does not restrict the
relationship between covariates and factor exposures to be linear, as is implied by
Equation (10.12). According to their problem formulation, identifying latent factors
and their time-varying loadings can be conceived of as an autoencoder that consists
of two feedforward neural networks. One network models factor loadings from a
large set of firm characteristics; the other uses excess returns to flexibly construct
latent factors. Alternatively, a set of characteristic-managed portfolios can be used to
model the latent factors. At the output layer of the autoencoder, the 𝛽𝑡 ,𝑖,𝑘 estimates
(specific to time 𝑡, asset 𝑖, and factor 𝑘) of the first network combine with the 𝐾
factors constructed by the second network, thereby providing excess return estimates.
This model architecture is based on the no-arbitrage condition and thus parallels the
SDF identification strategy by Chen et al. (2021) outlined in Section 10.2 of this
chapter.
The term autoencoder signifies that the model’s target variables, the excess
returns of different stocks, are also input variables. Hence, the model first encodes
information contained in the excess returns (by extracting latent factors from them),
before decoding them again to arrive at excess return predictions. Gu et al. (2021)
then confirm that IPCA results as a special case of their autoencoder specification.
The autoencoder is trained to minimize the MSE between realized excess returns
and their predictions, and an 𝐿 1 -penalty together with early stopping helps avoid
overfitting. Using monthly excess returns for all NYSE-, NASDAQ-, and AMEX-
traded stocks from March 1957 to December 2016, as well as 94 different firm-level
characteristics, Gu et al. (2021) retrain the autoencoder on an annual basis to provide
out-of-sample predictions. This retraining helps ensure the model can adapt to changes
in the factors or their loadings, but it constitutes a computationally expensive step,
especially compared with alternative identification strategies, such as IPCA.
Contrasting the out-of-sample performance of their network-based approach with
that of competing strategies, including IPCA, Gu et al. (2021) establish that the
autoencoder outperforms IPCA in terms of both the out-of-sample 𝑅 2 and the out-
of-sample Sharpe ratio. Thus, the increased flexibility with which factors and their
loadings can be described by the autoencoder translates into improved predictive
abilities.
352 Sönksen
10.4.4 Kozak, Nagel, and Santosh’s (2020) Regularized Bayesian

Approach
Kozak, Nagel and Santosh (2020) present a Bayesian alternative to Kelly et al.’s
(2019) IPCA approach. They argue against characteristic-sparse SDF representations
and suggest constructing factors by first computing principal components (PCs) from
the vast set of cross-sectional stock return predictors, then invoking regularization
later, to select a small number of PCs to span the SDF. Building on their findings
in Kozak et al. (2018), the authors argue that a factor earning high excess returns
should also have a high variance. This theoretical consideration is incorporated in a
Bayesian prior on the means of the factor portfolios, which are notoriously hard to
estimate. As a consequence, more shrinkage gets applied to the weights of PC-factors
associated with low eigenvalues.
Kozak et al. (2020) identify the weights 𝝎, by pursuing an elastic net mean-variance
optimization:
1 ˆ + 𝜆1 ||𝝎|| 1 + 1 𝜆2 ||𝝎|| 2 ,
ˆ ′ 𝚺ˆ −1 ( 𝝁ˆ − 𝚺𝝎)
min ( 𝝁ˆ − 𝚺𝝎) 2 (10.13)
𝝎 2 2
where 𝚺ˆ is the variance-covariance matrix of PC-factors, and 𝝁ˆ refers to its mean.

The 𝐿 2 -penalty is inherited from the Bayesian prior; including the 𝐿 1 -penalty in
Equation (10.13) supports factor selection.
For their empirical analyses, Kozak et al. (2020) use different sets of stock
characteristics and transform them to account for interactions or higher powers before
extracting principal components. Comparing the performance of their PC-sparse
SDFs to that of models that rely on only a few characteristics, they find that PC-sparse
models perform better.
The pruning function used by Bryzgalova, Pelger and Zhu (2021a) in Equation
(10.11) is similar to Equation (10.13); and indeed, Bryzgalova, Pelger and Zhu argue
that their weights 𝝎ˆ and those proposed by Kozak et al. (2020) coincide in the case
of uncorrelated assets. Linking the Bayesian setup of Kozak et al. (2020) to the
IPCA approach of Kelly et al. (2019), Nagel (2021, p. 89) points out that the IPCA
methodology (see Section 10.4.2) requires a prespecification of the number of latent
factors, which could be understood within the framework of Kozak et al. (2020) “as a
crude way of imposing the prior beliefs that high Sharpe ratios are more likely to
come from major sources of covariances than from low eigenvalue PCs.”
10.4.5 Which Factors to Choose and How to Deal with Weak Factors?
As mentioned in the introduction, two important issues associated with working with
factor models are the choice of factors under consideration and how to deal with weak
factors. These problems are by no means new, and plenty of studies attempt to tackle
them from traditional econometric perspectives. However, the ever-increasing factor
zoo and number of candidate variables aggravate the issues and make applications
of machine learning methodologies, e.g., as they relate to regularization techniques,
highly attractive. One method frequently used in analyses of (latent) factors is PCA.
Interestingly, this technique is an inherent part of both classic econometrics and
also the toolbox of machine learning methods. For this reason, I highlight in this
subsection those studies as invoking machine learning methods that not only apply
PCA (and analyze its respective strengths and weaknesses), but that extend PCA by
some other component that is used in the context of machine learning, such as a
regularization or penalty term.
Amongst the econometric approaches, Bai and Ng (2002) propose panel criteria
to determine the right number of factors in a model in which both the time-series
and cross-sectional dimension are very large. Onatski (2012) analyzes the finite
sample distribution of principal components when factors are weak and proposes an
approximation of the finite sample biases of PC estimators in such settings. Based
on these findings, he develops an estimator for the number of factors for which
the PCA delivers reliable results and applies this methodology to U.S. stock return
data, leading him to reject the hypothesis that the Fama and French (1993) factors
span the entire space of factor returns. Instead of focusing on the number of factors,
Bailey, Kapetanios and Pesaran (2021) aim at estimating the strength of individual
(observed and unobserved) factors. For this purpose, they propose a measure of factor
strength that builds on the number of statistically significant factor loadings and
for which consistency and asymptotics can be established if the factors in question
are at least moderately strong. Monte Carlo experiments serve to study the small
sample properties of the proposed estimator. In an empirical analysis, Bailey et al.
evaluate the strength of the 146 factors assembled by Feng et al. (2020) and find
that factor strength exhibits a high degree of time variation and that only the market
factor qualifies as strong. On a related issue, Pukthuanthong, Roll and Subrahmanyam
(2018) propose a protocol for identifying genuine risk factors and find that many
characteristics-based factors do not pass their test.8 Notably, the market factor is
amongst those candidate factors that comply with the protocol.
Gospodinov, Kan and Robotti (2014) note that inference fails in linear factor
models containing irrelevant factors, in the sense that the irrelevant factors have a high
probability of being mistaken for being priced. Proposing a method that establishes
proper inference in such misspecified models, Gospodinov et al. (2014) find little
evidence that macro factors are important – a finding that conflicts with Chen et al.’s
(2021) application of LSTM models to macro factors, which reveals that these factors
are important but that it is crucial to consider them as time series. In line with some
of the studies previously mentioned, Gospodinov et al. identify the market factor as
one of few factors that appear to be priced.
Anatolyev and Mikusheva (2022) describe three major challenges in dealing with
factor models. First, factors might be weak, but still priced, such that the resulting
betas and estimation errors are of the same order of magnitude. Second, the error
terms might be strongly correlated in the cross-section (e.g., due to mismeasurement
8 The protocol comprises the correlation between factors and returns, the factor being priced in the
cross-section of returns, and also a sensible reward-to-risk ratio.
354 Sönksen
of the true factors), thereby interfering with both estimation and inference. Third, in
empirical applications, the number of assets or portfolios considered is often of a
similar size as the time-series dimension. Anatolyev and Mikusheva show that the
two-pass estimation procedure conventionally applied in the context of linear factor
models results in inconsistent estimates when confronted with the aforementioned
challenges. Relying on sample-splitting and instrumental variables regression, they
come up with a new estimator that is consistent and can be easily implemented.
However, the model under consideration by Anatolyev and Mikusheva (2022) does
not account for time-varying factor loadings.9
Also on the topic of weak factors, but more closely related to machine learning
methods, Lettau and Pelger (2020b) extend PCA by imposing economic structure
through a penalty on the pricing error, thereby extracting factors not solely based on
variation, but also on the mean of the data. They name this approach to estimating
latent factors risk-premium PCA (RP-PCA) and argue that it allows to fit both the
cross-section and time series of expected returns. Furthermore, they point out that
RP-PCA – in contrast to conventional PCA – allows to identify weak factors with high
Sharpe ratios. RP-PCA is described in more detail in Appendix 10.6; the statistical
properties of the estimator are derived in Lettau and Pelger (2020a). In an alternative
approach of combining financial economic structure and the flexibility of machine
learning techniques, Feng, Polson and Xu (2021) train a feedforward neural network
to study a characteristics-sorted factor model. Their objective function focuses on
the sum of squared pricing errors and thus resembles an equally weighted version
of the GRS test statistic by Gibbons, Ross and Shanken (1989). Importantly, Feng
et al. pay particular attention to the hidden layers of the neural network, arguing
that these can be understood to generate deep-learning risk factors from the firm
characteristics that serve as inputs. Furthermore, they propose an activation function
for the estimation of long-short portfolio weights. Whilst similar to IPCA and RP-PCA,
an important difference between PCA-based approaches and that by Feng et al. is that
the application of a neural network allows for nonlinearities on the firm characteristics
instead of extracting linear components.
Whilst Lettau and Pelger (2020b) impose sparsity regarding the number of factors
(but notably not regarding the number of characteristics from which these factors
might be constructed), there are other studies which assume a sparse representation in
terms of characteristics. For example, DeMiguel, Martín-Utrera, Uppal and Nogales
(2020) are concerned with the issue of factor selection for portfolio optimization. They
incorporate transaction costs into their consideration and – using 𝐿 1 -norm penalties
– find that the number of relevant characteristics increases compared to a setting
without transaction costs. Also focusing on individual characteristics, Freyberger et
al. (2020) address the need to differentiate between factor candidates that contain
incremental information about average returns and others, with no such independent
informational content. They propose to use a group LASSO for the purpose of
selecting characteristics and use these characteristics in a nonparametric setup to
model the cross-section of expected returns, thereby avoiding strong functional form
9 An overview of recent econometric developments in factor model literature is provided by Fan, Li
and Liao (2021).
assumptions. Freyberger et al. assemble a data set of 62 characteristics and find that
only 9 to 16 of these provide incremental information in the presence of the other
factors and that the predictive power of characteristics is strongly time-varying.
Giglio and Xiu (2021) propose a three-step approach that invokes PCA to deal
with the issue of omitted factors in asset pricing models. They argue that standard
estimators of risk premia in linear asset pricing models are biased if some priced
factors are omitted, and that their method can correctly recover the risk premium of
any observable factor in such a setting. The approach augments the standard two-pass
regression method by PCA, such that the first step of the procedure is the construction
of principal components of test asset returns to recover the factor space. In a related
study, Giglio, Liao and Xiu (2021) address the challenges of multiple testing of many
alphas in linear factor models. With multiple testing, there is a danger of producing a
high number of false positive results just by chance. Additionally, omitted factors – a
frequent concern with linear asset pricing models – hamper existing false discovery
control approaches (e.g., Benjamini & Hochberg, 1995). Giglio, Liao and Xiu develop
a framework that exploits various machine learning methods to deal with omitted
factors and missing data (using matrix completion) and provide asymptotic theory
required for their proposed estimation and testing strategy.
Using Bayesian techniques that shrink weak factors according to their correlation
with asset returns, Bryzgalova, Huang and Julliard (2021) put forward a unified
framework for analyzing linear asset pricing models, which allows for traded and
non-traded factors and possible model misspecification. Their approach can be applied
to the entire factor zoo and is able to identify the dominant model specification – or,
in the absence of such – will resort to Bayesian model averaging. Considering about
2.25 quadrillion models, Bryzgalova, Huang and Julliard find that there appears to be
no unique best model specification. Instead, hundreds of different linear factor models
exhibit almost equivalent performance. They also find that only a small number of
factors robustly describe the cross-section of asset returns. Pelger and Xiong (2020)
combine nonparametric kernel projection with PCA and develop an inferential theory
for state-varying factor models of a large cross-sectional and time-series dimension.
In their study, factor loadings are functions of the state-process, thus increasing
the model’s flexibility and, compared with constant factor models, making it more
parsimonious regarding the number of factors required to explain the same variation
in the data. They derive asymptotic results and develop a statistical test for changes
in the factor loadings in different states. Applying their method to U.S. data, Pelger
and Xiong (2020) find that the factor structures of U.S. Treasury yields and S&P 500
stock returns exhibit a strong time variation.
In work related to Kelly et al. (2019), Kim, Korajczyk and Neuhierl (2021) separate
firm characteristics’ ability to explain the cross-section of asset returns into a risk
component (factor loadings) and a mispricing component, whilst allowing for a
time-varying functional relationship – a notable difference to IPCA, where the time
variation of factor loadings results from changing characteristics. To do so, they
extend the projected principal components approach proposed by Fan, Liao and Wang
(2016). Applying their technique to U.S. equity data, they find that firm characteristics
are informative regarding potential mispricing of stocks.
356 Sönksen
10.5 How Machine Learning Can Predict in Empirical Asset

Pricing
Assessing return predictability has always been an active pursuit in finance. A

comprehensive survey of studies that provide return predictions using – to a large extent
– more traditional econometric approaches is provided by Rapach and Zhou (2013).
Lewellen (2015) assesses return predictability using Fama-MacBeth regressions.
With the advent of the second generation of machine learning methods in empirical
finance, predictability literature experienced another boost.
Because this chapter centers on machine learning in the context of asset pricing,
it makes sense to start the discussion of return predictability with the reformulated
basic asset pricing Equation (10.4). As mentioned in the introduction, the conditional
𝑖 ) could be conceived of as a function of time 𝑡 variables,
covariance cov𝑡 (𝑚 𝑡+1 , 𝑅𝑡+1
and one could consider flexible functional forms and models to provide MSE-optimal
excess return forecasts. This notion provides the starting point for Gu, Kelly and
Xiu (2020), who examine a variety of machine learning methods, including artificial
neural networks, random forests, gradient-boosted regression trees, and elastic nets
– to provide those flexible functional forms. They are prudent to avoid overfitting
by working out a dynamic training and validation scheme. Considering a vast stock
universe that also includes penny stocks, the authors find that feedforward networks
are particularly well suited for excess return prediction at the one-month investment
horizon. Grammig, Hanenberg, Schlag and Sönksen (2021) compare the data-
driven techniques considered by Gu et al. (2020) with option-based approaches for
approximating stock risk premia (and thus MSE-optimal forecasts from a theoretical
point of view). The latter are based on financial economic theory and advocated by
Martin and Wagner (2019). Grammig et al. (2021) also assess the potential of hybrid
strategies that employ machine learning algorithms to pin down the approximation
error inherent in the theory-based approach. Their results indicate that random
forests in particular offer great potential for improving the performance of the pure
theory-based approach.
The IPCA approach outlined in Section 10.4 also can serve prediction purposes.
Kelly, Moskowitz and Pruitt (2021) use IPCA focusing on momentum and address the
question to what extent the momentum premium can be explained by time-varying risk
exposure. They show that stock momentum (and other past return characteristics that
predict future returns) help predict future realized betas, but that momentum no longer
significantly contributes to understanding conditional expected stock returns once the
conditional factor risk channel has been accounted for. Furthermore, Büchner and
Kelly (2022) adopt IPCA to predict option returns. Compared with other asset classes,
options pose particular challenges for return prediction, because their short lifespans
and rapidly changing characteristics, such as moneyness (i.e., the intrinsic value of an
option in its current state), make them ill-suited for off-the-shelf methods. Using IPCA
in this context allows to view all attributes of an option contract as pricing-relevant
characteristics that translate into time-varying latent factor loadings. Moreover, Kelly,
Palhares and Pruitt (2021) exploit IPCA to model corporate bond returns using a
five-factor model and time-varying factor loadings. They find that this approach
outperforms competing empirical strategies for bond return prediction. With another
assessment of bond return predictability, Bianchi, Büchner and Tamoni (2021) report
that tree-based approaches and neural networks, two highly nonlinear approaches,
prove particularly useful for this purpose. Studying the neural network forecasts in
more detail, Bianchi et al. further find these forecasts to exhibit countercyclicality
and to be correlated with variables that proxy for macroeconomic uncertainty and
(time-varying) risk aversion.
Wu, Chen, Yang and Tindall (2021) use different machine learning approaches for
cross-sectional return predictions pertaining to hedge fund selections, using features
that contain hedge fund-specific information. Like Gu et al. (2020), they find that
the flexibility of feedforward neural networks is particularly well suited for return
predictions. Wu et al. (2021) note that their empirical strategy outperforms prominent
hedge fund research indices almost constantly. Guijarro-Ordonez, Pelger and Zanotti
(2021) seek an optimal trading policy by exploiting temporal price differences among
similar assets. In addition to proposing a framework for such statistical arbitrage,
they use a convolutional neural network combined with a transformer to detect
commonalities and time-series patterns from large panels. The optimal trading
strategy emerging from these results outperforms competing benchmark approaches
and delivers high out-of-sample Sharpe ratios; the profitability of arbitrage trading
appears to be non-declining over time. Cong, Tang, Wang and Zhang (2021) apply deep
reinforcement learning directly to optimize the objectives of portfolio management,
instead of pursuing a traditional two-step approach (i.e., training machine learning
techniques for return prediction first, then translating the results into portfolio
management considerations), such that they employ multisequence, attention-based
neural networks. Studying return predictability at a high frequency, Chinco, Clark-
Joseph and Ye (2019) consider the entire cross-section of lagged returns as potential
predictors. They find that making rolling one-minute-ahead return forecasts using
LASSO to identify a sparse set of short-lived predictors increases both out-of-sample
fit and Sharpe ratios. Further analysis reveals that the selected predictors tend to be
related to stocks with news about fundamentals.
The application of machine learning techniques to empirical asset pricing also
encompasses unstructured or semi-structured data. For example, Obaid and Pukthu-
anthong (2022) predict market returns using the visual content of news. For this
purpose, they develop a daily market-level investor sentiment index that is defined as
the fraction of negative news photos and computed from a large sample of news media
images provided by the Wall Street Journal. The index negatively predicts next day’s
market returns and captures a reversal in subsequent days. Obaid and Pukthuanthong
compare their index with an alternative derived from text data and find that both
options appear to contain the same information and serve as substitutes. Relying on
news articles distributed via Dow Jones Newswires, Ke, Kelly and Xiu (2021) also
use textual data, but their focus is on the prediction of individual stock returns. To this
end, they propose a three-step approach to assign sentiment scores on an article basis
without requiring the use of pre-existing sentiment dictionaries. Ke et al. find that
their proposed text-mining approach detects a reasonable return predictive signal in
358 Sönksen
the data and outperforms competing commercial sentiment indices. Jiang, Kelly and
Xiu (2021) apply convolutional neural networks (CNN) to images of stock-level price
charts to extract price patterns that provide better return predictions than competing
trend signals. They argue that the detected trend-predictive signals generalize well to
other time-scales and also to other financial markets. In particular, Jiang et al. transfer
a CNN trained using U.S. data to 26 different international markets; in 19 of these 26
cases, the transferred model produced higher Sharpe ratios than models trained on
local data. This result implies that the richness of U.S. stock market data could be
meaningfully exploited to assist the analysis of other financial markets with fewer
stocks and shorter time series.
Avramov, Cheng and Metzker (2021) offer a skeptical view of predictability. Their
objective is not about finding the best method for trading purposes or about conducting
a comprehensive comparison between different machine learning approaches. Rather,
the authors aim at assessing whether machine learning approaches can be successfully
applied for the purpose of a profitable investment strategy under realistic economic
restrictions, e.g., as they relate to transaction costs. In doing so, Avramov et al.
build on prior work by Gu et al. (2020), Chen et al. (2021), Kelly et al. (2019),
and Gu et al. (2021) and compare the performance of these approaches on the full
universe of stocks considered in each of the studies, respectively, to that obtained
after reducing the stock universe by microcaps, firms without credit rating coverage,
and financially distressed firms. They find that the predictability of deep learning
methods deteriorates strongly whilst that of the IPCA approach, which assumes a
linear relationship between firm characteristics and stock returns and is outperformed
by the other techniques when evaluated on the full universe of stocks, is somewhat less
affected. This leads the authors to conclude that the increased flexibility of the deep
learning algorithms proves especially useful for the prediction of difficult-to-value
and difficult-to-arbitrage stocks. Further analysis reveals that the positive predictive
performance hinges strongly on periods of low market liquidity. Additionally, all
machine learning techniques under consideration imply a high portfolio turnover,
which would be rather expensive given realistic trading costs. However, Avramov
et al. (2021) also find that machine learning-based trading strategies display less
downside risk and yield considerable profit in long positions. Machine learning
methods successfully identify mispriced stocks consistent with most anomalies and
remain viable in the post-2001 period, during which traditional anomaly-based trading
strategies weaken.
Brogaard and Zareei (2022) are also concerned with evaluating whether machine
learning algorithms’ predictive strengths translate into a realistic setting that includes
transaction costs. In particular, they use machine learning techniques to contribute
to the discussion whether technical trading rules can be profitably applied by
practitioners; specifically, whether a practitioner could have identified a profitable
trading rule ex ante. This notion is contradicted by the efficient market hypothesis
according to which stock prices contain all publicly available information. One key
difference between Brogaard and Zareei’s (2022) study and that by Avramov et
al. (2021) is that the former focus on the identification of profitable trading rules.
Such trading rules are manifold – trading based on momentum constitutes one of
them – and they are easy to implement. In contrast, the approaches considered by
Avramov et al. use vast sets of firm characteristics that are in some cases related
to trading rules, but not necessarily so. Exploiting a broad set of different machine
learning techniques that include both evolutionary genetic algorithms and standard
loss-minimizing approaches, Brogaard and Zareei search for profitable trading rules.
Controlling for data-snooping and transaction costs, they indeed identify such trading
rules, but find their out-of-sample profitability to be decreasing over time.
The Nobel Prize in Economics shared by E. Fama, L.P. Hansen, and R. Shiller in
2013 made it clear that empirical asset pricing represents a highly developed field
within the economics discipline, with contributions that help define the forefront of
empirical method development. A rich data environment – both quantitatively and
qualitatively – and the interest from both academia and practice provides a fertile
ground.
The notable spread of data-intensive machine learning techniques in recent
years shows that research in empirical asset pricing remains as innovative and
vibrant as ever. It is noteworthy that the adoption of the methods did not entail just
improved measurement without theory. Instead, the integration of machine learning
techniques has proceeded in such a way that unresolved methodological issues
have been addressed creatively. This chapter provides a review of prominent, recent
contributions, thus revealing how machine learning techniques can help identify the
stochastic discount factor, the elusive object of core interest in asset pricing, as well as
how pertinent methods can be employed to improve efforts to test and evaluate asset
pricing approaches, and how the development of conditional factor models benefits
from the inclusion of machine learning techniques. It also showed how machine
learning can be applied for asset return prediction, keeping in mind the limitations
implied by theory and the challenges of a low signal-to-noise environment.
Appendix 1: An Upper Bound for the Sharpe Ratio
The upper bound of the Sharpe ratio can be derived directly from the basic asset
pricing equation. Assume that a law of total expectation has been applied to Equation
(10.3), such that the expectation is conditioned down to:
𝑒
E[𝑚 𝑡+1 𝑅𝑡+1 ] =0
𝑒 𝑒
cov[𝑚 𝑡+1 , 𝑅𝑡+1 ] + E[𝑚 𝑡+1 ]E[𝑅𝑡+1 ] =0
𝑒 𝑒
E[𝑚 𝑡+1 ]E[𝑅𝑡+1 ] = −cov(𝑚 𝑡+1 , 𝑅𝑡+1 )
𝑒 𝑒 𝑒
E[𝑚 𝑡+1 ]E[𝑅𝑡+1 ] = −𝜌(𝑚 𝑡+1 , 𝑅𝑡+1 )𝜎(𝑚 𝑡+1 )𝜎(𝑅𝑡+1 ),
360 Sönksen
where 𝜎(·) refers to the standard deviation, and 𝜌(·, ·) denotes a correlation. Further
reformulation yields:
𝑒 ]
E[𝑅𝑡+1 𝜎(𝑚 𝑡+1 ) 𝑒
𝑒 ) =− 𝜌(𝑚 𝑡+1 , 𝑅𝑡+1 ).
𝜎(𝑅𝑡+1 E[𝑚 𝑡+1 ]
Because |𝜌(·, ·)| ≤ 1, and E[𝑚 𝑡+1 ] > 0, we obtain:
|E[𝑅𝑡+1
𝑒 ]|
𝜎(𝑚 𝑡+1 )
𝑒 ) ≤ .
𝜎(𝑅𝑡+1 E[𝑚 𝑡+1 ]
Appendix 2: A Comparison of Different PCA Approaches
For this overview of key differences among the various PCA-related techniques
mentioned in the chapter, denote as X an (𝑁 × 𝑇) matrix of excess returns, F is an
(𝑇 × 𝐾) matrix of (latent) factors, X̃ and F̃ are the demeaned counterparts, and Z is a
(𝑇 × 𝑃) matrix of (observed) characteristics. I assume the general factor structure:
X = F𝚲′ + e
and try to identify F and 𝚲 from it.
Principal Component Analysis (PCA)

With conventional PCA, factor loadings 𝚲 are obtained by using PCA directly on
the sample variance-covariance matrix of X, which is 𝑇1 X′X − X̄′X̄, where X̄ denotes
sample means. Regressing X on 𝚲 ˆ results in estimates of the factors. Alternatively,
estimates of 𝚲 and F̃ might derive from
𝑁 𝑇
1 ∑︁ ∑︁ ˜ 2
min 𝑋𝑡 ,𝑖 − 𝐹˜𝑡′Λ𝑖 .
𝚲, F̃ 𝑁𝑇 𝑖=1 𝑡=1
Risk-Premium Principal Component Analysis (RP-PCA)

Lettau and Pelger (2020a) extend the conventional PCA by overweighting the mean,
such that they apply PCA not to 𝑇1 X′X − X̄′X̄ but instead to 𝑇1 X′X + 𝛾 X̄′X̄, where 𝛾
can be interpreted as a penalty parameter. To see this, note that we could alternatively
consider the problem:
𝑁 𝑇 𝑁
1 ∑︁ ∑︁ ˜ 2 1 ∑︁ ¯ 2
min 𝑋𝑡 ,𝑖 − 𝐹˜𝑡′Λ𝑖 + (1 + 𝛾) 𝑋𝑖 − 𝐹¯ ′Λ𝑖 .
𝚲, F̃ 𝑁𝑇 𝑖=1 𝑡=1 𝑁 𝑖=1
The first part of the objective function deals with minimizing unexplained variation,
and the second part imposes a penalty on pricing errors.
References 361
Projected Principal Component Analysis

Fan et al. (2016) consider daily data and propose, instead of applying PCA to the
variance-covariance matrix of X, projecting X on a set of time-invariant
asset-specific characteristics Z. Then the PCA would be applied to the
variance-covariance matrix of the thus smoothed X̂.
Instrumented Principal Component Analysis (IPCA)

The IPCA approach introduced by Kelly et al. (2019) is described in detail in
Section 10.4.2 of this Chapter. It is related to projected principal component analysis,
but Kelly et al. (2019) consider time-varying instruments – an important difference,
because it translates into time-varying factor loadings.
Supervised Principal Component Analysis (SPCA)

Giglio, Xiu and Zhang (2021) introduce supervised principal component analysis
to counteract issues associated with weak factors. The term supervised comes from
supervised machine learning. The authors propose that, rather than using PCA on the
variance-covariance matrix of the entire X, they should select stocks that exhibit a
strong correlation (in absolute value) with the instruments in Z. So, instead of using
all 𝑁 stocks, Giglio, Xiu and Zhang (2021) limit themselves to a fraction 𝑞 of the
sample and select those stocks with the strongest correlation:

1
𝐼ˆ = 𝑖 X̃ [𝑖 ] Z̃′ ≥ 𝑐 𝑞 ,

𝑇
where 𝑐 𝑞 denotes the (1 − 𝑞) correlation quantile. Then the PCA can be conducted
on X̃ [ 𝐼ˆ] .
References
Ahn, D.-H., Conrad, J. & Dittmar, R. F. (2009). Basis Assets. The Review of
Financial Studies, 22(12), 5133–5174. Retrieved from http://www.jstor.org/
stable/40468340
Anatolyev, S. & Mikusheva, A. (2022). Factor Models with Many Assets: Strong
Factors, Weak Factors, and the Two-Pass Procedure. Journal of Econometrics,
229(1), 103–126. Retrieved from https://www.sciencedirect.com/science/
article/pii/S0304407621000130 doi: 10.1016/j.jeconom.2021.01.002
Ang, A., Liu, J. & Schwarz, K. (2020). Using Stocks or Portfolios in Tests of Factor
Models. Journal of Financial and Quantitative Analysis, 55(3), 709–750. doi:
10.1017/S0022109019000255
Avramov, D., Cheng, S. & Metzker, L. (2021). Machine Learning versus Economic
Restrictions: Evidence from Stock Return Predictability. Retrieved from
http://dx.doi.org/10.2139/ssrn.3450322 (forthcoming: Management Science)
362 Sönksen
Bai, J. & Ng, S. (2002). Determining the Number of Factors in Approximate Factor
Models. Econometrica, 70(1), 191–221. doi: 10.1111/1468-0262.00273
Bailey, N., Kapetanios, G. & Pesaran, M. H. (2021). Measurement of Factor Strength:
Theory and Practice. Journal of Applied Econometrics, 36(5), 587–613. doi:
10.1002/jae.2830
Bakalli, G., Guerrier, S. & Scaillet, O. (2021). A Penalized Two-Pass Regression
to Predict Stock Returns with Time-Varying Risk Premia. Retrieved from
http://dx.doi.org/10.2139/ssrn.3777215 (Working Paper, accessed February
23, 2022)
Benjamini, Y. & Hochberg, Y. (1995). Controlling the False Discovery Rate: A
Practical and Powerful Approach to Multiple Testing. Journal of the Royal
Statistical Society: Series B (Methodological), 57(1), 289–300. Retrieved
from https://rss.onlinelibrary.wiley.com/doi/abs/10.1111/j.2517-6161.1995
.tb02031.x doi: https://doi.org/10.1111/j.2517-6161.1995.tb02031.x
Bianchi, D., Büchner, M. & Tamoni, A. (2021). Bond Risk Premiums with
Machine Learning. The Review of Financial Studies, 34(2), 1046–1089. doi:
10.1093/rfs/hhaa062
Brogaard, J. & Zareei, A. (2022). Machine Learning and the Stock Market.
Retrieved from http://dx.doi.org/10.2139/ssrn.3233119 (forthcoming: Journal
of Financial and Quantitative Analysis)
Bryzgalova, S., Huang, J. & Julliard, C. (2021). Bayesian Solutions for the Factor
Zoo: We Just Ran Two Quadrillion Models. Retrieved from http://dx.doi.org/
10.2139/ssrn.3481736 (Working Paper, accessed February 22, 2022)
Bryzgalova, S., Pelger, M. & Zhu, J. (2021a). Forest through the Trees: Building
Cross-Sections of Stock Returns. Retrieved from http://dx.doi.org/10.2139/
ssrn.3493458 (Working Paper, accessed February 23, 2022)
Bryzgalova, S., Pelger, M. & Zhu, J. (2021b). Internet Appendix for Forest
through the Trees: Building Cross-Sections of Stock Returns. Retrieved from
https://ssrn.com/abstract=3569264
Büchner, M. & Kelly, B. (2022). A Factor Model for Option Returns. Journal
of Financial Economics, 143(3), 1140–1161. Retrieved from https://www
.sciencedirect.com/science/article/pii/S0304405X21005249 doi: 10.1016/
j.jfineco.2021.12.007
Chen, L., Pelger, M. & Zhu, J. (2021). Deep Learning in Asset Pricing. Retrieved
from http://dx.doi.org/10.2139/ssrn.3350138 (Working Paper, accessed July
27, 2021)
Chinco, A., Clark-Joseph, A. D. & Ye, M. (2019). Sparse Signals in the Cross-Section
of Returns. The Journal of Finance, 74(1), 449–492. doi: 10.1111/jofi.12733
Cochrane, J. (1996). A Cross-Sectional Test of an Investment-Based Asset Pricing
Model. Journal of Political Economy, 104(3), 572–621. Retrieved from
Cochrane, J. (2005). Asset Pricing. Princeton University Press, New Jersey.
Cong, L. W., Tang, K., Wang, J. & Zhang, Y. (2021). AlphaPortfolio: Direct
Construction Through Deep Reinforcement Learning and Interpretable AI.
Retrieved from http://dx.doi.org/10.2139/ssrn.3554486 (Working Paper,
References 363
accessed February 23, 2022)

DeMiguel, V., Martín-Utrera, A., Uppal, R. & Nogales, F. (2020, 1st). A Transaction-
Cost Perspective on the Multitude of Firm Characteristics. Review of Financial
Studies, 33(5), 2180–2222. Retrieved from https://academic.oup.com/rfs/
article-abstract/33/5/2180/5821387 doi: 10.1093/rfs/hhz085
Duffie, D. & Singleton, K. J. (1993). Simulated Moments Estimation of Markov
Models of Asset Prices. Econometrica, 61(4), 929–952. doi: 10.2307/2951768
Fama, E. F. & French, K. R. (1993). Common Risk Factors in the Returns on Stocks
and Bonds. Journal of Financial Economics, 33(1), 3–56. Retrieved from
https://www.sciencedirect.com/science/article/pii/0304405X93900235 doi:
10.1016/0304-405X(93)90023-5
Fan, J., Li, K. & Liao, Y. (2021). Recent Developments in Factor Models and
Applications in Econometric Learning. Annual Review of Financial Economics,
13(1), 401–430. doi: 10.1146/annurev-financial-091420-011735
Fan, J., Liao, Y. & Wang, W. (2016). Projected Principal Component Analysis in Factor
Models. The Annals of Statistics, 44(1), 219 – 254. doi: 10.1214/15-AOS1364
Feng, G., Giglio, S. & Xiu, D. (2020). Taming the Factor Zoo: A Test of New Factors.
The Journal of Finance, 75(3), 1327–1370. doi: 10.1111/jofi.12883
Feng, G., Polson, N. & Xu, J. (2021). Deep Learning in Characteristics-Sorted Factor
Models. Retrieved from http://dx.doi.org/10.2139/ssrn.3243683 (Working
Paper, accessed February 22, 2022)
Freyberger, J., Neuhierl, A. & Weber, M. (2020). Dissecting Characteristics
Nonparametrically. The Review of Financial Studies, 33(5), 2326–2377. doi:
10.1093/rfs/hhz123
Gagliardini, P., Ossola, E. & Scaillet, O. (2016). Time-Varying Risk Premium in
Large Cross-Sectional Equity Data Sets. Econometrica, 84(3), 985–1046. doi:
10.3982/ECTA11069
Gagliardini, P., Ossola, E. & Scaillet, O. (2019). A Diagnostic Criterion for
Approximate Factor Structure. Journal of Econometrics, 212(2), 503–521. doi:
10.1016/j.jeconom.2019.06.001
Gibbons, M. R., Ross, S. A. & Shanken, J. (1989). A Test of the Efficiency
of a Given Portfolio. Econometrica, 57(5), 1121–1152. Retrieved from
Giglio, S., Kelly, B. T. & Xiu, D. (2022). Factor Models, Machine Learning, and Asset
Pricing. Retrieved from http://dx.doi.org/10.2139/ssrn.3943284 (forthcoming:
Annual Review of Financial Economics)
Giglio, S., Liao, Y. & Xiu, D. (2021). Thousands of Alpha Tests. The Review of
Financial Studies, 34(7), 3456–3496. doi: 10.1093/rfs/hhaa111
Giglio, S. & Xiu, D. (2021). Asset Pricing with Omitted Factors. Journal of Political
Economy, 129(7), 1947–1990. doi: 10.1086/714090
Giglio, S., Xiu, D. & Zhang, D. (2021). Test Assets and Weak Factors. Retrieved from
23, 2022)
Gospodinov, N., Kan, R. & Robotti, C. (2014). Misspecification-Robust Inference
in Linear Asset-Pricing Models with Irrelevant Risk Factors. The Review of
364 Sönksen
Financial Studies, 27(7), 2139–2170. doi: 10.1093/rfs/hht135

Grammig, J., Hanenberg, C., Schlag, C. & Sönksen, J. (2021). Diverging Roads:
Theory-Based vs. Machine Learning-Implied Stock Risk Premia. Retrieved from
22, 2022)
Gu, S., Kelly, B. & Xiu, D. (2020). Empirical Asset Pricing via Machine Learning.
The Review of Financial Studies, 33(5), 2223–2273. doi: 10.1093/rfs/hhaa009
Gu, S., Kelly, B. & Xiu, D. (2021). Autoencoder Asset Pricing Models. Journal of
Econometrics, 222(1), 429–450. doi: 10.1016/j.jeconom.2020.07
Guijarro-Ordonez, J., Pelger, M. & Zanotti, G. (2021). Deep Learning Statistical
Arbitrage. Retrieved from http://dx.doi.org/10.2139/ssrn.3862004 (Working
Paper, accessed July 27, 2021)
Hall, A. (2005). Generalized Method of Moments. Oxford University Press.
Hansen, L. P. (1982). Large Sample Properties of Generalized Method of Moments
Estimators. Econometrica, 50(4), 1029–1054. Retrieved from http://www.jstor
.org/stable/1912775
Hansen, L. P. & Jagannathan, R. (1991). Implications of Security Market Data for
Models of Dynamic Economies. Journal of Political Economy, 99(2), 225–262.
Hansen, L. P. & Jagannathan, R. (1997). Assessing Specification Errors in Stochastic
Discount Factor Models. The Journal of Finance, 52(2), 557–590. doi:
10.1111/j.1540-6261.1997.tb04813.x
Hansen, L. P. & Richard, S. F. (1987). The Role of Conditioning Information in
Deducing Testable Restrictions Implied by Dynamic Asset Pricing Models.
Econometrica, 55(3), 587–613. Retrieved from http://www.jstor.org/stable/
1913601
Hansen, L. P. & Singleton, K. J. (1982). Generalized Instrumental Variables
Estimation of Nonlinear Rational Expectations Models. Econometrica, 50(2),
1296–1286.
Harvey, C. R., Liu, Y. & Zhu, H. (2015). ...and the Cross-Section of Expected Returns.
The Review of Financial Studies, 29(1), 5–68. doi: 10.1093/rfs/hhv059
Jegadeesh, N., Noh, J., Pukthuanthong, K., Roll, R. & Wang, J. (2019). Empirical
Tests of Asset Pricing Models with Individual Assets: Resolving the Errors-in-
Variables Bias in Risk Premium Estimation. Journal of Financial Economics,
article/pii/S0304405X19300431 doi: 10.1016/j.jfineco.2019.02.010
Jiang, J., Kelly, B. T. & Xiu, D. (2021). (Re-)Imag(in)ing Price Trends. Retrieved from
22, 2022)
Ke, Z., Kelly, B. T. & Xiu, D. (2021). Predicting Returns with Text Data. Re-
trieved from http://dx.doi.org/10.2139/ssrn.3389884 (Working Paper, accessed
February 22, 2022)
Kelly, B. T., Moskowitz, T. J. & Pruitt, S. (2021). Understanding Momentum and
Reversal. Journal of Financial Economics, 140(3), 726–743. Retrieved from
https://www.sciencedirect.com/science/article/pii/S0304405X21000878 doi:
References 365
10.1016/j.jfineco.2020.06.024
Kelly, B. T., Palhares, D. & Pruitt, S. (2021). Modeling Corporate Bond Returns.
accessed February 22, 2022)
Kelly, B. T., Pruitt, S. & Su, Y. (2019). Characteristics are Covariances: A
Unified Model of Risk and Return. Journal of Financial Economics, 134(3),
501–524. Retrieved from https://www.sciencedirect.com/science/article/pii/
S0304405X19301151 doi: https://doi.org/10.1016/j.jfineco.2019.05.001
Kelly, B. T., Pruitt, S. & Su, Y. (2020). Instrumented Principal Component
Analysis. Retrieved from http://dx.doi.org/10.2139/ssrn.2983919 (Working
Paper, accessed February 23, 2022)
Kim, S., Korajczyk, R. A. & Neuhierl, A. (2021). Arbitrage Portfolios. The Review
of Financial Studies, 34(6), 2813–2856. doi: 10.1093/rfs/hhaa102
Korsaye, S. A., Quaini, A. & Trojani, F. (2019). Smart SDFs. Retrieved from
22, 2022)
Kozak, S., Nagel, S. & Santosh, S. (2018). Interpreting Factor Models. The Journal
of Finance, 73(3), 1183–1223. doi: 10.1111/jofi.12612
Kozak, S., Nagel, S. & Santosh, S. (2020). Shrinking the Cross-Section. Journal
of Financial Economics, 135(2), 271–292. Retrieved from https://www
.sciencedirect.com/science/article/pii/S0304405X19301655 doi: 10.1016/
j.jfineco.2019.06.008
Lettau, M. & Ludvigson, S. (2001). Resurrecting the (C)CAPM: A Cross-Sectional
Test When Risk Premia Are Time-Varying. Journal of Political Economy,
109(6), 1238–1287. Retrieved from http://www.jstor.org/stable/10.1086/
323282
Lettau, M. & Pelger, M. (2020a). Estimating Latent Asset-Pricing Factors. Journal
of Econometrics, 218(1), 1–31. Retrieved from https://www.sciencedirect
.com/science/article/pii/S0304407620300051 doi: https://doi.org/10.1016/
j.jeconom.2019.08.012
Lettau, M. & Pelger, M. (2020b). Factors That Fit the Time Series and Cross-Section
of Stock Returns. The Review of Financial Studies, 33(5), 2274–2325. doi:
10.1093/rfs/hhaa020
Lewellen, J. (2015). The Cross-Section of Expected Stock Returns. Critical Finance
Review, 4(1), 1–44. doi: 10.1561/104.00000024
Lewellen, J., Nagel, S. & Shanken, J. (2010). A Skeptical Appraisal of Asset
Pricing Tests. Journal of Financial Economics, 96(2), 175–194. Retrieved
from https://www.sciencedirect.com/science/article/pii/S0304405X09001950
doi: https://doi.org/10.1016/j.jfineco.2009.09.001
Lintner, J. (1965). The Valuation of Risk Assets and the Selection of Risky
Investments in Stock Portfolios and Capital Budgets. The Review of Economics
and Statistics, 47, 13–37.
Martin, i. W. R. & Wagner, C. (2019). What Is the Expected Return on a Stock? The
Journal of Finance, 74(4), 1887–1929. doi: https://doi.org/10.1111/jofi.12778
366 Sönksen
Mossin, J. (1966). Equilibrium in a Capital Asset Market. Econometrica, 34,

768–783.
Nagel, S. (2021). Machine Learning in Asset Pricing. Princeton University Press,
New Jersey.
Obaid, K. & Pukthuanthong, K. (2022). A Picture is Worth a Thousand Words:
Measuring Investor Sentiment by Combining Machine Learning and Photos
from News. Journal of Financial Economics, 144(1), 273–297. Retrieved from
https://www.sciencedirect.com/science/article/pii/S0304405X21002683 doi:
https://doi.org/10.1016/j.jfineco.2021.06.002
Onatski, A. (2012). Asymptotics of the Principal Components Estimator of Large
Factor Models with Weakly Influential Factors. Journal of Econometrics,
article/pii/S0304407612000449 doi: https://doi.org/10.1016/j.jeconom.2012
.01.034
Pelger, M. & Xiong, R. (2020). State-Varying Factor Models of Large Dimensions.
accessed July 27, 2021)
Pukthuanthong, K., Roll, R. & Subrahmanyam, A. (2018). A Protocol for Factor
Identification. The Review of Financial Studies, 32(4), 1573–1607. doi:
10.1093/rfs/hhy093
Rapach, D. & Zhou, G. (2013). Chapter 6 - Forecasting Stock Returns. In
G. Elliott & A. Timmermann (Eds.), Handbook of Economic Forecasting
(Vol. 2, pp. 328–383). Elsevier. Retrieved from https://www.sciencedirect.com/
science/article/pii/B9780444536839000062 doi: https://doi.org/10.1016/
B978-0-444-53683-9.00006-2
Raponi, V., Robotti, C. & Zaffaroni, P. (2019). Testing Beta-Pricing Models Using
Large Cross-Sections. The Review of Financial Studies, 33(6), 2796–2842.
doi: 10.1093/rfs/hhz064
Sharpe, W. F. (1964). Capital Asset Prices: A Theory of Market Equilibrium
under Conditions of Risk. The Journal of Finance, 19(3), 425-–442. doi:
10.1111/j.1540-6261.1964.tb02865.x
Singleton, K. J. (2006). Empirical Dynamic Asset Pricing. Princeton University
Press, New Jersey.
Weigand, A. (2019). Machine Learning in Empirical Asset Pricing. Financial Markets
and Portfolio Management, 33(1), 93–104. doi: 10.1007/s11408-019-00326-3
Wu, W., Chen, J., Yang, Z. B. & Tindall, M. L. (2021). A Cross-Sectional
Machine Learning Approach for Hedge Fund Return Prediction and Selection.
Management Science, 67(7), 4577–4601. doi: 10.1287/mnsc.2020.3696
Appendix A
Terminology
A.1 Introduction
As seen in this volume, terminologies used in the Machine Learning literature often
have equivalent counterparts in statistics and/or econometrics. These differences in
terminologies create potential hurdles in making machine learning methods more
readily applicable in econometrics. The purpose of this Appendix is to provide a list
of terminologies typically used in machine learning and explain them in the language
of econometrics or, at the very least, explain in a language that is more familiar to
economists or econometricians.
A.2 Terms
Adjacency matrix. A square matrix which can represent nodes and edges in a
network. See Chapter 6.
Agents. Entities in an econometric analysis (e.g., individuals, households, firms).
Bagging. Using Bootstrapping for selecting (multiple) data and AGGregating the
results.
Boosting. Model averaging: sequentially learn from models: use residuals from each
models and fit data to them in a systematic way.
Bots. A program that interacts with systems or users.
Causal forest. A causal forest estimator averages several individual causal trees
grown on different samples (just as a standard random forest averages several
individual regression trees). See Chapter 3.
Causal tree. A regression tree used to estimate heterogeneous treatment effects. See
Chapter 3.
in Theoretical and Applied Econometrics 53, https://doi.org/10.1007/978-3-031-15149-1
368 A Terminology
Classification problem. Modelling/Predicting discrete nominal outcomes, sometimes

the random variables are ordinal but not cardinal. See Chapter 2.
Clustering. Algorithms for grouping data points.
Confusion matrix. A 2 × 2 matrix measuring the prediction performance of a binary
choice model. See Chapter 2.
Convolutional Neural Network. A class of neural networks with an architecture
that has at least one convolution layer, which consists of performing matrix products
between the matrix in the input or previous layer and other matrices called convolution
filters.
Covariates. Explanatory Variables. See Chapter 1.
Cross-fitting. A variant of the DML procedure in which the sample is used in a
similar way as in a cross-validation exercise. See Chapter 3.
Cross-validation. Method used to estimated the tuning parameter (based on dividing
the sample randomly into partitions and randomly select from these to estimate the
tuning parameter). See Chapter 1.
Crowd-sourced data. Data collected with a participatory method, with the help of a
large group of people.
Debiased Machine Learning (DML). A way to employ ML methods to estimate
average treatment effect parameters without regularization bias. See Chapter 3.
Deep learning. Subfield of ML grouping algorithms organized in layers (an input
layer, an output layer, and hidden layers connecting the input and output layers).
Dimensionality reduction/embedding. Algorithms which aim at representing the
data in a lower dimensional space while retaining useful information.
Domain generalization. Also known as out-of-distribution generalization. A set of
techniques aiming to create the best prediction models on random variables which
joint distributions are different, but related, to those in the training set.
Double Machine Learning (DML). Alternative terminology for debiased ML. The
adjective ’double’ is due to the fact that in some simple settings, such as the partially
linear regression model, debiased parameter estimation requires the implementation
of two ML procedures. See Chapter 3.
Edge. Connection between nodes in a network. See Chapter 6.
Elastic Net. A Regularised Regression Method which regularizer is an affine combin-
ation of LASSO and Ridge. See Chapter 1.
Ensemble method. Combine predictions from different models. A form of models
averaging.
Feature. Covariate or explanatory variable.
A.2 Terms 369
Feature engineering. Transforming variables and manipulating (imputing) observa-

tions.
Feature extraction. Obtaining features from the raw data that will be used for
preditction.
Forest. A collection of trees. See Chapter 2.
Graph Laplacian eigenmaps. A matrix factorization technique primarily used to
learn lower dimensional representations of graphs. See Chapter 6.
Graph representation learning. Assigning representation vectors to graph elements
(e.g.,: nodes, edges) containing accurate information on the structure of a large graph.
See Chapter 6.
Hierarchical bayesian geostatistical models.
Honest approach. A sample splitting approach to estimating a causal tree. One part
of the sample is used to find the model specification and the other part to estimate
the parameter(s) of interest. This method ensures unbiased parameter estimates and
standard inference. See Chapter 3.
Instance. A data point (or sometimes a small subset of the database). Can also be
called Records.
𝐾 Nearest Neighbors. Algorithms to predict a group or label based on the distance
of a data point to the other data points.
Labelled data. Each observation is linked with a particular value of the response
variable (opposite of unlabelled data).
LASSO. The Least Absolute Shrinkage and Selection Operator, or LASSO imposes a
penalty function that limits the size of its estimators by absolute value. See Chapter 1.
Learner. Model.
Loss function. A function mapping the observed data and model estimates to loss.
Learning is usually equivalent to minimizing a loss function.
Machine. Algorithm (usually used for learning).
Network/graph. A set of agents (nodes) connected by edges. See Chapter 6.
Neural network. Highly nonlinear model.
Node. Agents in a network. See Chapter 6.
Open-source data. Data publicly available for analysis and redistribution.
Overfitting. When predictions are fitted too closely to a certain set of data, resulting
in potentially failure to predict with other sets of data reliably.
Post-selection inference. Computing standard errors, confidence intervals, p-values,
etc., after using an ML method to find a model. See Chapter 3.
370 A Terminology
Principal component analysis. An algorithm which uses a linear mapping to

maximize variance in the lower dimensional space.
Predictors. Features, covariates or explanatory variables. See Chapter 1.
Prune. Methods to control the size of a tree to avoid overfitting. See Chapter 2.
Random Forest. A forest with each tree consists of random selection of regressors
(covariates). See Chapter 2.
Regularization. Estimator with regularizer. See Chapter 1.
Regularizer. Penalty function. See Chapter 1.
Scoring. Re-estimation of a model on some new/additional data. Not to be confused
with the score (vector).
Shrinkage. Restrict the magnitude of the parameter vector for purpose of identifying
the important features/covariates.
Spatiotemporal deep learning. Deep learning using spatial panel data.
Stepwise selection. Fitting regression models by adding or eliminating variables
through automatic procedure based on their significance, typically determined via
F-tests.
Supervised learning. Estimation with/of a model.
Target. Dependent variable.
Test set. Out-of-Sample (subset of data used for prediction evaluation).
Training. Estimation of a (predictive) model.
Training set. With-in-sample (subset of data used for estimation).
Transfer learning. Algorithms to store results from one ML problem and apply them
to solve other related or similar ML problems.
Tree. Can be viewed as a threshold regression model. See Chapter 2.
Tuning parameter. Parameter defining the constrain in a shrinkage estimator. See
Chapter 1.
Unlabelled data. Each observation is not linked with a particular value of the
response variable (opposite of labelled data).
Unstructured data. Data that are not organized in a predefined data format such as
tables or graphs.
Unsupervised learning. Estimation without a model. Often related to non-parametric
techniques (No designated dependent variable).
Variable importance. A visualisation showing the impact of excluding a particular
variable on the prediction performance, typically measured in MSE or Gini index.
See Chapter 2.
A.2 Terms 371
Wrapper feature selector. A method to identify relevant variables in a regression

setting using subset selection.

Sanet ST

Uploaded by

Copyright:

Available Formats

Sanet ST

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sanet ST

Uploaded by

Copyright:

Available Formats

Advanced Studies in Theoretical and Applied Econometrics 53

Econometrics with Machine

ISSN 1570-5811 ISSN 2214-7977 (electronic)

Emory University, Atlanta, USA, June 2022 Esfandiar Maasoumi

applications to portfolio selection problems via a state-space framework. Financial

Perth, June 2022 Felix Chan

1 Linear Econometric Models with Machine Learning . . . . . . . . . . . . . . 1

2.2.1 Regularization with Nonlinear Least Squares . . . . . . . . . . 44

4 Forecasting with Machine Learning Methods . . . . . . . . . . . . . . . . . . . . 111

5.12 A Final Word on the Importance of Study Design in Mitigating Bias168

8.2.2 Nodewise Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258

10 Machine Learning for Asset Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . 337

10.4.2 Kelly, Pruitt, and Su’s (2019) Instrumented Principal

Maria Victoria Anauati

Felix Chan and László Mátyás

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1

an economics/econometrics interpretation point of view. Let us call them control

1.2 Shrinkage Estimators and Regularizers

𝑦 𝑖 = x𝑖′ 𝛽 0 + 𝑢 𝑖 , 𝑢 𝑖 ∼ 𝐷 (0, 𝜎𝑢2 ), 𝑖 = 1, . . . , 𝑁, (1.1)

𝛽ˆ 𝑂𝐿𝑆 = (X′X) −1 X′y (1.3)

𝛽ˆ = arg min 𝑔 (𝛽𝛽 ; y, X) (1.5)

𝛽ˆ = arg min 𝑔 (𝛽𝛽 ; y, X) + 𝜆𝑝 (𝛽𝛽 ;𝛼

A somewhat subtle difference between the Lagrangian as defined Equation (1.7)

1.2.1 𝑳𝜸 norm, Bridge, LASSO and Ridge

A particularly interesting class of regularizers is called the Bridge estimator as defined

𝛽ˆ 𝑅𝑖𝑑𝑔𝑒 = (X′X + 𝜆I) −1 X′y. (1.9)

(a) LASSO (b) Ridge

(c) Elastic Net

Fig. 1.1: Contour plots of LASSO, Ridge and Elastic Net

When 𝛾 ≠ 2, there is no closed form solution to the associated constrained optimization

1.2.2 Elastic Net and SCAD

𝑝(𝛽𝛽 ; 𝛼) = 𝛼1 ||𝛽𝛽 || 1 + 𝛼2 ||𝛽𝛽 || 22 , 𝛼 ∈ [0, 1]. (1.10)

of (𝛼1 , 𝛼2 ) is to be determined by the researchers, along with 𝜆. Thus, Elastic Net

𝛽ˆ 𝐿 𝐴𝑆𝑆𝑂,𝑖 = sgn 𝛽ˆ𝑂𝐿𝑆,𝑖 | 𝛽ˆ𝑂𝐿𝑆,𝑖 | − 𝜆 ,

the bias determined by 𝜆, or equivalently, 𝑐. This should not be a surprise, since if

where 𝐼 ( 𝐴) is an indicator function that equals 1 if 𝐴 is true and 0 otherwise and

(a) LASSO Penalty (b) Ridge Penalty

(c) Elastic Net Penalty (d) SCAD Penalty

1.2.3 Adaptive LASSO

where 𝑤 𝑗 > 0 for all 𝑗 = 1, . . . , 𝑝 are weights to be pre-determined by the researchers.

𝛽ˆ 𝑎𝑑𝑎 = arg min (y − X𝛽𝛽 ) ′ (y − X𝛽𝛽 ) + 𝜆w′ |𝛽𝛽 |, (1.15)

which is a standard LASSO problem with 𝛽ˆ 𝑗 = 𝜃ˆ 𝑗 /𝑤 𝑗 .

1.2.4 Group LASSO

is to create 𝑀 dummy variables,4 each representing a single category. This leads

where the penalty function is the root of the quadratic form

4 Assuming there is no intercept.

This section provides a brief overview of the computation of various shrinkage

1.3.1 Computation and Least Angular Regression

It should be clear that each shrinkage estimator introduced above is a solution to

5 This has been extracted from Hastie et al. (2009)

The computation of LASSO requires a small modification to Step 4 above, namely:

1.3.2 Cross Validation and Tuning Parameters

Step 5. The tuning parameter can then be chosen based on

𝜆ˆ = arg min 𝐶𝑉 (𝜆).

1.4 Asymptotic Properties of Shrinkage Estimators

Valid statistical inference often relies on the asymptotic properties of estimators.

1.4.1 Oracle Properties

where Σ is the variance-covariance matrix of the following estimator

𝛽ˆ 𝑜𝑟 𝑎𝑐𝑙𝑒 = arg min 𝑔(𝛽𝛽 ). (1.20)