Forecasting by factors, by variables, by both, or neither?

Jennifer L. Castle† , Michael P. Clements and David F. Hendry∗

†
Magdalen College and Institute for New Economic Thinking at the
Oxford Martin School, University of Oxford, UK

Economics Department, Warwick University, UK
?
Economics Department and Institute for New Economic Thinking at the
Oxford Martin School, University of Oxford, UK

June 29, 2011

Abstract
We forecast US GDP and inflation over 1-, 4- and 8-step horizons using the dataset from Stock
and Watson (2009), with factors, variables, both, and neither. Autometrics handles perfect collinear-
ity and more regressors than observations, enabling all principal components and variables to be in-
cluded for model selection, jointly with using impulse-indicator saturation (IIS) for multiple breaks.
Empirically, factor models are more useful for 1-step ahead forecasts than at longer horizons, when
selecting over variables tends to be better. Accounting for in-sample breaks and outliers using IIS
is useful. Recursive updating helps, but recursive selection leads to greater variability, and neither
outperforms autoregressions.

JEL classifications: C51, C22.

Keywords: Model selection; Factor models; Forecasting; Impulse-indicator saturation; Autometrics

1 Introduction and historical background

There are three venerable traditions in economic forecasting based respectively on economic-theory de-
rived empirical econometric models, ‘indicator’ or ‘factor’ approaches combining many sources of in-
formation, and mechanistic approaches.1
Members of the first group are exemplified by early models like Smith (1927, 1929) and Tinbergen
(1930), smaller systems in the immediate post-war period (such as Klein, 1950, Tinbergen, 1951, Klein,
Ball, Hazlewood and Vandome, 1961), leading onto large macro-econometric models (Duesenberry,
Fromm, Klein and Kuh, 1969, and Fair, 1970, with a survey in Wallis, 1989), and now including both
dynamic stochastic general equilibrium (DSGE) models widely used at Central Banks (see e.g, Smets
and Wouters, 2003), and global models, first developed by project Link (see e.g., Waelbroeck, 1976) and
more recently, global vector autoregressions (GVARs: see Dees, di Mauro, Pesaran and Smith, 2007,
Pesaran, Schuerman and Smith, 2009, and Ericsson, 2010).
∗
This research was supported in part by grants from the Open Society Institute and the Oxford Martin School. Contact
details: jennifer.castle@magd.ox.ac.uk, m.p.Clements@warwick.ac.uk and david.hendry@nuffield.ox.ac.uk.
1
It is a great pleasure to contribute a paper on economic forecasting to a Festschrift in honor of Professor Hashem Pesaran,
who has made so many substantive contributions to this important topic. DFH has known Hashem for nearly 40 years, and
has benefited in countless ways from his vast range of publications on almost every conceivable topic in econometrics, theory
and applied. That output has received nearly 20,000 citations, achieved while first creating then editing the Journal of Applied
Econometrics since its foundation in 1986. We all look forward to many more salient insights from him.

1
 The second approach commenced with the ABC curves of Persons (1924), followed by leading indi-
cators as in Zarnowitz and Boschan (1977) with critiques in Diebold and Rudebusch (1991) and Emerson
and Hendry (1996). Factor analytic and principal component methods have a long history in statistics
and psychology (see e.g., Spearman, 1927, Cattell, 1952, Anderson, 1958, Lawley and Maxwell, 1963,
Joreskog, 1967, and Bartholomew, 1987) and have seen some distinguished applications in economics
(e.g., Stone, 1947, for an early macroeconomic application; and Gorman, 1956, for a microeconomic
one). Diffusion indices and factor models are now quite widely used for economic forecasting: see e.g.,
Stock and Watson (1989, 1999, 2009), Forni, Hallin, Lippi and Reichlin (2000), Peña and Poncela (2004,
2006), and Schumacher and Breitung (2008).
The third set includes methods like exponentially weighted moving averages, denoted EWMA, the
closely related Holt–Winters approach (see Holt, 1957, and Winters, 1960), damped trend (see e.g.,
Fildes, 1992), and autoregressions, including the general time-series approach in Box and Jenkins (1970).
Some members of this class were often found to dominate in forecasting competitions, such as Makri-
dakis, Andersen, Carbone, Fildes et al. (1982) and Makridakis and Hibon (2000).
Until recently, while the first two approaches often compared their forecasts with various ‘naive’
methods selected from the third group, there was little direct comparison between them, and almost no
studies included both. Here, we consider the reasons for that lacuna, and explain how it can be remedied.
The structure of the paper is as follows. Section 2 describes some of the issues that are likely to
bear on the topic of this paper, including the role of measurement errors. Section 3 relates the ‘external’
variables, denoted {zt }, to factors {ft }. Section 4 compares variable-based and factor-based models.
Section 5 develops the analysis of forecasting from factor models with a taxonomy of sources of forecast
error in the empirically relevant case of non-stationary processes. Section 6 addresses the problem of
systematic forecast failure to which equilibrium-correction formulations are prone in the face of location
shifts. Section 7 discusses model selection with both factors and variables, and section 8 illustrates the
analysis using US GDP and inflation forecasts. Section 9 concludes.

2 Setting the scene

There are a number of interacting issues that need to be addressed in an analysis of forecasting, whatever
device is used. The complexity of these issues, and the way they interact, means that an answer to the
question in the title of this paper ‘forecasting by factors, by variables, both, or neither?’ is likely to be
context specific. Even though general guidelines might prove hard to come by, it is fruitful to consider
these issues and how they affect our research question. We consider eight aspects: (i) the pooling of
both variables and factors in forecasting models; (ii) the role of in-sample model selection in that setting;
(iii) whether or not breaks over the forecast horizon are unanticipated; (iv) the role of more versus less
information in forecasting; (v) the type of forecasting model being used, specifically whether or not it is
an equilibrium-correction mechanism (EqCM); (vi) measurement errors in the data, especially near the
forecast origin; (vii) how to evaluate the ‘success or failure’ of forecasts; (viii) the nature of the DGP
itself. We briefly consider these in turn.

2.1 Pooling of information

Factor models are a way of forecasting using a large number of predictors, as opposed to pooling over
the forecasts of a large number of simple, often single-predictor, models. When there are many variables
in the set from which factors are formed (the ‘external’ variables), including both sets will often result
in the number of candidate variables, N , being larger than the sample size, T . This problem may have
seemed insurmountable in the past, but now is not. Let zt denote the set of n ‘external’ variables’
from which the factors ft = Bzt (say) are formed, then ft , . . . ft−s , zt , . . . zt−s comprise the initial set

2
of candidate variables. Automatic model selection can use multi-path searches to eliminate irrelevant
variables by exploring all feasible paths with mixtures of expanding and contracting block searches, so
can handle settings with both perfect collinearity and N > T , as shown in Hendry and Krolzig (2005)
and Doornik (2009b). The simulations in Castle, Doornik and Hendry (2011a) show the feasibility of
such an approach when N > T in linear dynamic models. Hence we are not forced at the outset to
allow only a small number of factors, or just the factors and a few lags of the variable being forecast,
say, as candidates. When the number of candidate variables exceeds the sample size, model selection is
unavoidable, so we consider that issue next.

2.2 Model selection

The ‘general-to-specific’ (Gets) search algorithm in Autometrics within PcGive (see Doornik, 2007,
2009a, and Hendry and Doornik, 2009) seeks the local data generating process (denoted LDGP), namely
the DGP for the set of variables under consideration (see e.g., Hendry, 2009) by formulating a general un-
restricted model (GUM) that nests the LDGP, and checking its congruence when feasible (estimable once
N  T and perfect collinearities are removed). Search thereafter ensures congruence, so all selected
models are valid restrictions of the GUM, and should parsimoniously encompass the feasible GUM.
Location shifts are removed in-sample by impulse-indicator saturation (IIS: see Hendry, Johansen and
Santos, 2008, Johansen and Nielsen, 2009, and the simulation  studies in Castle,
Doornik and Hendry,
2011c), which also addresses possible outliers. Thus, if 1{j=t} , t = 1, . . . , T denotes the complete
set of T impulse indicators, we allow for ft , . . . ft−s , zt , . . . zt−s and 1{j=t} , t = 1, . . . , T all being
included in the initial set of candidate variables to which multi-path search is applied, so N > T will al-
ways occur when IIS is used. The in-sample feasibility of this approach is established in Castle, Doornik
and Hendry (2011b). Here we are concerned with the application of models selected in this way to a
forecasting context when the DGP is non-stationary due to structural breaks. Since there seem to be few
analyses of how well a factor forecasting approach would then perform (see however, Stock and Watson,
2009, and Corradi and Swanson, 2011), we explore its behavior facing location shifts at the forecast
origin.

2.3 Unanticipated location shifts

Third, ex ante forecasting is fundamentally different from ex post modeling when unanticipated location
shifts can occur. Breaks can always be modeled after the event (at worst by indicator variables), but will
cause forecast failure when not anticipated. Clements and Hendry (1998, 1999) proposed a general theory
of economic forecasting for a world of structural breaks using mis-specified models, and emphasized that
it had radically different implications from a forecasting theory based on stationarity and well-specified
models (as in Klein, 1971, say). Moreover, those authors also establish that breaks other than location
shifts are less pernicious for forecasting (though not for policy analyses). Pesaran and Timmermann
(2005) and Pesaran, Pettenuzzo and Timmermann (2006) consider forecasting time series subject to
multiple structural breaks, and Pesaran and Timmermann (2007) examine the use of moving windows
in that context. Castle, Fawcett and Hendry (2010, 2011) investigate how breaks themselves might be
forecast, and if not, how to forecast during breaks, but draw somewhat pessimistic conclusions due to
the limited information that will be available at the time any location shift occurs. Thus, we focus the
analysis on the impacts of unanticipated location shifts in factor-based forecasting models.

2.4 Role of information in forecasting

Fourth, factor models can be interpreted as a particular form of ‘pooling of information’, in contrast to
the ‘pooling of forecasts’ literature discussed in (e.g.) Hendry and Clements (2004). Pooling information

3
ought to dominate pooling forecasts, each of which is based on limited information, except when all vari-
ables are orthogonal (see e.g, Granger, 1989). However, the taxonomy of forecast errors in Clements and
Hendry (2005b) suggests that incomplete information by itself is unlikely to play a key role in forecast
failure (except if that information would forecast breaks). Consequently, using large amounts of data
may not correct one of the main problems confronting forecasters, namely location shifts, unless that
additional information is directly pertinent to forecasting breaks. Moreover, although we use Gets model
selection from a very general initial candidate set, embedding available theory specifications, combined
with congruence as a basis for econometric modeling, when forecasting facing location shifts, it cannot
be that causal models will dominate non-causal (see Clements and Hendry, 1998) nor that congruent
modeling helps (see e.g., Allen and Fildes, 2001). Conversely, Makridakis and Hibon (2000) conclude
that parsimonious models do best in forecasting competitions, but Clements and Hendry (2001) argue
that such findings are conflated with robustness to location shifts because most of the parsimonious mod-
els evaluated also happened to be relatively robust to location shifts compared to their non-parsimonious
contenders.2 Since more information cannot lower predictability, and omitting crucial explanatory vari-
ables will both bias parameter estimates and lead to an inferior fit, the jury remains out on the benefits of
more versus less information when forecasting.

2.5 Equilibrium-correcting behavior

Fifth, factor models are often equilibrium correction in form, so they suffer from the general non-
robustness to location shifts of that class of model. However, the principles of robust-model formulation
discussed in Clements and Hendry (2005b) apply, and any equilibrium-correction system, whether based
on variables or factors (or both), could be differenced prior to forecasting, thereby embedding the result-
ing model in a second-differenced forecasting device. Castle et al. (2010) show that how a given model
is used in the forecast period matters, and explore various transformations that reduce systematic forecast
failure after location shifts. Section 6 provides a more extensive discussion.

2.6 Measurement errors

Sixth, many of the ‘solutions’ to systematic forecast failure induced by location shifts exacerbate the ad-
verse effects of data measurement errors near the forecast origin: for example, differencing doubles their
impact. Conversely, averaging mitigates the effects of random measurement errors, so as one method
of averaging over variables, factors might help mitigate data errors. Forecasting models which explic-
itly account for data revisions offer an alternative solution. These include modeling the different vin-
tage estimates of a given time-observation as a vector autoregression (see, e.g., Garratt, Lee, Mise and
Shields, 2008, 2009, and Hecq and Jacobs, 2009, following on from Patterson, 1995, 2003), as well as
the approach of Kishor and Koenig (2010) (building on earlier contributions by Howrey, 1978, 1984,
and Sargent, 1989). For the latter, a VAR is estimated on post-revision data, necessitating stopping the
estimation sample short of the forecast origin, and the model forecasts of the periods up to the origin
are combined with lightly-revised data for these periods via the Kalman filter to obtain post-revision
estimates. The forecast is then conditioned on these estimates of what the revised values of the latest
data will be. Clements and Galvão (2011) provide some evidence on the efficacy of these strategies for
forecasting US output growth and inflation, albeit using information sets consisting only of lags (and
different vintage estimates) of the variable being forecast.
2
Parsimonious models need not be robust–to see that the two characteristics are distinct, consider using as the forecast an
estimate of the unconditional mean of the process to date. No model specification/selection/estimation is required, apart from
the calculation of a sample mean, suggesting a simple forecasting device, which is nevertheless highly susceptible to location
shifts.

4
 The frequency of macroeconomic data can also affect its accuracy, as can nowcasting (see e.g.,
Castle, Fawcett and Hendry, 2009, and Bánbura, Giannone and Reichlin, 2011) and ‘real time’ (versus
ex post) forecasting (on the latter, see e.g., Croushore, 2006, and Clements and Galvão, 2008). Empirical
evidence suggests that the magnitudes of data measurement errors are larger in the most recent data, in
other words, in the data on which the forecast is being conditioned (hence the Kishor and Koenig, 2010,
idea of stopping the model estimation period early, and attempting to predict the ‘final’ estimates of the
most recent data), as well as during turbulent periods (Swanson and van Dijk, 2006), which might favour
factor models over other approaches that do not explicitly attempt to take data revisions into account.

2.7 Forecast evaluation

Next, there is a vast literature on how to evaluate the ‘success or failure’ of forecasts (see among many
others, Leitch and Tanner, 1991, Pesaran and Timmermann, 1992, Clements and Hendry, 1993a, Granger
and Pesaran, 2000a, 2000b, Pesaran and Skouras, 2002), as well as using forecasts to evaluate models
(see e.g., West, 1996, West and McCracken, 1998, Hansen and Timmermann, 2011, with a sceptical
view in Castle and Hendry, 2011b), forecasting methods (Giacomini and White, 2006), and economic
theory (Clements and Hendry, 2005a). As a first exercise in forecasting from models selected from both
variables and factors, below we just report descriptive statistics of forecast performance.

2.8 Nature of the DGP

Finally, the nature of the DGP itself matters greatly to the success of a specific forecasting model or
method. In particular, the factor model would be expected to do well if the ‘basic’ driving forces are
primarily factors, in the sense that a few factors account for a large part of the variance of the variables
of interest. The ideal case for factor model forecasting is where the DGP is:

xt = Υ (L) ft + et
ft = Φ (L) ft−1 + η t

where xt is n × 1, ft is m × 1, Υ(L) and Φ(L) are n × m and m × m, and n  m so that the low-
dimensional ft drives the co-movements of the high-dimensional xt . The latent factors are assumed here
to have a VAR representation. Suppose in addition that the mean-zero ‘idiosyncratic’ errors et satisfy
E[ei,t ej,t−k ] = 0 all k unless i = j (allowing the individual errors to be serially correlated), and that
E[η t et−k ] = 0 for all k.
It then follows that given the ft , each variable in xt , say xi,t , can be optimally forecast using only the
ft and lags of xi,t (xi,t−1 , xi,t−2 etc). If we let λi (L)0 denote the ith row of Υ(L), then:
 
Et [xi,t+1 | xt , ft , xt−1 , ft−1 , . . .] = Et λi (L)0 ft+1 + ei,t+1 | xt , ft , xt−1 , ft−1 , . . .
 
= Et λi (L)0 ft+1 | xt , ft , xt−1 , ft−1 , . . .
+ Et [ei,t+1 | xt , ft , xt−1 , ft−1 , . . .]
 
= Et λi (L)0 ft+1 | ft , ft−1 , . . . + Et [ei,t+1 | ei,t , ei,t−1 . . .]
= γ (L)0 ft + δ (L) xi,t

under the assumptions we have made (see Stock and Watson, 2011, for a detailed discussion). Absent
structural breaks, the model with the appropriate factors and lags of xi would deliver the best forecasts
(in population: ignoring parameter estimation uncertainty). The results of Faust and Wright (2007),
among others, suggest that the factor structure may not be a particularly good representation of the
macroeconomy. Our empirical approach allows that the ‘basic’ driving forces may be variables or factors,

5
and that there may be non-linearities (captured by linear approximations), as well as the many possible
non-stationarities noted above. We assume the DGP originates in the space of variables, with factors
being potentially convenient approximations that parsimoniously capture linear combinations of effects.
Although non-linearity can be tackled explicitly along with all the other complications (see e.g., Castle
and Hendry, 2011a), we only analyze linear DGPs here.
Thus, we consider forecasting from linear models selected in-sample from (a) a large set of variables,
(b) over those variables’ principal components (PCs), and (c) over a candidate set including both, in each
case with IIS, so the initial model will necessarily have N > T , and in the third case will be perfectly
collinear. We exploit the ability of automatic model selection to operate successfully in such a setting, as
well as to select despite more candidate variables than observations.

3 Relating ‘external’ variables to factors

Consider a vector of n stochastic variables {zt } that are weakly stationary over t = 1, . . . , T . For
specificity, we assume that zt is generated by a first-order vector autoregression (VAR) with deterministic
term π:
zt = π + Πzt−1 + vt (1)
where Π has all its eigenvalues inside the unit circle, and vt ∼ INn [0, Ωv ], where n < T . From (1):

E [zt ] = π + ΠE [zt−1 ] = π + Πµ = µ

where µ = (In − Π)−1 π. The principal-component description of zt is:

zt = Ψf t + et (2)

where ft ∼ IDm [κ, P] is a latent vector of dimension m ≤ n, so Ψ is n × m, with et ∼ IDn [0, Ωe ],

E[ft e0t ] = 0 and E [et e0t ] = Ωe . When E [ft ] = κ and E [et ] = 0, under weak stationarity from (2):

E [zt ] = ΨE [ft ] + E [et ] = Ψκ = µ (3)

Then:
     
E (zt − µ) (zt − µ)0 = ΨE (ft − κ) (ft − κ)0 Ψ0 + E et e0t
= ΨPΨ0 + Ωe = M (4)

say, where P is an m × m diagonal matrix and hence zt ∼ Dn [µ, M]. Let:

M = BΛB0 (5)

where B0 B = In , so B−1 = B0 and the eigenvalues are ordered from the largest downwards with:
 0   
0 B1 Λ11 0
B = and Λ = , (6)
B02 0 Λ22

where Λ11 is m × m, with B01 MB1 = Λ11 and:

BΛB0 = B1 Λ11 B01 + B2 Λ22 B02 .

Consequently, from (2) and (6):

B0 (zt − µ) = B0 (Ψ (ft − κ) + et ) = ft − κ (7)

6
If only m linear combinations actually matter, so n − m do not, the matrix B01 weights the zt to produce
the relevant principal components where:
B01 (zt − µ) = f1,t − κ1 (8)
In (7), we allow for the possibility that n = m, so ft is the complete set of principal components entered
in the candidate selection set, of which only f1,t are in fact relevant to explaining yt .

4 Variable-based and factor-based models

The postulated in-sample DGP for yt is:
yt = β 0 + β 0 zt−1 + ρyt−1 + t (9)
 
where |ρ| < 1 and t ∼ IN 0, σ 2
. Integrated-cointegrated systems can be reduced to this framework
analytically, albeit posing greater difficulties empirically. Under weak stationarity in-sample:
E [yt ] = β 0 + β 0 E [zt−1 ] + ρE [yt−1 ] = β 0 + β 0 µ + ρδ = δ (10)


so δ = β 0 + β 0 µ / (1 − ρ) and (9) can be expressed as:
yt − δ = β 0 (zt−1 − µ) + ρ (yt−1 − δ) + t (11)
or as an EqCM when that is a useful reparametrization. In general, only a subset of the zt−1 will matter
substantively, and we denote that by za,t−1 , so the remaining variables are not individually significant at
relevant sample sizes, leading to:
yt − δ ' β 0a (za,t−1 − µa ) + ρa (yt−1 − δ) + ut (12)
However, that does not preclude that known linear combinations of the omitted variables might also be
significant if added.
Alternatively, following up that last comment, from (7):
yt − δ = β 0 B (ft−1 − κ) + ρ (yt−1 − δ) + t = τ 0 (ft−1 − κ) + ρ (yt−1 − δ) + t (13)
where again only a subset may matter, namely the f1,t−1 in (8), so that:
yt − δ ' τ 01 (f1,t−1 − κ1 ) + ρ1 (yt−1 − δ) + wt (14)
Finally, there is the possibility that when both variables and their principal components are allowed,
some zb,t−1 of the za,t−1 and some f2,t−1 of the f1,t−1 are retained to provide closer, yet more parsimo-
nious, approximations to the behavior of yt in-sample:
yt − δ = β 0b (zb,t−1 − µb ) + τ 02 (f2,t−1 − κ2 ) + ρb (yt−1 − δ) + t (15)
In practice, there may well have been location shifts and outliers in-sample, so we also include IIS
during model selection. Thus, a vector of deterministic terms (such as intercepts, location shifts, and
indicator variables) denoted qt with Q1t = (q1 . . . qt ) is allowed, as well as longer lags, so the sequential
1 q ] (when that exists).
conditional expectation of yt at time t is denoted Et [yt |Z1t−1 , Yt−1 t
An important special case is when the DGP for yt is an autoregression, so that none of the zi,t−1
actually matter. When yt is just an AR(1), say, then:
yt = γ 0 + γ 1 yt−1 + vt .
Searching over factors alone, excluding yt−1 , might lead to the retention of many of the ft to approximate
yt−1 , especially if zt includes yt , so the starting model always includes yt−1 , allowing for simpler models
when they are the DGP. When searching over 2n variables and factors at significance level α, then 2αn
will be adventitiously significant on average (one retained by chance when n = 100 and α = 0.005, or
α = 0.0025 with IIS at T = 150, correpsonding to critical values of cα ' 2.85 and cα ' 3.10).

7
5 Forecasting from factors when variables matter
The aim is to forecast the scalar {yT +h } over a forecast horizon h = 1, . . . , H, from a forecast origin
at T , at which point the information set consists of Z1T = (z1 . .
. zT ) and (y1 . . . yT ). Forecast accuracy
is to be judged by a criterion function Ce u bT +1|T . . . u
bT +H|T , which we take to depend only on the
forecast errors ubT +h|T = yT +h − ybT +h|T , where ‘smaller’ values of Ce (·) are preferable. Even so,
unless the complete joint density is known, evaluation outcomes depend on the specific transformation
considered.
Once in-sample estimates of the factors {b ft } are available, one-step forecasts can be generated from
estimates of the selected equation:3
   
δ + τb 01 b
ybT +1|T = b f1,T − κ b1 + b ρ ybT − bδ (16)

where ybT is the ‘flash’ estimate of the forecast origin value. Multi-step estimation can be used to obtain
the values of the coefficients in the forecasting device (see e.g., Clements and Hendry, 1996, Bhansali,
2002, and Chevillon and Hendry, 2005, for overviews), so for h-step ahead forecasts:
   
ybT +h|T = b
δ (h) + τb 01,(h) b ρ(h) ybT − b
b1 + b
f1,T − κ δ (h) (17)

in which case u bT +h|T = yT +h − ybT +h|T will generally be a moving average process of order (h − 1)
(denoted MA(h − 1)).
Existing taxonomies of sources of forecast errors have analyzed a range of open and closed models in
variables, so here we consider the factor model when the DGP has a factor structure, as in (13). The DGP
depends on zt−1 and yt−1 , although not all the variables zi,t−1 need enter the DGP, and the forecasting
model is allowed to incorporate a subset of the factors. Our taxonomy of forecast errors focuses attention
on what are likely to be the principle sources of forecast bias and forecast-error variance. Following
earlier work, we begin by allowing location shifts as the only source of instability over the forecast
horizon, but then consider the impact of a shift in the parameter vector that determines the impact of the
factors on yt . Stock and Watson (2009) consider the effects of instabilities in the forecasting model–that
is, in the effects of the factors on yt –but as we show, a key determinant of forecasting performance is the
impact of location shifts. We let the DGP change at T to:

yT +h = δ ∗ + β 0 (zT +h−1 − µ∗ ) + ρ (yT +h−1 − δ ∗ ) + T +h (18)

for h = 1, . . . , H. Mapping to principal components yields:

yT +h = δ ∗ + τ 0 (fT +h−1 − κ∗ ) + ρ (yT +h−1 − δ ∗ ) + T +h (19)

where for now τ and ρ remain at their in-sample values during the forecast period.
We derive the 1-step forecast-error taxonomy, which highlights the key factors, and allows us to
separately distinguish 11 sources of forecast error. Calculating the forecast error as (19) minus (16), for
h = 1, gives rise to:
   
bT +1|T = δ∗ − b
u δ + τ 0 (fT − κ∗ ) − τb 01 bf1,T − κb1
 
ρ ybT − b
+ ρ (yT − δ ∗ ) − b δ + T +1 .

Using τ 01 (κ∗1 − κ1 ) + τ 02 (κ∗2 − κ2 ) = τ 0 (κ∗ − κ), we derive the forecast error reported in table 1.

3
Estimates b
f1,t of ft using principal components B01 (zt − µ) depend on the scaling of the zt , so are often based on the
correlation matrix.

8
 Table 1: Factor model taxonomy of forecast errors, u
bT +1|T = . . .

(1 − ρ) (δ∗ − δ) [A] equilibrium-mean shift

−τ 0 (κ∗ −κ)  [B] factor-mean shift
+ (1 − ρ) δ − b δ [C] equilibrium-mean estimation
−τ 01 (κ1−κ b 1) [D] factor-mean estimation
+ρ (yT − ybT )  [E] flash estimate error
+τ 0 f1,T − b
1 f1,T [F] factor estimate error
+τ 02 (f2,T− κ2 )  [G] factor approximation error
+ (τ 1 − τb 1 )0 b
f1,T − κ1 [H] factor estimation covariance
 
+ (ρ − bρ) ybT − b δ [I] flash estimation covariance
0
+ (τ 1 − τb 1 ) (κ1 − κ
b1) [J] parameter estimation covariance
+T +1 [K] innovation error

Taking expectations assuming near unbiased parameter estimates, and neglecting terms of Op T −1 :
 
E u yT ]) + τ 01 (f1,T − E[b
bT +1|T ' (1 − ρ) (δ ∗ − δ) − τ 0 (κ∗ − κ) + ρ (yT − E[b f1,T ]) (20)
which indicates that sources [A] and [B] in table 1 are primary determinants of forecast bias, although
data and factor estimation errors ([E] and [F]) also contribute. These last two and all the remaining
terms contribute to the forecast-error variance. The factor approximation error does not enter (20) as
E [f2,T ] = κ2 . Even when [E] and [F] are negligible, the equilibrium-mean and factor-mean shifts could
be large. For example, if in (1):
π ∗ = π + 1(t≥T ) θ for h = 1, . . . , H (21)
so that the intercept in the unmodeled variables representation undergoes a permanent shift at T , then as:
π = (In − Π) Ψκ
when Π and Ψ are constant, κ will shift, and for n = m:
κ∗ = Ψ−1 (In − Π)−1 π ∗ = κ + 1(t≥T ) Ψ−1 (In − Π)−1 θ (22)
Thus, forecast-error biases are entailed by equilibrium-mean shifts within the forecasting model of yT +1
(i.e., δ ∗ 6= δ) or in the external variables entering its DGP (κ∗ 6= κ) irrespective of the inclusion or
exclusion of the associated factors, whereas the approximation error by itself does not induce such a
problem. This outcome is little different from a model based directly on the zt (rather than ft ) where
shifts in their equilibrium mean can also induce forecast failure yet omission does not exacerbate that
problem (see Hendry and Mizon, 2011, for a general taxonomy of systems with unmodeled variables).
Consider now the possibility that τ and ρ change value for the forecast period, so that in place of
(19) the DGP is given by:
yT +1 = δ∗ + τ ∗0 (fT − κ∗ ) + ρ∗ (yT − δ ∗ ) + T +1 (23)
Without constructing a detailed taxonomy, the key impacts can be deduced. Relative to the baseline case
illustrated in table 1, the change in τ induces an additional error term:


τ ∗0 (fT − κ∗ ) − τ 0 (fT − κ∗ ) = τ ∗0 − τ 0 (fT − κ∗ )

9
so that the slope change will interact with the location shift, but in its absence will be relatively benign–
this additional term will not contribute to the bias when κ∗ = κ, suggesting the primacy of location
shifts. In a similar fashion, the change in persistence of the process (the shift in ρ) only affects the
forecast bias if the mean of yt also changes over the forecast period. To see this, the additional term in
the forecast error when ρ shifts is:
(ρ∗ − ρ) (yT − δ ∗ )
which has a zero expectation when the shift in ρ does not cause a shift in δ, so δ ∗ = δ.
Finally, it is illuminating to consider the principal sources of forecast error for an AR(1) model, as
this model serves as the benchmark against which the selected factor-and-variable models in section 8 are
to be compared. For the sake of brevity, we ignore factors of secondary importance, such as parameter
estimation uncertainty and data mis-measurement, and construct the forecast error for the AR(1):

yt = δ + γ (yt−1 − δ) + ut (24)

when the forecast period DGP is given by (19). Notice that the omission of the factors will typically
change the autoregressive parameter γ, so that γ need not equal ρ, but the long-run mean is the in-sample
period value of δ. Denoting the forecast error from the AR(1) model by vbT +1|T , we obtain:

vbT +1|T = (1 − ρ) (δ∗ − δ) − τ 0 (κ∗ − fT ) + (ρ − γ) (yT − δ)

with a forecast bias of:  

E vbT +1|T = (1 − ρ) (δ ∗ − δ) − τ 0 (κ∗ − κ) ,
matching the two leading terms in (20) for the bias of the factor-forecasting model. Hence whether we
include the ‘correct’ set of factors, a subset of these, or none at all will have no effect on the bias of the
forecasts (at the level of abstraction we are operating at here). This affirms the importance of location
shifts and the relative unimportance of forecasting model mis-specification (as in e.g., Clements and
Hendry, 2006).

6 The equilibrium-correction problem

Section 5 assumes a single forecast origin, but forecasting is rarely viewed as a one-off venture, and
of interest is the performance of the competing models as the origin moves through time. Although all
models will fail when there is a location shift which is unknown when the forecast is made, of interest
is the speed and extent to which forecasts recover as the origin moves forward in time from the break
point. A feature of the ‘equilibrium-correction’ class of models, to which (16) belongs, is their lack of
adaptability over time. To see this, note that (16) could be rewritten for 1-step forecasts as:
   
∆byT +1|T = τb 01 b b 1 + (b
f1,T − κ ρ − 1) ybT − b δ
 
so that E ∆b
yT +1|T ' 0, whereas the DGP is given by:

∆yT +1 = τ 0 (fT − κ∗ ) + (ρ − 1) (yT − δ ∗ ) + T +1 (25)

with an expected value which is non-zero when there are locations shifts:

E [∆yT +1 ] = τ 0 E [fT − κ∗ ] + (ρ − 1) E [yT − δ ∗ ] = τ 01 (κ1 − κ∗1 ) + (ρ − 1) (δ − δ ∗ ) (26)

Thus shifts in the deterministic terms will induce forecast failure, principally because they are embedded
in ∆yT +1 , but not in forecasts of this quantity. The class of equilibrium-correction models is such that

10
this problem persists as the origin is extended forward. For example, forecasting T + 2 from T + 1 even
for known in-sample parameters, accurate data and no approximation error, we find:

∆b yT +2|T +1 = τ 01 (κ1 − κ∗1 ) + (ρ − 1) (δ − δ ∗ ) + T +1

uT +2|T +1 = ∆yT +2 − ∆b

This generic difficulty for EqCMs suggests using a robust forecasting device approach which exploits
(26), as in:
∆eyT +2|T +1 = ∆yT +1 + τb 01 ∆b
f1,T + (b
ρ − 1) ∆b
yT
Again, under the simplifying assumptions (known in-sample parameters, etc), and denoting the forecast
error by ∆e
uT +2|T +1 = ∆yT +2 − ∆eyT +2|T +1 , using (25) gives:

uT +2|T +1 = τ 0 (fT +1 − κ∗ ) + (ρ − 1) (yT +1 − δ ∗ ) + T +2

∆e
− τ 0 (fT − κ∗ ) − (ρ − 1) (yT − δ ∗ ) − T +1
= τ 01 ∆f1,T +1 + (ρ − 1) ∆yT +1 + ∆T +2 (27)

which is less dependent on the location shifts.

To the extent that most factor models are also EqCMs, location shifts could have two impacts. The
first is when breaks affect the mapping between the original variables’ information and the derived factors
(i.e., changes in the weights). This is addressed in Stock and Watson (2009), who find a relatively
innocuous effect. Breaks in the coefficients of zero-mean variables or factors in forecasting models also
appear less problematic.
However, breaks due to location shifts within any EqCM forecasting model will induce systematic
mis-forecasting, and the above analysis applies equally to factor-based models (as illustrated in section
5). In the empirical forecasting exercise in section 8 below, the variables are already differenced once,
so large shifts in equilibrium means are unlikely, and hence such formulations already embody a partial
robustness to previous location shifts. Indeed, if in place of (24), the differenced-data version is used,
then forecasting T + 2 from T + 1:

∆e
yT +2|T +1 = γ∆yT +1

when:
∆yT +2 = τ ∗0 ∆fT +1 + ρ∗ ∆yT +1 + ∆T +2
we have:
veT +2|T +1 = τ ∗0 ∆fT +1 + (ρ∗ − γ) ∆yT +1 + ∆T +2
which is close to (27).

7 Automatic Model Selection

The primary comparison of interest is between automatic selection over variables as against PC-based
factor models in terms of forecasting. Factors are often regarded as necessary to summarize a large
amount of information, but automatic selection procedures show this is unnecessary. Selection will
place a zero weight on variables that are insignificant in explaining variation in the dependent variable
according to a pre-specified critical value, whereas principal components will place a small, but non-zero
weight on variables that have a low correlation with other explanatory variables.
One advantage of using an automatic model selection algorithm is that it enables us to remain ag-
nostic initially about the LDGP. If the data are generated by a few latent factors that capture underlying

11
movements in the economy such as business cycles, then principal components should be used to fore-
cast future outcomes. On the other hand, if the data are generated by individual disaggregated economic
variables then these should form the forecasting model. By including both explanations jointly the data
can determine the most plausible structure.
A further advantage of model selection is that arbitrary methods to select the relevant principal com-
ponents are not needed. Various methods have been proposed in the literature but most take the principal
components that explain the most variation between the set of explanatory variables, not the most vari-
ation between the explanatory variables and the dependent variable. This would require the correlation
structure between the regressors and the dependent variable to be similar to the correlation structure
within the regressors (see e.g., Castle et al., 2011b). Instead, by selecting PCs based on their statistical
significance, we capture the latter correlation. In the empirical application we find the retained PCs tend
not to be the first few PCs, suggesting that the correlation structure does differ between the dependent
variable and the disaggregates.
The model selection algorithm used is Autometrics, which undertakes a multi-path tree search, com-
mencing from the general model with all potential regressors including variables, factors and lags of both
as well as impulse indicators, and eliminates insignificant variables while ensuring a set of pre-specified
diagnostic tests are satisfied in the reduction procedure, checking the subsequent reductions with encom-
passing tests. Variables are eliminated if they are statistically insignificant at the chosen criterion whilst
ensuring the resulting model is still congruent and encompassing (see Doornik, 2008). There are various
methods to speed up the search procedure which involve joint testing.
The multi-path tree search enables perfectly collinear sets of regressors to be included jointly. While
the general model is not estimable initially, the search procedure proceeds by excluding one of the
perfectly-collinear variables initially so selection is undertaken within a subset of the candidate set, but
the multi-path search allows that excluded variable to be included in a different path search, with another
perfectly-singular variable being dropped. Autometrics uses expanding as well as contracting searches
which enables regressors initially excluded to return within different candidate sets. This ‘sieve’ contin-
ues until N < T and there are no perfect singularities. The standard tree search selection can then be
applied: see Doornik (2009a, 2009b).

8 Forecasting US GDP and Inflation

The empirical forecasting exercise compares the forecast performance of regression models based on
principal components, variables, or both. We forecast quarterly GDP growth and the quarterly change in
inflation over the period 1997–2006, as well as considering the corresponding level forecasts for GDP
and quarterly inflation. Models are selected in-sample using Autometrics, with all variables and their
principal components included in the candidate set.
The AR benchmark models against which factor model forecasts are often compared have typically
been difficult to beat systematically over this forecast period. For example, in terms of forecasting
inflation, Stock and Watson (2010) argue that simple univariate models, such as a random-walk model,
or the time-varying unobserved components model of Stock and Watson (2007), are competitive with
models with explanatory variables. Stock and Watson (2003) are relatively downbeat about the usefulness
of leading indicators for predicting output growth: also see Clements and Galvão (2009) for evidence
using higher-frequency data.

8.1 Data
The data are 144 seasonally adjusted quarterly time series for the United States, over 1959q1–2006q4
taken from Stock and Watson (2009). There are n = 109 disaggregates in the dataset which are used as

12
the candidate set of regressors and are also the set of variables used to form the principal components.
All data (usually in logs) are transformed to remove unit roots by taking first or second differences, as
described in Stock and Watson (2009) Appendix Table A1. The estimation sample spans T =1962q3–
2006q4, after transformations and lags, with the forecast horizon spanning H =1997q1–2006q4.

8.1.1 Principal Components

b the n × n sample
Let Z denote the (T + H) × n matrix of transformed disaggregated variables, and Ω
correlation matrix. The eigenvalue decomposition is:
b =B
Ω bΛbB
b0 (28)

where Λ b is the diagonal matrix of ordered eigenvalues (λ b1 ≥ . . . ≥ λ

bn ≥ 0) and B b1 , . . . , b
b = (b bn)
b 0 b
is the corresponding matrix of eigenvectors, with B B = In . The sample principal components are
computed as:
b b 0e
ft = B zt (29)
where Z e = (e zT )0 denotes the standardized data, zej,t = (zj,t − z j ) /e
z1 , . . . , e σ xj ∀j = 1, . . . , n where
P h P i1/2
T T 2
z j = T1 t=1 zj,t and σ ezj = T1 t=1 (zj,t − z j ) . When the principal components are estimated
in-sample, h = 0, whereas h = 1, . . . , H for recursive estimation of the principal components.

8.2 Impulse-indicator saturation

A number of authors have assessed the forecast performance of factor models over this period, and Stock
and Watson (2011) review studies which explicitly consider the impact of breaks on factor model fore-
casts. One of the key studies is Stock and Watson (2009) who find ‘considerable evidence of instability
in the factor model; the indirect evidence suggests instability in all elements (the factor loadings, the
factor dynamics, and the idiosyncratic dynamics).’ (Stock and Watson, 2009, p. 197). They identify a
break in 1984, associated with the Great Moderation (see, e.g., McConnell and Perez-Quiros, 2000), and
find the coefficients of the factors in the forecasting models are not constant across this period.
However, they argue that the factors can be reasonably well estimated by PCs even when the individ-
ual loadings are subject to instability, and more accurate factor model forecasts are found to result from
estimating the factors on the whole sample, but the forecasting models only on the period after 1984. As
such, we estimate the principal components over the full sample period, but rather than restricting the
estimation sample to post-1984, we use IIS to account for parameter stability. Instead of imposing a sin-
gle break at this point, we test for the presence of multiple breaks and outliers jointly with the selection
procedure by applying IIS. Although this procedure adds an impulse indicator for every observation to
the candidate regressor set, there is little efficiency loss under the null of no breaks, yet the procedure
has power to detect both outliers and location shifts when there are breaks: see Castle et al. (2011c).

8.3 Forecasting models

The forecasting models are obtained by undertaking selection on the general unrestricted model (GUM):
Jb
X X Jb
n X X Jb
n X T
X
∆yt = γ 0 + γ j ∆yt−j + β i,j ∆zi,t−j + φk,j fk,t−j + δl 1{l=t} + t (30)
j=Ja i=1 j=Ja k=1 j=Ja l=1

where ∆yt is the first difference of log real gross domestic product or the quarterly change in quarterly
inflation.

13
 Forecasting models are obtained by undertaking selection on (30) using Autometrics where we set:
(i) φ = 0, i.e. select over variables only;
(ii) β = 0, select over factors only; and
(iii) φ 6= 0 and β 6= 0, i.e. jointly select variables and factors;
where the intercept and lags of the dependent variable are included in all models. For the three forecasting
specifications we consider:
(a) δ = 0, no IIS; and
(b) δ 6= 0, with IIS,
resulting in six forecasting model specifications.
Three forecast horizons are recorded, including 1-step, 4-step and 8-step ahead direct forecasts. For
the 1-step ahead forecasts we set Ja = 1 and Jb = 4, allowing for 4 lags of the dependent and exogenous
regressors. For 4-step ahead direct forecasts we set Ja = 4 and Jb = 7, and 8-step ahead forecasts set
Ja = 8 and Jb = 11.
One of the problems that factor forecasts face is the need to difference to stationarity in order to com-
pute the principal components. This implies that any structural breaks in the levels will be differenced
out. Many variables are second differenced to obtain stationarity, so breaks in growth rates will also be
removed. A consequence of differencing is that the resulting forecasts will be robust to breaks after they
occur. However, accurate forecasting of the differences is insufficient for levels, as an insignificant one-
direction error can cumulate to an important under or over forecast: e.g., 1% per quarter cumulates to
4% over a year. Hence, we also report the implied log-level forecasts. As the forecasts are direct h-step
forecasts rather than dynamic forecasts, we let the forecast origin, T , roll forward rather than remain
fixed at 1996q4. The levels forecasts for log GDP and quarterly inflation are computed as:
h
X
ybT +k+h = ∆b
yT +k+h + yT +k for k = 0, . . . , H − h (31)
i=1

where h = 4 and 8 for the 4-step and 8-step ahead forecasts respectively in (31) (1-step forecast errors
are the same for levels and differences). For the 4-step forecasts, we evaluate over 37 forecasts as we
require forecasts for the period 1997q1–1997q4 to obtain the level forecast of 1997q4, and likewise for
the 8-step ahead forecasts we evaluate over 33 forecasts.
There are T = 138 observations in-sample. For the variables only or factor forecasts, there are
N = 441 regressors in (30) with no IIS and N = 579 with IIS. We set the significance level for selection
at α = 1% with no IIS, so approximately 4 regressors would be retained on average under the null that
no regressors are relevant, and set α = 0.5% for IIS. When both variables and factors are included, there
are N = 877 regressors with no IIS, so we set α = 0.5%, and with IIS N = 1015, so α = 0.25%. Table
2 summarizes the selection significance levels, with null retention numbers in parentheses. Overfitting
is not a concern despite commencing with N  T . Parsimony can be achieved by a tighter significance
level, with the cost being a loss of power for regressors with a significance level less than the critical
value. As a check, we consider a super-conservative strategy by setting the significance levels even
tighter, as shown in the last row of table 2.

Variables Factors Both

No IIS IIS No IIS IIS No IIS IIS
Number of regressors 441 579 441 579 877 1015
Conservative (null) 1% (4.4) 0.5% (3) 1% (4.4) 0.5% (3) 0.5% (4.4) 0.25% (2.5)
Tighter (null) 0.5% (2.2) 0.1% (0.6) 0.5% (2.2) 0.1 (0.6)% 0.1% (0.9) 0.05% (0.5)

Table 2: Significance levels used for model selection.

14
 We also compute three benchmark forecasts including the random walk and AR(1) forecasts com-
puted directly and iteratively:

yTRW
∆b +k+h = ∆yT +k (32)
AR(D) b +βb ∆yT +k
∆b
yT +k+h =β 0 1 (33)
h−1
X
AR(I)
∆b
yT +k+h = γ bi1 + γ
b0 γ bh1 ∆yT +k (34)
i=1

for k = 0, . . . , H − h and h = 1, 4 and 8.

8.4 Results
We first consider in-sample selection and estimation, where the forecasting model is selected and esti-
mated over t = 1, . . . , T , with 40 forecasts computed over k = 1, . . . , H, resulting in 90 forecasting
models. We evaluate the forecasts on RMSFE, with the caveat that the MSFE criterion may not result
in a definitive ranking, as that measure is not invariant to non-singular, scale preserving linear transfor-
mations, see Clements and Hendry (1993a, 1993b).

8.4.1 In-sample estimation and selection for GDP

First we consider the forecast performance for GDP of factors, variables or both when the forecasting
model is selected and estimated in-sample. Table 3 records the in-sample model fit and number of
retained regressors for selection with IIS. In-sample, the model fit is generally better for the variables
than factors. Tightening the significance level results in fewer retained regressors and a worse model fit
as expected. Few dummies are retained on average, but the retained dummies are clustered around 1984,
supporting the identification of a break there by Stock and Watson (2009).

1-step 4-step 8-step

Cons Super Cons Super Cons Super
Variables
σ
b 0.559 0.669 0.629 0.711 0.702 0.799
No. regressors 9 5 14 7 9 5
No. dummies 2 1 1 2 5 4
Factors
σ
b 0.657 0.702 0.718 0.798 0.671 0.813
No. regressors 5 4 8 4 12 5
No. dummies 2 1 6 2 5 2
Both
σ
b 0.553 0.753 0.712 0.819 0.767 0.788
No. regressors 9 2 9 4 9 6
No. dummies 2 0 1 0 2 4

Table 3: In-sample results for GDP growth models selected with IIS: σ b = equation standard error, No.
regressors and No. dummies record the number of regressors and, as a subset, the number of dummies
retained, and Cons and Super are the conservative and super-conservative strategies respectively.

Table 4 records the average RMSFE for each of the forecasting models, averaged across horizon,
whether IIS is applied, and the selection significance level for GDP and GDP growth. The RMSFEs are

15
closely similar for the first three methods, but forecasting with factors is slightly preferable to forecasting
with variables for GDP growth, although it performs worse for the levels forecasts. However, the AR(1)
model, either iterative or direct, has the lowest average RMSFE.

Variables Factors Both RW AR(D) AR(I)

∆b yT +k 0.666 0.600 0.650 0.666 0.485 0.491
ybT +k 1.931 1.971 1.925 3.509 1.324 1.336

Table 4: RMSFE (×100) for GDP and quarterly GDP growth, with benchmark Random Walk, direct
AR(1) [AR(D)] and iterative AR(1) [AR(I)] forecasts.

We deconstruct the results for the first 3 methods in figures 1 and 2 by analyzing forecast performance
over the forecast horizon in panel (a), whether IIS was applied or not in panel (b), and the tightness of
the selection criterion in panel (c). We plot the results for the 1-step ahead levels forecasts in figure 2 for
comparison, despite their being identical to those in figure 1.
Variables Factors Both

0.0075 0.0075

0.0050 0.0050
0.0050

0.0025 0.0025 0.0025

1−step 4−step 8−step IIS no IIS conservative super−

conservative
Figure 1: Average RMSFE for GDP growth (∆b
yT +h )
For GDP growth, the factor model performs best at short horizons but worst at longer horizons.
The variable model improves in forecast accuracy at the 8-step horizon. There is some benefit to IIS
but few indicators are retained on average given the tight significance level used for selection (no more
than six), and the indicators retained are roughly the same for the factors and variables models. A
tighter significance level does improve the variable model forecasts, suggesting parsimonious models are
preferable here, but there is little improvement to the forecasts in the factor model.
Variables Factors Both
0.04
0.02 0.02

0.02 0.01 0.01

1−step 4−step 8−step IIS no IIS conservative super−

conservative
Figure 2: Average RMSFE for log GDP (b
yT +h )
In levels, worsening factor forecasts as the horizon increases is evident. IIS yields greater improve-
ments as the few dummies that are retained will be translated to level shifts capturing the shift in 1984.

16
 Figures 3 and 4 record the distributions of forecast errors for variables (panel a), factors (panel b) and
both (panel c) for GDP growth and the level of GDP respectively.4 In growth rates, the forecast errors are
close to normal for variables and factors models. The levels forecasts have a fatter upper tail and some
evidence of bimodality, but there are no significant differences between the variables and factors model
forecast errors.

Variables Factors Both

75
50 50

50

25 25
25

−0.025 0.000 0.025 −0.02 0.00 0.02 −0.025 0.000 0.025

Figure 3: Distributions of forecast errors for GDP growth averaging across horizon, IIS/no IIS and
selection strategy.

40
Variables Factors 30 Both
30
30

20 20
20

10 10
10

−0.05 0.00 0.05 0.10 −0.05 0.00 0.05 −0.05 0.00 0.05

Figure 4: Distributions of forecast errors for log GDP averaging across horizon, IIS/no IIS and selection
strategy.

8.4.2 In-sample estimation and selection for inflation

We next consider the results for inflation. Table 5 records the in-sample model fit and number of retained
variables for selection with IIS. In-sample, selecting over variables results in a better model fit than
factors or both at all horizons, in keeping with the GDP results. Direct models at longer horizons do
not always result in a worse fit, and neither does a tighter significance level, despite fewer regressors
retained for the super-conservative strategy. Few dummies are retained on average suggesting that the
differencing to stationarity has removed most breaks, although more dummies are retained at longer
horizons as direct models exclude factors and variables at lags shorter than the forecast horizon.
Table 6 reports the forecast results for inflation, averaged across horizon, selection significance level
and whether IIS is applied. In differences, including both variables and factors outperforms the individual
4
For GDP growth there are 480 forecast errors for each model specification. For the levels distributions, we include the
1-step levels forecasts (which are identical to the difference forecast errors), resulting in 440 forecast errors.

17
models, but not the AR(1) models, either direct or iterative, as in GDP growth. Variables perform better
than factors, particularly in levels, due to worsening forecast performance of factors at longer horizons.

1-step 4-step 8-step

Cons Super Cons Super Cons Super
Variables
σ
b 0.240 0.283 0.293 0.290 0.227 0.296
No. regressors 20 12 17 13 21 12
No. dummies 2 1 4 8 8 3
Factors
σ
b 0.263 0.336 0.299 0.318 0.311 0.327
No. regressors 15 7 20 14 21 20
No. dummies 5 1 4 13 7 6
Both
σ
b 0.266 0.284 0.302 0.356 0.284 0.334
No. regressors 14 11 12 10 13 8
No. dummies 4 3 4 3 1 3

Table 5: In-sample results for inflation forecasting models selected with IIS: σ
b = equation standard
error, No. regressors and No. dummies record the number of regressors and, as a subset, the number of
dummies retained, and Cons and Super are conservative and super-conservative strategies respectively.

Variables Factors Both RW AR(D) AR(I)

∆b yT +k 0.535 0.582 0.491 0.577 0.416 0.415
ybT +k 0.687 0.981 0.656 0.469 0.444 0.464

Table 6: RMSFE for quarterly inflation and the quarterly change in inflation, with benchmark Random
Walk, direct AR(1) [AR(D)] and iterative AR(1) [AR(I)] forecasts.

To examine the results more closely, figures 5 and 6 record results for the forecast horizon in panel
(a), whether IIS was applied or not in panel (b), and the selection significance level in panel (c).
Variables Factors Both
0.75

0.50 0.50
0.50

0.25 0.25
0.25

1−step 4−step 8−step IIS no IIS conservative super−

conservative
Figure 5: Average RMSFE for annual change in inflation (∆b
yT +k )
In differences, variables models and both factors and variables models are roughly constant over the
forecast horizon but the factor models perform worst on average as the horizon increases. There are
some gains to IIS for variable models or factor models, but the gains are small. Tightening the selection
criterion improves the forecasts, with the more parsimonious models for the variable models and factor

18
 Variables Factors Both

1.0 1.0
1.5

1.0
0.5 0.5

0.5

1−step 4−step 8−step IIS no IIS conservative super−

conservative
Figure 6: Average RMSFE for quarterly inflation (b
yT +k )

models yielding a smaller RMSFE. Factors models selected using a super-conservative strategy are best
on a RMSFE criterion at a 1-step horizon, both with and without IIS, but also the worst over the longer
8-step horizon. Thus, factor models appear more useful for the shorter forecasting horizons here.
Transforming to levels highlights the worsening forecast accuracy of the factor model over the fore-
cast horizon. The transformation can result in differing rankings, as can be seen at the 8-step horizon
where the variable model is preferred in levels but the combined model is preferred in differences, albeit
by small magnitudes. Also, a looser strategy is preferred for the factor models as opposed to the other
models. The differences between the combined and variable models are small because most of the re-
tained regressors for the combined model are variables. However, including many additional factors in
selection is not costly when undertaken at a conservative or super-conservative strategy.
Figures 7 and 8 record the distributions of forecast errors for the variables, factors, and combined
models. The forecast errors are approximately normally distributed for the variables models. In levels,
the factors model has a fatter lower tail, but again there is little systematic difference in forecast errors
for the different models.

1.0
Variables Factors Both

0.75
0.75

0.50 0.50 0.5

0.25 0.25

−2 0 2 −2.5 0.0 2.5 0 2

Figure 7: Distributions of forecast errors for the quarterly change in inflation averaging across horizon,
IIS/no IIS and selection strategy.

8.4.3 Selection and estimation method

Finally, we assess whether the selection and estimation method used impacts on forecast accuracy. We
compare four methods:
(1) the forecasting model is fixed after selection and estimation in-sample over t = 1, . . . , T ;
(2) the model is selected over t = 1, . . . , T , but both parameters and eigenvalues for the principal com-
ponents are recursively re-estimated over each forecast horizon, T = 1, . . . , T + h;

19
 Variables Factors 0.75 Both
0.75
0.50

0.50 0.50

0.25
0.25 0.25

−2.5 0.0 2.5 −5.0 −2.5 0.0 2.5 −2 0 2

Figure 8: Distributions of forecast errors for quarterly inflation averaging across horizon, IIS/no IIS and
selection strategy.

(3) the models is both selected and re-estimated recursively over each forecast horizon, t = 1, . . . , T + h,
and the eigenvalues re-calculated, so the model specification can change with each new observation; and
(4) partial recursive selection and estimation, where the previous selected model is forced, in that selec-
tion occurs only over the additional variables or factors, so previously retained variables are not dropped
even if insignificant.
Figure 9 records the RMSFEs for GDP growth with IIS using the conservative strategy for each
estimation method. There are no clear advantages to using any specific selection and estimation method.
Recursive estimation marginally outperforms in-sample estimation but there are cases where in-sample
estimation is preferred, for example the factor model at 4-step and the variable model at 8-step. Recursive
selection and estimation can be worse at longer horizons, with partial recursive selection providing a
slightly more robust selection method. In general, the differencing and IIS has accounted for the in-
sample breaks and so there is little difference between selection and estimation methods for GDP growth.
Factors Variables Both

1−step 4−step 8−step

0.0075 0.0075 0.0075

0.0050 0.0050 0.0050

0.0025 0.0025 0.0025

n on e
n on e

n on e

io ti iv
io ti iv

io ti iv

e
io e

io e
io e
e

pl

rs ial
ct a rs
ct a rs

ct a rs
at rsiv

at rsiv
pl

at rsiv

pl

e
rs ial
rs ial

n
n

am

iv
le timcu
le timcu

le timcu

e
am

cu rt
m

iv
iv

cu rt

re pa
cu rt

se es re
se es re

se es re
timcu

timcu
timcu

sa

−s
−s

re pa
re pa

es re

es re
es re

−
in

in
in

Figure 9: Average RMSFE for GDP growth with IIS using the conservative strategy.
Figure 10 records the RMSFEs for the quarterly change in inflation with IIS using the conservative
strategy for each estimation method. At 1-step and 4-step there is very little difference in RMSFEs
between the selection and estimation strategies, but at 8-step there is weak evidence to suggest recursive
estimation or partial recursive estimation and selection is preferred, particularly for variables.

9 Conclusion
There have been many analyses of the forecast performance of either factor models or regression models,
but few examples of the joint consideration of factors and variables. Automatic model selection can allow

20
 Factors Variables Both

1−step 4−step 8−step

0.50 0.50 0.50

0.25 0.25 0.25

n on e

n on e
io ti iv
n on e

io ti iv
io e

io e
io e

io ti iv

at rsiv

at rsiv
ct a rs
at rsiv
e

ct a rs
pl
ct a rs

n
pl

rs ial
le timcu

rs ial
rs ial

le timcu
pl

e
e
le timcu

am
m

iv
tim c u

tim c u
se es re

iv
tim c u

iv

cu rt
se es re
m
cu rt
cu rt
se es re
a

re pa
−s

re pa
es re
re pa

es re
es re
−s

−s
in
in

in
Figure 10: Average RMSFE for the quarterly change in inflation with IIS using the conservative strategy.

for more regressors than observations, perfect collinearities and multiple breaks and outliers, so the set
of candidate regressors can include both factors, as measured by their static principal components, and
variables. We consider which methods perform best in a forecasting context, including naive devices.
One of the key explanations for forecast failure is that of structural breaks. When the underlying
data generating process shifts, but the forecasting model remains unchanged, forecast failure will result.
As both variables and factor models are in the class of equilibrium-correction models, they both face
the problem of non-robustness to location shifts. In our empirical example, we use impulse-indicator
saturation to account for breaks and outliers in-sample, noting that IIS could be used to implement
intercept corrections if an indicator variable were retained for the last in-sample observation. We find
there is some advantage to using IIS for forecasting in differences as the unconditional mean is better
estimated, but as the data are differenced to estimate the principal components, few impulse-indicators
are retained. Backing out levels forecasts does highlight the non-stationarity due to level shifts (e.g., the
structural break in 1984), and a further extension would be to consider modeling and selecting variables
in levels, augmented by the stationary principal components which may pick up underlying latent variable
dynamics.
The empirical application considered both GDP and inflation, and their differences, using Automet-
rics to select forecasting models that include either principal components (PCs), individual variables, or
both. The results are mixed but suggest that factor models are more useful for 1-step ahead forecasting,
but their relative performance declines as the forecast horizon increases. For direct multi-step forecasting,
Autometrics selection over variables (or variables and PCs) tends to forecast better than factor forecasts,
suggesting that there are benefits to selecting the weights based on the correlation with yt+h . There is
little evidence to suggest that recursive estimation or recursive estimation and selection is better. A more
robust alternative of partial recursive estimation and selection is proposed. There are gains to using IIS,
but a tight significance is needed to control the retention rate.
The ability to undertake model selection jointly on factors and variables avoids imposing a model
specified in either just variables or factors, and circumvents the need for arbitrary selection of factors.
Thus it is a useful tool, both for in-sample modeling and for forecasting out-of-sample. Whether the
data are best described by latent factors or observable variables will depend on the phenomena being
analysed, and can be determined by the data itself using model selection techniques.

References
Allen, P. G., and Fildes, R. A. (2001). Econometric forecasting strategies and techniques. In Armstrong,
J. S. (ed.), Principles of Forecasting, pp. 303–362. Boston: Kluwer Academic Publishers.

21
Anderson, T. W. (1958). An Introduction to Multivariate Statistical Analysis. New York: John Wiley &
Sons.
Bánbura, M., Giannone, D., and Reichlin, L. (2011). Nowcasting. in Clements, and Hendry (2011),
Ch. 7.
Bartholomew, D. J. (1987). Latent Variable Models and Factor Analysis. New York: Oxford University
Press.
Bhansali, R. J. (2002). Multi-step forecasting. in Clements, and Hendry (2002), pp. 206–221.
Box, G. E. P., and Jenkins, G. M. (1970). Time Series Analysis: Forecasting and Control. San Francisco:
Holden-Day.
Castle, J. L., Doornik, J. A., and Hendry, D. F. (2011a). Evaluating automatic model selection. Journal
of Time Series Econometrics, 3 (1), DOI: 10.2202/1941–1928.1097.
Castle, J. L., Doornik, J. A., and Hendry, D. F. (2011b). Model selection in equations with many ‘small’
effects. Discussion paper 528, Economics Department, Oxford University.
Castle, J. L., Doornik, J. A., and Hendry, D. F. (2011c). Model selection when there are multiple breaks.
Journal of Econometrics, forthcoming.
Castle, J. L., Fawcett, N. W. P., and Hendry, D. F. (2009). Nowcasting is not just contemporaneous
forecasting. National Institute Economic Review, 210, 71–89.
Castle, J. L., Fawcett, N. W. P., and Hendry, D. F. (2010). Forecasting with equilibrium-correction models
during structural breaks. Journal of Econometrics, 158, 25–36.
Castle, J. L., Fawcett, N. W. P., and Hendry, D. F. (2011). Forecasting Breaks and During Breaks. In
Clements, and Hendry (2011), Ch. 11.
Castle, J. L., and Hendry, D. F. (2011a). Automatic selection of non-linear models. In Wang, L., Gar-
nier, H., and Jackman, T.(eds.), System Identification, Environmental Modelling and Control, p.
forthcoming. New York: Springer.
Castle, J. L., and Hendry, D. F. (2011b). On not evaluating economic models by forecast outcomes.
Discussion paper, 538, Economics Department, Oxford University.
Castle, J. L., and Shephard, N.(eds.)(2009). The Methodology and Practice of Econometrics. Oxford:
Oxford University Press.
Cattell, R. B. (1952). Factor Analysis. New York: Harper.
Chevillon, G., and Hendry, D. F. (2005). Non-parametric direct multi-step estimation for forecasting
economic processes. International Journal of Forecasting, 21, 201–218.
Clements, M. P., and Galvão, A. B. (2008). Macroeconomic forecasting with mixed-frequency data:
Forecasting output growth in the United States. Journal of Business and Economic Statistics, 26,
546–554.
Clements, M. P., and Galvão, A. B. (2009). Forecasting US output growth using leading indicators: An
appraisal using MIDAS models. Journal of Applied Econometrics, 24, 1187–1206.
Clements, M. P., and Galvão, A. B. (2011). Improving real-time estimates of output gaps and inflation
trends with multiple-vintage VAR models. Discussion paper, Department of Economics, Univer-
sity of Warwick.
Clements, M. P., and Hendry, D. F. (1993a). On the limitations of comparing mean squared forecast
errors (with discussion). Journal of Forecasting, 12, 617–637.
Clements, M. P., and Hendry, D. F. (1993b). On the limitations of comparing mean squared forecast
errors: A reply. Journal of Forecasting, 12, 669–676.
Clements, M. P., and Hendry, D. F. (1996). Multi-step estimation for forecasting. Oxford Bulletin of

22
 Economics and Statistics, 58, 657–684.
Clements, M. P., and Hendry, D. F. (1998). Forecasting Economic Time Series. Cambridge: Cambridge
University Press.
Clements, M. P., and Hendry, D. F. (1999). Forecasting Non-stationary Economic Time Series. Cam-
bridge, Mass.: MIT Press.
Clements, M. P., and Hendry, D. F. (2001). Explaining the results of the M3 forecasting competition.
International Journal of Forecasting, 17, 550–554.
Clements, M. P., and Hendry, D. F.(eds.)(2002). A Companion to Economic Forecasting. Oxford: Black-
wells.
Clements, M. P., and Hendry, D. F. (2005a). Evaluating a model by forecast performance. Oxford Bulletin
of Economics and Statistics, 67, 931–956.
Clements, M. P., and Hendry, D. F. (2005b). Guest editors’ introduction: Information in economic
forecasting. Oxford Bulletin of Economics and Statistics, 67, 713–753.
Clements, M. P., and Hendry, D. F. (2006). Forecasting with breaks. In Elliott, G., Granger, C., and
Timmermann, A.(eds.), Handbook of Economic Forecasting, Volume 1. Handbook of Economics
24, pp. 605–657: Elsevier, Horth-Holland.
Clements, M. P., and Hendry, D. F.(eds.)(2011). Oxford Handbook of Economic Forecasting. Oxford:
Oxford University Press.
Corradi, V., and Swanson, N. R. (2011). Testing for factor model forecast and structural stability. Work-
ing paper, Economics Department, Warwick University.
Croushore, D. (2006). Forecasting with real-time macroeconomic data. In Elliott, G., Granger, C., and
Timmermann, A.(eds.), op.cit., pp. 961–982.
Dees, S., di Mauro, F., Pesaran, M. H., and Smith, L. V. (2007). Exploring the international linkages in
the EURO area: A global VAR analysis. Journal of Applied Econometrics, 22, 1–38.
Diebold, F. X., and Rudebusch, G. D. (1991). Forecasting output with the composite leading index: An
ex ante analysis. Journal of the American Statistical Association, 86, 603–610.
Doornik, J. A. (2007). Object-Oriented Matrix Programming using Ox 6th ed. London: Timberlake
Consultants Press.
Doornik, J. A. (2008). Encompassing and automatic model selection. Oxford Bulletin of Economics and
Statistics, 70, 915–925.
Doornik, J. A. (2009a). Autometrics. In Castle, and Shephard (2009), pp. 88–121.
Doornik, J. A. (2009b). Econometric model selection with more variables than observations. Working
paper, Economics Department, University of Oxford.
Duesenberry, J. S., Fromm, G., Klein, L. R., and Kuh, E.(eds.)(1969). The Brookings Model: Some
Further Results. Amsterdam: North-Holland.
Emerson, R. A., and Hendry, D. F. (1996). An evaluation of forecasting using leading indicators. Journal
of Forecasting, 15, 271–291.
Ericsson, N. R. (2010). Improving GVARs. Discussion paper, Federal Reserve Board of Governors,
Washington, D.C.
Fair, R. C. (1970). A Short-Run Forecasting Model of the United States Economy. Lexington: D.C.
Heath.
Faust, J., and Wright, J. H. (2007). Comparing Greenbook and reduced form forecasts using a large
realtime dataset. NBER Working Paper Series 13397.
Fildes, R. (1992). The evaluation of extrapolative forecasting methods. International Journal of Fore-

23
 casting, 8, 81–98.
Forni, M., Hallin, M., Lippi, M., and Reichlin, L. (2000). The generalized factor model: Identification
and estimation. Review of Economics and Statistics, 82, 540–554.
Garratt, A., Lee, K., Mise, E., and Shields, K. (2008). Real time representations of the output gap.
Review of Economics and Statistics, 90, 792–804.
Garratt, A., Lee, K., Mise, E., and Shields, K. (2009). Real time representations of the UK output gap in
the presence of model uncertainty. International Journal of Forecasting, 25, 81–102.
Giacomini, R., and White, H. (2006). Tests of conditional predictive ability. Econometrica, 74, 1545 –
1578.
Gorman, W. M. (1956). Demand for related goods. Discussion paper, Agricultural Experimental Station,
Iowa.
Granger, C. W. J. (1989). Combining forecasts - Twenty years later. Journal of Forecasting, 8, 167–173.
Granger, C. W. J., and Pesaran, M. H. (2000a). A decision-theoretic approach to forecast evaluation. In
Chon, W. S., Li, W. K., and Tong, H.(eds.), Statistics and Finance: An Interface, pp. 261–278.
London: Imperial College Press.
Granger, C. W. J., and Pesaran, M. H. (2000b). Economic and statistical measures of forecasting accu-
racy. Journal of Forecasting, 19, 537–560.
Hansen, P. R., and Timmermann, A. (2011). Choice of sample split in forecast evaluation. Working
paper, Economics Department, Stanford University.
Hecq, A., and Jacobs, J. P. A. M. (2009). On the VAR-VECM representation of real time data. Discussion
paper, mimeo, University of Maastricht, Department of Quantitative Economics.
Hendry, D. F. (2009). The methodology of empirical econometric modeling: Applied econometrics
through the looking-glass. In Mills, T. C., and Patterson, K. D.(eds.), Palgrave Handbook of
Econometrics, pp. 3–67. Basingstoke: Palgrave MacMillan.
Hendry, D. F., and Clements, M. P. (2004). Pooling of forecasts. Econometrics Journal, 7, 1–31.
Hendry, D. F., and Doornik, J. A. (2009). Empirical Econometric Modelling using PcGive: Volume I.
London: Timberlake Consultants Press.
Hendry, D. F., Johansen, S., and Santos, C. (2008). Automatic selection of indicators in a fully saturated
regression. Computational Statistics, 33, 317–335. Erratum, 337–339.
Hendry, D. F., and Krolzig, H.-M. (2005). The properties of automatic Gets modelling. Economic
Journal, 115, C32–C61.
Hendry, D. F., and Mizon, G. E. (2011). An open-model forecast-error taxonomy. Working paper 552,
Economics Department, Oxford University.
Holt, C. C. (1957). Forecasting seasonals and trends by exponentially weighted moving averages. ONR
Research Memorandum 52, Carnegie Institute of Technology, Pittsburgh, Pennsylvania.
Howrey, E. P. (1978). The use of preliminary data in econometric forecasting. Review of Economics and
Statistics, 60, 193–200.
Howrey, E. P. (1984). Data revisions, reconstruction and prediction: an application to inventory invest-
ment. The Review of Economics and Statistics, 66, 386–393.
Johansen, S., and Nielsen, B. (2009). An analysis of the indicator saturation estimator as a robust regres-
sion estimator. In Castle, and Shephard (2009), pp. 1–36.
Joreskog, K. G. (1967). Some contributions to maximum likelihood factor analysis. Psychometrika, 32.
Kishor, N. K., and Koenig, E. F. (2010). VAR estimation and forecasting when data are subject to
revision. Journal of Business and Economic Statistics. Forthcoming.

24
Klein, L. R. (1950). Economic Fluctuations in the United States, 1921–41. No. 11 in Cowles Commission
Monograph. New York: John Wiley.
Klein, L. R. (1971). An Essay on the Theory of Economic Prediction. Chicago: Markham Publishing
Company.
Klein, L. R., Ball, R. J., Hazlewood, A., and Vandome, P. (1961). An Econometric Model of the UK.
Oxford: Oxford University Press.
Lawley, D. N., and Maxwell, A. E. (1963). Factor Analysis as a Statistical Method. London: Butterworth
and Co.
Leitch, G., and Tanner, J. E. (1991). Economic forecast evaluation: Profits versus the conventional error
measures. American Economic Review, 81, 580–590.
Makridakis, S., Andersen, A., Carbone, R., Fildes, R., et al. (1982). The accuracy of extrapolation (time
series) methods: Results of a forecasting competition. Journal of Forecasting, 1, 111–153.
Makridakis, S., and Hibon, M. (2000). The M3-competition: Results, conclusions and implications.
International Journal of Forecasting, 16, 451–476.
McConnell, M. M., and Perez-Quiros, G. P. (2000). Output fluctuations in the United States: What has
changed since the early 1980s?. American Economic Review, 90, 1464–1476.
Patterson, K. D. (1995). An integrated model of the data measurement and data generation processes
with an application to consumers’ expenditure. Economic Journal, 105, 54–76.
Patterson, K. D. (2003). Exploiting information in vintages of time-series data. International Journal of
Forecasting, 19, 177–197.
Peña, D., and Poncela, P. (2004). Forecasting with nonstationary dynamic factor models. Journal of
Econometrics, 119, 291–321.
Peña, D., and Poncela, P. (2006). Nonstationary dynamic factor analysis. Journal of Statistical Planning
and Inference, 136, 1237–1257.
Persons, W. M. (1924). The Problem of Business Forecasting. No. 6 in Pollak Foundation for Economic
Research Publications. London: Pitman.
Pesaran, M. H., Pettenuzzo, D., and Timmermann, A. (2006). Forecasting time series subject to multiple
structural breaks. Review of Economic Studies, 73, 1057–1084.
Pesaran, M. H., Schuerman, T., and Smith, L. V. (2009). Forecasting economic and financial variables
with global VARs. International Journal of Forecasting, 25, 642–675.
Pesaran, M. H., and Skouras, S. (2002). Decision-based methods for forecast evaluation. in Clements,
and Hendry (2002), pp. 241–267.
Pesaran, M. H., and Timmermann, A. (1992). A simple nonparametric test of predictive performance.
Journal of Business and Economic Statistics, 10, 461–465.
Pesaran, M. H., and Timmermann, A. (2005). Small sample properties of forecasts from autoregressive
models under structural breaks. Journal of Econometrics, 129, 183–217.
Pesaran, M. H., and Timmermann, A. (2007). Selection of estimation window in the presence of breaks.
Journal of Econometrics, 137, 134–161.
Sargent, T. J. (1989). Two models of measurements and the investment accelerator. Journal of Political
Economy, 97, 251–287.
Schumacher, C., and Breitung, J. (2008). Real-time forecasting of German GDP based on a large factor
model with monthly and quarterly data. International Journal of Forecasting, 24, 386–398.
Smets, F., and Wouters, R. (2003). An estimated stochastic dynamic general equilibrium model of the
Euro Area. Journal of the European Economic Association, 1, 1123–1175.

25
Smith, B. B. (1927). Forecasting the volume and value of the cotton crop. Journal of the American
Statistical Association, 22, 442–459.
Smith, B. B. (1929). Judging the forecast for 1929. Journal of the American Statistical Association, 24,
94–98.
Spearman, C. (1927). The Abilities of Man. London: Macmillan.
Stock, J. H., and Watson, M. W. (1989). New indexes of coincident and leading economic indicators.
NBER Macro-Economic Annual, 351–409.
Stock, J. H., and Watson, M. W. (1999). A comparison of linear and nonlinear models for forecasting
macroeconomic time series. In Engle, R. F., and White, H.(eds.), Cointegration, Causality and
Forecasting, pp. 1–44. Oxford: Oxford University Press.
Stock, J. H., and Watson, M. W. (2003). How Did Leading Indicator Forecasts Perform During the 2001
Recession. Federal Reserve Bank of Richmond, Economic Quarterly, 89/3, 71–90.
Stock, J. H., and Watson, M. W. (2007). Why has U.S. Inflation Become Harder to Forecast?. Journal
of Money, Credit and Banking, Supplement to Vol. 39, 3–33.
Stock, J. H., and Watson, M. W. (2009). Forecasting in dynamic factor models subject to structural
instability. In Castle, and Shephard (2009), Ch. 7.
Stock, J. H., and Watson, M. W. (2010). Modelling Inflation after the Crisis. NBER Working Paper
Series, 16488.
Stock, J. H., and Watson, M. W. (2011). Dynamic factor models. In Clements, and Hendry (2011), Ch. 2.
Stone, J. R. N. (1947). On the interdependence of blocks of transactions. Journal of the Royal Statistical
Society, 8, 1–32. Supplement.
Swanson, N. R., and van Dijk, D. (2006). Are statistical reporting agencies getting it right? data ratio-
nality and business cycle asymmetry. Journal of Business and Economic Statistics, 24, 24–42.
Tinbergen, J. (1930). Determination and interpretation of supply curves: An example [Bestimmung
und Deutung von Angebotskurven: ein Beispiel]. Zeitschrift fur Nationalökonomie, 1, 669–679.
Reprinted in Hendry, D. F. and Morgan, M. S. (1995), The Foundations of Econometric Analysis.
Cambridge: Cambridge University Press.
Tinbergen, J. (1951). Business Cycles in the United Kingdom 1870–1914. Amsterdam: North-Holland.
Waelbroeck, J. K. (ed.)(1976). The Models of Project LINK. Amsterdam: North-Holland Publishing
Company.
Wallis, K. F. (1989). Macroeconomic forecasting: A survey. Economic Journal, 99, 28–61.
West, K. D. (1996). Asymptotic inference about predictive ability. Econometrica, 64, 1067–1084.
West, K. D., and McCracken, M. W. (1998). Regression-based tests of predictive ability. International
Economic Review, 39, 817–840.
Winters, P. R. (1960). Forecasting sales by exponentially weighted moving averages. Management
Science, 6, 324–342.
Zarnowitz, V., and Boschan, C. (1977). Cyclical indicators: An evaluation and new leading indexes. In
Handbook of Cyclical Indicators, pp. 170–183. Washington, USA: U.S. Department of Commerce.

26