Artificial intelligence and machine learning in financial

services

Market developments and financial stability implications

1 November 2017
The Financial Stability Board (FSB) is established to coordinate at the international level the
work of national financial authorities and international standard-setting bodies in order to
develop and promote the implementation of effective regulatory, supervisory and other
financial sector policies. Its mandate is set out in the FSB Charter, which governs the
policymaking and related activities of the FSB. These activities, including any decisions
reached in their context, shall not be binding or give rise to any legal rights or obligations
under the FSB’s Articles of Association.

Contacting the Financial Stability Board

Sign up for e-mail alerts:
www.fsb.org/emailalert Follow the FSB on
Twitter: @FinStbBoard
E-mail the FSB at: fsb@fsb.org

Copyright © 2017 Financial Stability Board. Please refer to: http://www.fsb.org/terms_conditions

Contents Page
Executive Summary....................................................................................................................1
Introduction.................................................................................................................................3
1. Background and definitions..........................................................................................3
2. Drivers...........................................................................................................................7
3. Selected use cases.......................................................................................................10
3.1 Customer-focused uses: credit scoring, insurance and client-facing chatbots............11
3.1.1 Credit scoring applications..........................................................................................12
3.1.2 Use for pricing, marketing and managing insurance policies.....................................13
3.1.3 Client-facing chatbots.................................................................................................14
3.2 Operations-focused uses.............................................................................................15
3.2.1 Capital optimisation use case......................................................................................15
3.2.2 Model risk management (back-testing and model validation) and stress testing.......16
3.2.3 Market impact analysis (modelling of trading out of big positions)...........................17
3.3 Trading and portfolio management.............................................................................18
3.3.1 AI and machine learning in trading execution............................................................18
3.3.2 Scope for the use of AI and machine learning in portfolio management....................18
3.4 AI and machine learning in regulatory compliance and supervision..........................19
3.4.1 RegTech: applications by financial institutions for regulatory compliance...............20
3.4.2 Uses for macroprudential surveillance and data quality assurance.............................21
3.4.3 SupTech: uses and potential uses by central banks and prudential authorities...........21
3.4.4 Uses by market regulators for surveillance and fraud detection.................................23
4. Micro-financial analysis..............................................................................................24
4.1 Possible effects of AI and machine learning on financial markets.............................24
4.2 Possible effects of AI and machine learning on financial institutions........................25
4.3 Possible effects of AI and machine learning on consumers and investors.................27
4.4 Current regulatory considerations regarding the use of AI and machine learning.....28
5. Macro-financial analysis.............................................................................................29
5.1 Market concentration and systemic importance of institutions...................................29
5.2 Potential Market Vulnerabilities.................................................................................30
5.3 Networks and interconnectedness...............................................................................31
5.4 Other implications of AI and machine learning applications......................................31
6. Conclusions and implications for financial stability...................................................32
iii
Glossary.....................................................................................................................................35
Annex A: Legal issues around AI and machine learning..........................................................37
Annex B: AI ethics....................................................................................................................39
Annex C: Drafting team members............................................................................................40

iv
Executive Summary
Artificial intelligence (AI) and machine learning are being rapidly adopted for a range of
applications in the financial services industry. As such, it is important to begin considering the
financial stability implications of such uses. Because uses of this technology in finance are in
a nascent and rapidly evolving phase, and data on usage are largely unavailable, any analysis
must be necessarily preliminary, and developments in this area should be monitored closely.
Many applications, or “use cases”, of AI and machine learning already exist. The adoption of
these use cases has been driven by both supply factors, such as technological advances and
the availability of financial sector data and infrastructure, and by demand factors, such as
profitability needs, competition with other firms, and the demands of financial regulation.
Some of the current and potential use cases of AI and machine learning include:
- Financial institutions and vendors are using AI and machine learning methods to
assess credit quality, to price and market insurance contracts, and to automate client
interaction.
- Institutions are optimising scarce capital with AI and machine learning techniques, as
well as back-testing models and analysing the market impact of trading large
positions.
- Hedge funds, broker-dealers, and other firms are using AI and machine learning to
find signals for higher (and uncorrelated) returns and optimise trading execution.
- Both public and private sector institutions may use these technologies for regulatory
compliance, surveillance, data quality assessment, and fraud detection.
With the FSB FinTech framework,1 our analysis reveals a number of potential benefits and
risks for financial stability that should be monitored as the technology is adopted in the
coming years and as more data becomes available. In some cases, these observations are also
contained in the FSB report on regulatory and supervisory issues around FinTech.2 They are:
- The more efficient processing of information, for example in credit decisions,
financial markets, insurance contracts, and customer interaction, may contribute to a
more efficient financial system. The RegTech and SupTech applications of AI and
machine learning can help improve regulatory compliance and increase supervisory
effectiveness.
- At the same time, network effects and scalability of new technologies may in the
future give rise to third-party dependencies. This could in turn lead to the emergence
of new systemically important players that could fall outside the regulatory perimeter.
- Applications of AI and machine learning could result in new and unexpected forms of
interconnectedness between financial markets and institutions, for instance based on
the use by various institutions of previously unrelated data sources.

1
FSB (2016), “Fintech: Describing the Landscape and a Framework for Analysis,” March [unpublished].
2
FSB (2017), “Financial Stability Implications from FinTech, Supervisory and Regulatory Issues that Merit Authorities
Attention,” June.

1
 - The lack of interpretability or “auditability” of AI and machine learning methods
could become a macro-level risk. Similarly, a widespread use of opaque models may
result in unintended consequences.
- As with any new product or service, there are important issues around appropriate risk
management and oversight. It will be important to assess uses of AI and machine
learning in view of their risks, including adherence to relevant protocols on data
privacy, conduct risks, and cybersecurity. Adequate testing and ‘training’ of tools with
unbiased data and feedback mechanisms is important to ensure applications do what
they are intended to do.
Overall, AI and machine learning applications show substantial promise if their specific risks
are properly managed. The concluding section gives preliminary thoughts on governance and
development of models, as well as auditability by institutions and supervisors.

2
Introduction
This report analyses possible financial stability implications of the use of artificial intelligence
(AI) and machine learning in financial services. It was drafted by a team of experts from the
FSB Financial Innovation Network (FIN). The report draws on discussions with firms; 3
academic research; public and private sector reports; and ongoing work at FSB member
institutions.4 The report analyses potential financial stability implications of the growing use
of AI by financial institutions. Given the relative novelty of many applications, and the
paucity of data on adoption, it is necessarily a horizon-scanning piece.
The report is structured as follows. In section 1, the key concepts of the report are defined,
and some background is given on the development of AI and machine learning for financial
applications. Section 2 describes supply and demand factors driving the adoption of these
techniques in financial services. Section 3 describes four sets of use cases: (i) customer-
focused applications; (ii) operations-focused uses; (iii) trading and portfolio management; and
(iv) regulatory compliance and supervision. Section 4 contains a micro-analysis of the effects
of adoption on financial markets, institutions and consumers. Section 5 gives a macro-analysis
of effects on the financial system. Finally, section 6 concludes with an assessment of
implications for financial stability.

1. Background and definitions

Researchers in computer science and statistics have developed advanced techniques to obtain
insights from large disparate data sets. Data may be of different types, from different sources,
and of different quality (structured and unstructured data). These techniques can leverage the
ability of computers to perform tasks, such as recognising images and processing natural
languages, by learning from experience. The application of computational tools to address
tasks traditionally requiring human sophistication is broadly termed ‘artificial intelligence’
(AI). As a field, AI has existed for many years. However, recent increases in computing
power coupled with increases in the availability and quantity of data have resulted in a
resurgence of interest in potential applications of artificial intelligence.5 These applications
are already being used to diagnose diseases, translate languages, and drive cars; and they are
increasingly being used in the financial sector as well.

3
FIN held two workshops on this topic. The first workshop was held on 4 April 2017 in San Francisco. The second
workshop was held on 27 June 2017 in Basel. The participants at these workshops included representatives from 7
financial institutions, six artificial intelligence firms, three large tech firms and two industry organisations from North
America, Europe and Asia. In addition, drafting team members and the FSB secretariat conducted bilateral conversations
with relevant private sector contacts across a range of jurisdictions.
4
The report draws on some examples from specific private firms involved in FinTech. These examples are not exhaustive
and do not constitute an endorsement by the FSB for any firm, product or service. Similarly, they do not imply any
conclusion about the status of any product or service described under applicable law. Rather, such examples are included
for purposes of illustration of new and emerging business models in the markets studied.
5
Various financial regulatory authorities have defined the big data phenomenon as a confluence of factors, including the
ubiquitous collection of data from a variety of sources, the plummeting cost of data storage and powerful capacity to
analyse data. See, e.g., U.S. Federal Trade Commission Report, January (2016,), “Big Data: A tool for Inclusion or
Exclusion?” January, p. 1; EBA, EIOPA and ESMA (2016), “European Joint Committee Discussion Paper on the Use of
Big Data by Financial Institutions,” JC 2016 86, p. 7.

3
There are many terms that are used in describing this field, so some definitions are needed
before proceeding. ‘Big data’ is a term for which there is no single, consistent definition, but
the term is used broadly to describe the storage and analysis of large and/or complicated data
sets using a variety of techniques including AI.6 Analysis of such large and complicated
datasets is often called ‘big data analytics.’ A key feature of the complexity relevant in big
data sets analytics often relates to the amount of unstructured or semi-structured data
contained in the datasets.
This report defines AI as the theory and development of computer systems able to perform
tasks that traditionally have required human intelligence. AI is a broad field, of which
‘machine learning’ is a sub-category.7 Machine learning may be defined as a method of
designing a sequence of actions to solve a problem, known as algorithms, 8 which optimise
automatically through experience and with limited or no human intervention. 9 These
techniques can be used to find patterns in large amounts of data (big data analytics) from
increasingly diverse and innovative sources. Figure 1 gives an overview.

Figure 1: A schematic view of AI, machine learning and big data analytics

Artificial intelligence

Supervised learning
Big data analytics
Machine learning
Reinforcement learning
Deep

Unsupervised learning

Many machine learning tools build on statistical methods that are familiar to most researchers.
These include extending linear regression models to deal with potentially millions of inputs,
or using statistical techniques to summarise a large dataset for easy visualisation. Yet
machine

6
Jonathan Stuart Ward and Adam Barker (2013), “Undefined By Data: A Survey of Big Data Definitions” Cornell
University, arXiv:1309.5821.
7
Examples of AI applications that are not machine learning include the computer science fields of ontology management,
or the formal naming and defining of terms and relationships by computers, as well as inductive and deductive logic and
knowledge representation. In this report, for completeness, we often refer to “AI and machine learning,” with the
understanding that many of the important recent advances are in the machine learning space.
8
An algorithm may be defined as a set of steps to be performed or rules to be followed to solve a mathematical problem.
More recently, the term has been adopted to refer to a process to be followed, often by a computer.
9
Arthur Samuel (1959), “Some Studies in Machine Learning Using the Game of Checkers,” IBM Journal: 211-229; Tom
Mitchell (1997), Machine Learning, New York: McGraw Hill; Michael Jordan and Tom Mitchell (2015), “Machine
learning: Trends, perspectives, and prospects,” Science 349(6245): 255-260. Samuel defined machine learning as the
“field of study that gives computers the ability to learn without being explicitly programmed,” while Mitchell defines it as
the “the question of how to build computers that improve automatically through experience.”

4
learning frameworks are inherently more flexible; patterns detected by machine learning
algorithms are not constrained to the linear relationships that tend to dominate economic and
financial analysis. In general, machine learning deals with (automated) optimisation,
prediction, and categorisation, not with causal inference. 10 In other words, classifying whether
the debt of a company will be investment grade or high yield one year from now could be
done with machine learning. However, determining what factors have driven the level of bond
yields would likely not be done using machine learning.
There are several categories of machine learning algorithms. These categories vary according
to the level of human intervention required in labelling the data:
• In ‘supervised learning’, the algorithm is fed a set of ‘training’ data that contains
labels on some portion of the observations. For instance, a data set of transactions may
contain labels on some data points identifying those that are fraudulent and those that
are not fraudulent. The algorithm will ‘learn’ a general rule of classification that it will
use to predict the labels for the remaining observations in the data set.
• ‘Unsupervised learning’ refers to situations where the data provided to the algorithm
does not contain labels. The algorithm is asked to detect patterns in the data by
identifying clusters of observations that depend on similar underlying characteristics.
For example, an unsupervised machine learning algorithm could be set up to look for
securities that have characteristics similar to an illiquid security that is hard to price. If
it finds an appropriate cluster for the illiquid security, pricing of other securities in the
cluster can be used to help price the illiquid security.
• ‘Reinforcement learning’ falls in between supervised and unsupervised learning. In
this case, the algorithm is fed an unlabelled set of data, chooses an action for each data
point, and receives feedback (perhaps from a human) that helps the algorithm learn.
For instance, reinforcement learning can be used in robotics, game theory, and self-
driving cars.11
• ‘Deep learning’ is a form of machine learning that uses algorithms that work in
‘layers’ inspired by the structure and function of the brain. Deep learning algorithms,
whose structure are called artificial neural networks, can be used for supervised,
unsupervised, or reinforcement learning.
Recently, deep learning has led to remarkable results in diverse fields, such as image
recognition and natural language processing (NLP). Deep learning algorithms are capable of
discovering generalisable concepts, such as encoding the concept of a ‘car’ from a series of
images. An investor might deploy an algorithm that recognises cars to count the number of
cars in a retail parking lot from a satellite image in order to infer a likely store sales figure for
a particular period. NLP allows computers to ‘read’ and produce written text or, when
combined

10
Here, prediction is understood as identifying something as likely before the event based on experience. Causal inference
and forecasting are done from a scientific perspective on the basis of analysis of the past.
11
This categorisation of machine learning algorithms is taken from Kolanovic, Marko and Rajesh Krishnamachari (2017),
“Big Data and AI Strategies: Machine Learning and Alternative Data Approach to Investing,” JP Morgan, May; and

5
Internet Society (2017), “Artificial Intelligence and Machine Learning: Policy Paper,” Internet Society White Paper,
April.

6
with voice recognition, to read and produce spoken language. This allows firms to automate
financial service functions previously requiring manual intervention.
Machine learning can be applied to different types of problems, such as classification or
regression analysis. Classification algorithms, which are far more frequently deployed in
practice, group observations into a finite number of categories. Classification algorithms are
probability-based, meaning that the outcome is the category for which it finds the highest
probability that it belongs to. An example might be to automatically read a sell-side report and
label it as ‘bullish’ or ‘bearish’ with some probability, or estimate an unrated company’s
initial credit rating. Regression algorithms, in contrast, estimate the outcome of problems that
have an infinite number of solutions (continuous set of possible outcomes). This outcome can
be accompanied with a confidence interval. Regression algorithms can be used for the pricing
of options. Regression algorithms can also be used as one intermediate step of classification
algorithm.
It is important to note what machine learning cannot do, such as determining causality.
Generally speaking, machine learning algorithms are used to identify patterns that are
correlated with other events or patterns. The patterns that machine learning identifies are
merely correlations, some of which are unrecognisable to the human eye. However, AI and
machine learning applications are being used increasingly by economists and others to help
understand complex relationships, along with other tools and domain expertise.
Many machine learning techniques are hardly new. Indeed, neural networks, the base concept
for deep learning, were first developed in the 1960s. 12 However after an initial burst of
excitement, machine learning and AI failed to live up to their promises and funding dissipated
for over a decade, in part because of the lack of sufficient computing power and data. There
was renewed funding and interest in applications in the 1980’s, during which many of the
research concepts were developed for later breakthroughs.13
By 2011 and 2012, driven by the vast increase in the computational power of modern
computers, machine learning algorithms, especially deep learning algorithms, began to
consistently win image, text, and speech recognition contests. Noticing this trend, major tech
companies began to acquire deep learning start-ups and rapidly accelerate deep learning
research.14 Also new is the scale of collection of big data, for example the ability to capture
data on the scale of every single credit card transaction or every word on the web, and even
‘mouse’ hovers over websites. Other advances have also helped, such as increased
interconnectedness of information technology resources with cloud computing architecture,
with which big data can now be organised and analysed. Using data sets of this size and
complexity and with the increase in computing power, machine learning algorithms results
have improved, some of which are highlighted in the sections that follow. This has also
spurred large

12
The ‘K-Nearest-Neighbour’ algorithm, which is the simple, intuitive machine learning algorithm taught at the start of
many undergraduate Computer Science classes, was proposed in its modern form in 1967. See Thomas Cover and Peter
Hart (1967), “Nearest Neighbor Pattern Classification,” IEEE transactions on information theory. 13.1: 21-27.
13
See Luke Dormehl (2016), Thinking Machines: The Quest for Artificial Intelligence--and Where It's Taking Us Next,
London: Penguin Books.
14
Tim Dettmers (2015), “Deep Learning in a Nutshell: History and Training,” Parallel Forall, December. Accessed on May
30, 2017. Web. https://devblogs.nvidia.com/parallelforall/deep-learning-nutshell-history-training/

7
investments in AI start-ups. The World Economic Forum reported that global investment in
AI start-ups rose from $282 million in 2011, to $2.4 billion in 2015. 15 The number of merger
and acquisition (M&A) deals in AI has also accelerated over this period (figure 2).

Figure 2: Global artificial intelligence merger and acquisition activity, 2012-2017

Source: CB Insights (2017), “The Race For AI: Google, Baidu, Intel, Apple In A Rush To Grab Artificial Intelligence
Startups,” Research Brief, July.

Many applications tend more toward ‘augmented intelligence,’ or an augmentation of human

capabilities, rather than a replacement of humans. Even as advancements in AI and machine
learning continue, including in the area of deep learning, most industries are not attempting to
fully replicate human intelligence. As noted by one industry observer “…a human in the loop
is essential: we are, unlike machines, able to take into account context and use general
knowledge to put AI-drawn conclusions into perspective.”16

2. Drivers
A variety of factors that have contributed to the growing use of FinTech generally have also
spurred adoption of AI and machine learning in financial services. 17 On the supply side,
financial market participants have benefitted from the availability of AI and machine learning

15
See Slaughter and May and ASI Data Science (2017), “Superhuman Resources: Responsible deployment of AI in
business,” Joint White Paper, June.
16
See Finextra and Intel (2017), “The Next Big Wave: How Financial Institutions Can Stay Ahead of the AI Revolution,”
May.
17
See FSB FinTech Issues Group (2017), p.10. AI and machine learning are also being adopted widely in sectors such as
health care, manufacturing, marketing, and many other areas.

8
tools developed for applications in other fields. These include availability of computing power
owing to faster processor speeds, lower hardware costs, and better access to computing power
via cloud services.18 Similarly, there is cheaper storage, parsing, and analysis of data through
the availability of targeted databases, software, and algorithms. There has also been a rapid
growth of datasets for learning and prediction owing to increased digitisation and the adoption
of web-based services.19 The declining cost of data storage and estimates of the volume of
global data sets are shown in figure 3.

Figure 3: Costs of storage and global data availability, 2009-2017

Source: Reinsel, Gantz and Rydning (2017); Klein (2017). One zettabyte is equal to one billion terabytes.

The same tools driving advances in machine learning in search engines and self-driving cars,
can be adopted in the financial sector. For example, entity recognition tools that enable search
engines to understand when a user is referring to Ford Motor Company, rather than fording a
river, are now used to quickly identify news or social media chatter relevant to publicly traded
firms. As more firms adopt these tools, the financial incentives to access new or additional
data and to develop faster and more accurate AI and machine learning tools may increase. In
turn, such adoption and development of tools may affect incentives for yet other firms.
A variety of technological developments in the financial sector have contributed to the
creation of infrastructure and data sets. The proliferation of electronic trading platforms has
been accompanied by an increase in the availability of high quality market data in structured
formats.20 In some countries, such as the United States, market regulators allow publicly
traded firms to use social media for public announcements. In addition to making digitised
financial

18
See IOSCO (2017), “Research Report on Financial Technologies (Fintech),” February, p. 6 on growth of computing power.
19
Institute of International Finance (2017), “Deploying RegTech Against Financial Crime,” Report of the IIF RegTech
Working Group, March; David Reinsel, John Gantz and John Rydning (2017), “Data Age 2025: The Evolution of Data to
Life-Critical,” IDC White Paper, April; Andy Klein (2017), “Hard Drive Cost Per Gigabyte,” Backblaze, July.
20
See IOSCO (2017), p. 41 for description of APIs and FIX protocols to assist in price discovery and market liquidity in
corporate bond markets.

9
 data available for machine learning, the computerisation of markets has made it possible for
AI algorithms to interact directly with markets, putting in real-time complex buy and sell
orders based on sophisticated decision-making, in many cases with minimal human
intervention. Meanwhile, retail credit scoring systems have become more common since the
1980s,21 and news has become machine readable since the 1990s. With the growth of data in
financial markets as well as datasets – such as online search trends, viewership patterns and
social media that contain financial information about markets and consumers – there are even
more data sources that can be explored and mined in the financial sector.
On the demand side, financial institutions have incentives to use AI and machine learning for
business needs. Opportunities for cost reduction, risk management gains, and productivity
improvements have encouraged adoption, as they all can contribute to greater profitability. In
a recent study, industry sources described priorities for using AI and machine learning as
follows: optimising processes on behalf of clients; working to create interactions between
systems and staff applying AI to enhance decision-making; and developing new products and
services to offer to clients.22 In many cases these factors may also drive ‘arms races’ in which
market participants increasingly find it necessary to keep up with their competitors’ adoption
of AI and machine learning, including for reputational reasons (hype).

Figure 4: Supply and demand factors of financial adoption of AI and machine learning

Technology Financial sector factors Profitability Competition Regulation

vements in computing Availability
power, data of
availability,
infrastructure
Potential
algorithms,
for
andcost
data
reduction,
costs,
to apply
etc. revenue
new
“Arms
techniques
gains,
race” with
improved
Prudential
other risk
financial
regulations,
management
institutions
data reporting,
and firms best execution, AML, etc

Supply factors Demand factors

There is also demand due to regulatory compliance.23 New regulations have increased the
need for efficient regulatory compliance, which has pushed banks to automate 24 and adopt
new analytical tools that can include use of AI and machine learning. Financial institutions are
seeking cost effective means of complying with regulatory requirements, such as prudential

21
Examples include the FICO score in the US and Canada, Schufa scoring in Germany and firm-specific scores in Japan.
22
Finextra and Intel (2017).
23
From the perspective of technology companies offering AI and machine learning solutions for regulatory compliance by
financial institutions, these factors can be considered supply-side.
24
It is important to distinguish between ‘automation’ or ‘intelligent automation’ and use of AI and machine learning. This is
a distinction that goes back to the use of algorithms to spread out orders in the 1990s. Most of those rule-based algorithms
are not considered machine learning algorithms by the financial community, in contrast to the types of algorithms used
today that do ‘learn’ from financial data, news and other data sources.

1
regulations, data reporting, best execution of trades, and rules on anti-money laundering and
combating the financing of terrorism (AML/CFT). Correspondingly, supervisory agencies are
faced with responsibility for evaluating larger, more complex and faster-growing datasets,
necessitating more powerful analytical tools to better monitor the financial sector. Figure 4
shows how these supply and demand factors fit together.
A number of developments could impact future adoption of a broad range of financial
applications of AI and machine learning. These developments include continued growth in the
number of data sources and the timeliness of access to data; growth in data repositories, data
granularity, variety of data types; and efforts to enhance data quality. Continued improvement
in hardware, as well as AI and machine learning software as a service, including open-source
libraries, will also impact continued innovation. Development in hardware includes
processing chips and quantum computing that enable faster and more powerful AI. These
developments could enable cheaper and broader access to AI and machine learning tools that
are increasingly powerful. They could make more sophisticated real-time insights possible on
larger datasets, such as real-time databases of online user behaviour or internet-of-things (IoT)
sensors located around the world.
At the same time, sophisticated software services are becoming more widely available. Some
of the software services are open source libraries made available in the past few years that
provide researchers with off-the-shelf-tools for machine learning. There are also a growing
variety of vendors that provide machine learning for financial market participants, including
some firms that scrape news and/or metadata and enable users to identify the specific features
(webpages viewed, etc.) that correlate with the events they are interested in predicting. As
services emerge to provide, clean, organise, and analyse these data for financial insights, the
cost to users of incorporating sophisticated insights may fall significantly. Thus, at the same
time, risks related to multiple users of the same information and techniques across the
financial sector could grow (see section 4).
The legal framework for relevant data will likely also impact the adoption of AI and machine
learning tools. Breaches of personal data or uses of data that are not in the interests of
consumers may be expected to lead to added data protection legislation (see annex A). In
addition, the development of new data standards, new data reporting requirements, or other
institutional changes in financial services can impact the adoption of AI and machine learning
in specific markets.

3. Selected use cases

Financial stability implications depend critically on the uses of AI and machine learning. To
assess these implications, questions to be considered would include which AI and machine
learning tools are being used to make which types of decisions, on what time scales, to
address which financial functions, and where and at what level human involvement is being
integrated.
AI and machine learning are being adopted for a number of purposes across the financial
system. Examples include:
• Sentiment indicators: Social media data analytics companies use AI and machine
learning techniques to provide ‘sentiment indicators’ to a number of financial
1
services

1
 players. Investor sentiment indicators are being developed and sold to banks, hedge
funds, high-frequency trading traders, and social trading and investment platforms.25
• Trading signals: Machine learning can help firms to increase productivity and to
reduce costs by quickly scanning and making decisions based on more sources of
information than a human can (see section 3.3). Therein also lies a limitation of
machine learning technology: by identifying and relying on patterns that were
predictive of outcomes in the past, these tools are susceptible to false information. 26
For example, there were market moves across equities, bonds, foreign exchange, and
commodities in April 2013 after trading algorithms reacted to a fraudulent news
Tweet announcing two explosions at the White House. These types of issues may be
exacerbated with more widespread use of machine learning.27
• AML/CFT and fraud detection: Seeking to increase productivity and simultaneously
reduce costs and risks, while complying with regulations, some firms use AI for
AML/CFT and fraud detection at financial institutions. 28 Firms can also use machine
learning for credit monitoring and risk mitigation purposes, (see section 3.4).
This section considers four sets of use cases of AI and machine learning. These are: (i)
customer-focused (or ‘front-office’) uses, including credit scoring, insurance, and client-
facing chatbots; (ii) operations-focused (or ‘back-office’) uses, including capital optimisation,
model risk management and market impact analysis; (iii) trading and portfolio management in
financial markets; and (iv) uses of AI and machine learning by financial institutions for
regulatory compliance (‘RegTech’) or by public authorities for supervision (‘SupTech’). For
each of the use cases, a few examples of active or potential use cases are given, alongside
estimates (where available) of the current adoption of technologies. The implications at the
micro and macro level are reserved for sections 4 and 5.

3.1 Customer-focused uses: credit scoring, insurance and client-facing chatbots

AI and machine learning are already being applied in the front office of financial institutions.
Large-scale client data are fed into new algorithms to assess credit quality and thus to price
loan contracts. Similarly, such data can help assess risks for selling and pricing insurance
policies. Finally, client interactions may increasingly be carried out by AI interfaces with so-
called ‘chatbots,’ or virtual assistance programs that interact with users in natural language.
This section considers each in turn.

25
See IOSCO (2017), p. 28. Social trading refers to a range of trading platforms that allow users to compare trading
strategies or copy the trading strategy of other investors (see glossary).
26
Of course, humans are also prone to error and manipulation. Machine learning algorithms may be superior to humans in
detecting patterns that have not occurred before. At the same time, the use of such algorithms for automation could allow
for a quicker and more large-scale dispersion of effects than for functions performed by individual human users.
27
Tero Karpp, and Kate Crawford (2015), “Social Media, Financial Algorithms and the Hack Crash,” Theory Culture &
Society 33(1): 73-92.
28
Bart van Liebergen (2017), “Machine Learning: A Revolution in Risk Management and Compliance?” The Capco
Institute Journal of Financial Transformation, April.

1
3.1.1 Credit scoring applications
Credit scoring tools that use machine learning are designed to speed up lending decisions,
while potentially limiting incremental risk. Lenders have long relied on credit scores to make
lending decisions for firms and retail clients. Data on transaction and payment history from
financial institutions historically served as the foundation of most credit scoring models.
These models use tools such as regression, decision trees, and statistical analysis to generate a
credit score using limited amounts of structured data. However, banks and other lenders are
increasingly turning to additional, unstructured and semi-structured data sources, including
social media activity, mobile phone use and text message activity, to capture a more nuanced
view of creditworthiness, and improve the rating accuracy of loans. Applying machine
learning algorithms to this constellation of new data has enabled assessment of qualitative
factors such as consumption behaviour and willingness to pay. The ability to leverage
additional data on such measures allows for greater, faster, and cheaper segmentation of
borrower quality and ultimately leads to a quicker credit decision. 29 However, the use of
personal data raises other policy issues, including those related to data privacy and data
protections.30
In addition to facilitating a potentially more precise, segmented assessment of
creditworthiness, the use of machine learning algorithms in credit scoring may help enable
greater access to credit. In traditional credit scoring models used in some markets, a potential
borrower must have a sufficient amount of historical credit information available to be
considered ‘scorable.’ In the absence of this information, a credit score cannot be generated,
and a potentially creditworthy borrower is often unable to obtain credit and build a credit
history. With the use of alternative data sources and the application of machine learning
algorithms to help develop an assessment of ability and willingness to repay, lenders may be
able to arrive at credit decisions that previously would have been impossible. 31 While this
trend may benefit economies with shallow credit markets, it could lead to non-sustainable
increases in credit outstanding in countries with deep credit markets.32 More generally, it has
not yet been proved that machine learning-based credit scoring models outperform traditional
ones for assessing creditworthiness.
Over the past several years, a host of FinTech start-up companies targeting customers not
traditionally served by banks have emerged. In addition to more commonly known online
lenders that lend in the United States, one firm is using an algorithmic approach to data
analysis and has expanded to overseas markets, particularly China, where the majority of
borrowers do not have credit scores. Another firm, based in London, is working to provide
credit scores for individuals with ‘thin’ credit files, using its algorithms and alternative data
sources to review loan applications rejected by lenders for potential errors. Additionally,
some companies are

29
Stefan Lessmann, Bart Baesens, Hsin-Vonn Seow, and Lyn Thomas (2015), “Benchmarking state-of-the art classification
algorithms for credit scoring: An update of research,” European Journal of Operational Research 247(1): 124-136.
30
See CGFS and FSB (2017), p. 26.
31
As an example, high-frequency online data on payments transactions can help to assess the creditworthiness of individuals
and small businesses.
32
For empirical evidence on the differential impact of credit inclusion on financial stability, see Ratna Sahay, Martin Čihák,
Papa N’Diaye, Adolfo Barajas, Srobona Mitra, Annette Kyobe, Yen Nian Mooi, and Seyed Reza Yousefi (2015),

1
“Financial Inclusion: Can It Meet Multiple Macroeconomic Goals?” IMF Staff Discussion Note 15/17. One conclusion is
that in countries with deep credit markets, greater credit-based financial inclusion can be associated with new
vulnerabilities.

1
drawing on the vast amounts of data housed at traditional banks to integrate mobile banking
apps with bank data and AI to assist with financial management and make financial
projections, which may be first steps to developing a credit history.
There are a number of advantages and disadvantages to using AI in credit scoring models. AI
allows massive amounts of data to be analysed very quickly. As a result, it could yield credit
scoring policies that can handle a broader range of credit inputs, lowering the cost of
assessing credit risks for certain individuals, and increasing the number of individuals for
whom firms can measure credit risk. An example of the application of big data to credit
scoring could include the assessment of non-credit bill payments, such as the timely payment
of cell phone and other utility bills, in combination with other data. Additionally, people
without a credit history or credit score may be able to get a loan or a credit card due to AI,
where a lack of credit history has traditionally been a constraining factor as alternative
indicators of the likelihood to repay have been lacking in conventional credit scoring models.
However, the use of complex algorithms could result in a lack of transparency to consumers.
This ‘black box’ aspect of machine learning algorithms may in turn raise concerns. When
using machine learning to assign credit scores make credit decisions, it is generally more
difficult to provide consumers, auditors, and supervisors with an explanation of a credit score
and resulting credit decision if challenged. Additionally, some argue that the use of new
alternative data sources, such as online behaviour or non-traditional financial information,
could introduce bias into the credit decision.33 Specifically, consumer advocacy groups point
out that machine learning tools can yield combinations of borrower characteristics that simply
predict race or gender, factors that fair lending laws prohibit considering in many jurisdictions
(see annex B). These algorithms might rate a borrower from an ethnic minority at higher risk
of default because similar borrowers have traditionally been given less favourable loan
conditions.
The availability of historical data across a range of borrowers and loan products is key to the
performance of these tools. Likewise, the availability, quality, and reliability of data on
borrower-product performance across a wide range of financial circumstances is also key to
the performance of these risk models. Also, the lack of data on new AI and machine learning
models, and the lack of information about the performance of these models in a variety of
financial cycles, has been noted by some authorities.34

3.1.2 Use for pricing, marketing and managing insurance policies

The insurance industry is using machine learning to analyse complex data to lower costs and
improve profitability. Since analysing data to drive pricing forms the core of insurance
business, insurance-related technology, sometimes called ‘InsurTech,’ often relies on analysis
of big data. Adoption of AI and machine learning applications in InsurTech is particularly
high in the United States, UK, Germany and China.35 Many applications involve
improvements to the

33
For a thorough treatment, see Cathy O’Neil (2016), Weapons of Math Destruction: How Big Data Increases
Inequality and Threatens Democracy, London: Allen Lane.
34
See U.S. Treasury Department (2016), “Opportunities and Challenges in Online Marketplace Lending,” May; CGFS and
FSB (2017), “FinTech Credit: Market Structure, Business Models and Financial Stability Implications,” May.

1
35
Jim Struntz (2017), “AI on the insurance frontline,” Accenture Insurance Blog, July.

1
underwriting process, assisting agents in sorting through vast data sets that insurance
companies have collected to identify cases that pose higher risk, potentially reducing claims
and improving profitability.36 Some insurance companies are actively using machine learning
to improve the pricing or marketing of insurance products by incorporating real-time, highly
granular data, such as online shopping behaviour or telemetrics (sensors in connected devices,
such as car odometers). Firms usually have access to those data through partnerships,
acquisitions, or non- insurance activities. In many cases, firms need to ask for an active
consent of the user whenever data protection regulation asks them to.
AI and machine learning applications can substantially augment some insurance sector
functions, such as underwriting and claims processing. In underwriting, large commercial
underwriting and life or disability underwriting can be augmented by AI systems based on
NLP. These applications can learn from training sets of past claims to highlight key
considerations for human decision-makers. Machine learning techniques can be used to
determine repair costs and automatically categorise the severity of vehicle accident damage. 37
In addition, AI may help reduce claims processing times and operational costs. 38 Insurance
companies are also exploring how AI and machine learning and remote sensors (connected
through the ‘internet of things’) can detect, and in some cases prevent, insurable incidents
before they occur, such as chemical spills or car accidents.
It seems likely that these methods will achieve greater adoption. According to private sector
estimates, global InsurTech investment totalled $1.7 billion in 2016. At the same time, 26 per
cent of insurers provide monetary or non-monetary support (for example, coaching) to digital
start-ups, and 17 per cent of insurers have an in-house venture capital fund or investment
vehicle targeting technology.39 While the use of machine learning has the potential to produce
more accurate pricing and risk assessment for insurance companies, there may be consumer
protection concerns that stem from potential data errors or the exclusion of some groups (see
section 5).

3.1.3 Client-facing chatbots

Chatbots are virtual assistants that help customers transact or solve problems. These
automated programmes use NLP to interact with clients in natural language (by text or voice),
and use machine learning algorithms to improve over time. Chatbots are being introduced by
a range of financial services firms, often in their mobile apps or social media. While many are
still in the trial phase, there is potential for growth as chatbots gain increasing usage,
especially among the younger generations, and become more sophisticated. The current
generation of chatbots in use by financial services firms is simple, generally providing
balance information or alerts to

36
See International Association of Insurance Supervisors, “FinTech Developments in the Insurance Industry,” 21 February
2017.
37
PWC (2016), “Top Issues: AI in Insurance: Hype or reality?” March.
38
IAIS (2017), “Report on FinTech Developments in the Insurance Industry,” February.
39
Accenture, “The Rise of Insurtech,” 2017.

1
customers, or answering simple questions. It is worth observing that the increasing usage of
chatbots is correlated with the increased usage of messaging applications.40
Chatbots are increasingly moving toward giving advice and prompting customers to act. In
addition to assisting customers of financial institutions in making financial decisions,
financial institutions can benefit by gaining information about their customers based on
interactions with chatbots. While outdated infrastructure for client data storage has slowed the
development of chatbots in financial institutions in many jurisdictions, Asian financial
institutions and regulators have developed more sophisticated chatbots that are currently in
active use. The insurance industry has also explored the use of chatbots to provide real-time
insurance advice.

3.2 Operations-focused uses

Financial institutions can use AI and machine learning tools for a number of operational (or
back-office) applications. Some of these applications include: (i) capital optimisation by
banks;
(ii) model risk management (back-testing and model validation); and (iii) market impact
analysis (modelling of trading out of big positions). This section considers each of these in turn.

3.2.1 Capital optimisation use case

Capital optimisation, or the maximisation of profits given scarce capital, is a traditional
function in running a bank that heavily relies on mathematical approaches. AI and machine
learning tools build on the foundations of computing capabilities, big data, and mathematical
concepts of optimisation to increase the efficiency, accuracy, and speed of capital
optimisation. Optimisation of bank’s regulatory capital with machine learning has been a
topic of interest from both academic and business professionals over the last several years. In
2012, a private sector observer noted that most banks said they had conducted meaningful
programmes to optimise risk-weighted assets (RWA) and had seen 5 to 15 per cent RWA
savings.41 Capital optimisation is also being done in the area of derivatives margin
optimisation such as margin valuation adjustment (MVA). 42 New regulations around clearing
and bilateral margining have increased the demand for sophisticated techniques for optimising
capital and initial margin.43
AI and machine learning could assist banks in optimising MVA, and recent research suggests
that work is being done in this area. 44 In the context of MVA optimisation, machine learning
tries to reduce the initial margin for derivatives by a combination of: (a) executing pairs of

40
According to one estimate, usage of the top 4 messaging services now surpasses the top 4 social network apps. See BI
Intelligence (2016), “Messaging apps are now bigger than social networks,” September.
41
McKinsey (2012), “Capital management, Banking’s new imperative,” McKinsey Working papers on Risk, Number 38,
November 2012.
42
Margin valuation adjustment tries to determine the funding cost of the initial margin posted for a derivatives transaction.
43
Kondratyev A, and G. Giorgidze, (2017), “MVA Optimisation with Machine Learning Algorithms, 23,” Social Science
Research Network, January 2017.
44
M Heusser and P Varhol, (2016) “An Intro to Genetic Algorithms,” InfoWorld, December 2016; Qinghai Bai (2010),
“Analysis of Particle Swarm Optimisation Algorithm,” Computer and Information Science 3(1), February. The first paper
discusses genetic algorithms, which are a method that optimises an output or outputs with successive derived equations
based on predetermined hypotheses. The second discusses particle swarm optimisation, which is a method and algorithm
1
based on swarm intelligence, derived from research on the movement behaviour of flocks of birds and schools of fish.
Both offer machine learning techniques to optimise funding costs in the constraints of the leverage ratio-implied capital
charge.

2
offsetting derivative trades; (b) executing offsetting strategies with the same dealer; (c)
novating trades from one dealer portfolio to another. Machine learning finds the best
combination of the initial margin reducing trades at a given time based on the degree of initial
margin reduction in the past under different combinations of those trades. A likely implication
of these advances in RWA and MVA optimisation is a reduction in the traditionally calibrated
regulatory capital and larger reliance on the non-optimisable capital regulatory tools.

3.2.2 Model risk management (back-testing and model validation) and stress testing
Academics and practitioners often cite back-testing and model validation as areas where
progress with AI and machine learning will likely be soon visible. 45 Banks are considering
machine learning to make sense of large, unstructured and semi-structured datasets and to
police the outputs of primary models. Back-testing is important because it is traditionally used
to evaluate how well banks’ risk models are performing. In the last years, US and European
prudential regulators focused on back-testing and validation used by banks by providing
guidance on model risk management.46 Using a range of financial settings for back-testing
allows for consideration of shifts in market behaviour and other trends, hopefully reducing the
potential for underestimating risk in such scenarios.47
Some applications are already live. For instance, one global corporate and investment bank is
using unsupervised learning algorithms in model validation. Its equity derivatives business
has used this type of machine learning to detect anomalous projections generated by its stress-
testing models. Each night, these models produce over three million computations to inform
regulatory, internal capital allocations and limit monitoring. A small fraction of these
computations are extreme, and knocked out of the normal distribution of results by a quirk of
the computation cycle or faulty data inputs. Unsupervised learning algorithms help model
validators in the ongoing monitoring of internal and regulatory stress-testing models, as they
can help determine whether those models are performing within acceptable tolerances or
drifting from their original purpose. They can also provide additional input to operational risk
models, such as the vulnerability of organisations to cyber-attacks.
Similarly, AI and machine learning techniques are also being applied to stress testing. The
increased use of stress testing following the financial crisis has posed challenges for banks as
they work to analyse large amounts of data for regulatory stress tests. One provider of AI and
machine learning tools has worked closely with a large financial institution to develop tools to
assist them in modelling their capital markets business for bank stress testing. The tools
developed aim to limit the number of variables used in scenario analysis for the loss given
default and probability of default models. By using unsupervised learning methods to review

45
Louie Woodall (2017), “Model risk managers eye benefits of machine learning,” Risk, April.
46
For example, see Federal Reserve Board (2011), “Supervisory guidance on model risk management,” SR Letter 11-7,
April; FDIC (2017), “Supervisory guidance on model risk management,” FIL-22-2017, June; ECB (2017), “Guide to the
Targeted Review of Internal Models,” February.
47
Model validation is defined in industry guidance as “the set of processes and activities intended to verify that models are
performing as expected, in line with their design objectives, and business uses [to identify] potential limitations and
assumptions, and assesses their possible impact.” See Clayton Mitchell (2016), “Model Validation: For Elements of
Determining the Accuracy of Your Model,” British Bankers Association, January.

2
large amounts of data, the tools can document any bias associated with selection of variables,
thereby leading to better models with greater transparency.

3.2.3 Market impact analysis (modelling of trading out of big positions)

AI and machine learning can complement conventional market impact models.48 Firms can
use AI to obtain more information from sparse historical models, or help identify non-linear
relationships in order flow. Machine learning can be used to create ‘trading robots’ that then
teach themselves how to react to market changes. Market impact analysis involves evaluating
the effect of a firm’s own trading on market prices. Since firms are concerned about the
impact of trades, especially of large trades, on market prices, more accurate estimation of this
impact is key to timing trades and minimising trading execution costs.
Firms are investigating using AI tools to assess the market impact of a given trade. The effect
of a firm’s own trading on market prices is notoriously hard to model, especially for less
liquid securities, where data on comparable past trades are scarce. AI tools may help by
augmenting models already in use, or by introducing a machine learning approach to
minimise trading impact on prices and liquidity. For the most active systematic funds, as
much as two-thirds of the gain on trades are estimated to be lost to market impact costs. 49 AI
tools may help by augmenting models already in use, or by introducing a machine learning
approach to minimise trading impact on prices and liquidity for trading both into and out of
large market positions, or as a part of every-day trading strategies.
Machine learning is often used to identify groups of bonds that behave similarly to each
other.50 By doing so, they can rely on many more data points, providing better estimates of
price movements when the market is thin. The resulting tool groups bonds into broad,
intuitively similar buckets and then, using cluster analysis, collects the most comparable
products together in each bucket, to score the liquidity of individual bonds.
Also, AI can be used to help identify how the timing of trades can minimise market impact.
Market impact models can be developed that describe how the effect of a trade depends on
previous trades as a starting point. The models attempt to avoid scheduling trades too closely
together to avoid having a market impact greater than the sum of its parts. These models can
be used to set out the best possible trading schedules for a range of scenarios and then tweak
the schedule as the real trade progresses, using supervised learning techniques to make the
short- term predictions determining those tweaks. Banks are also testing reinforcement
learning to teach artificial intelligence tools to react to order imbalance and queue position in
the limit order book.

48
Sebastian Day (2017), “Quants turn to machine learning to model market impact,” RISK Magazine, April.
49
Day (2017).
50
For example, one firm uses machine learning to assess the liquidity of bonds. Every bond is quantitatively measured
against a range of common features such as currency, duration, time to maturity and amount outstanding. Those
measurements determine its position within a theoretical multi-dimensional space. For example, trading 500 lots of an
obscure US Treasury bond, the tool will identify other US Treasury bonds the shortest distance away within that space.
The tool will then use their combined pool of data to calibrate the parametric model.

2
3.3 Trading and portfolio management
AI and machine learning techniques are active areas of research and development for asset
managers and trading firms. In addition to significant research and development (R&D), some
firms now use machine learning to devise trading and investment strategies. The extent to
which AI investment strategies are autonomous or incorporate human oversight varies on a
case-by- case basis. In this section, we distinguish between trading execution (primarily sell-
side) and portfolio management (buy-side).

3.3.1 AI and machine learning in trading execution

Trading firms are looking to AI and machine learning to use data to improve their ability to
sell to clients. For example, analysing past trading behaviour can help anticipate a client’s
next order. Trading generates large quantities of data, and this scale is typically required by
machine learning tools to work effectively. If the current trend to increasing use of voice-to-
text services continues, this will generate additional data from trades executed over the phone,
which can be integrated with the pool of data from electronic platforms.
AI and machine learning can more pro-actively manage risk exposures. Machine learning can
serve as a basis for risk modelling by exchanges to determine when members’ trading account
positions may have increased risk profiles that may warrant intervention. For large trading
firms such as banks, the use of a central risk trading book, or risk management techniques
based on big data analysis, have enabled these firms to manage risks and optimise their use of
capital by centralising the risks that arise from various parts of their businesses.
AI and machine learning can help compliance with trading regulations. A RegTech
application of AI to trading is voice-to-text technology powered by deep learning. This helps
firms meet pre-trade and post-trade transparency requirements for non-equity markets. 51

3.3.2 Scope for the use of AI and machine learning in portfolio management
In portfolio management, AI and machine learning tools are being used to identify new
signals on price movements and to make more effective use of the vast amount of available
data and market research than with current models. Machine learning tools work on the same
principles as existing analytical techniques used in systematic investing. The key task is to
identify signals from data on which predictions relating to price level or volatility can be
made, over various time horizons, to generate higher and uncorrelated returns.
Among asset managers, machine learning is used most extensively by systematic (‘quant’)
funds, most of which are hedge funds.52 An AI unit tends to sit within a larger team at an asset
manager to aid with portfolio construction. One view in the industry is that for AI and
machine learning to be effective, both traders and quants need to have good oversight and
understanding of the tools used. Many quant funds state that they are not comfortable fully
automating and

51
Under the Markets in Financial Instruments Directive (MiFID II), for instance, dealers designated as ‘systematic
internalisers’ will be unable to quote prices to a client for some bilateral trades unless they also simultaneously broadcast
the same quote to the wider market – a task that is realisable if the quote is provided on a screen, but more difficult via a
phone.
52
While these players have been using quantitative techniques since long before the recent advances in deep learning, they

2
are widely regarded as some of the most avid adopters of these techniques.

2
implementing a model if they cannot understand how a particular prediction is made. The
concentration of machine learning tools among quant funds reflects how machine learning is
fundamentally an approach to generating predictive power from data, which distinguishes it
from investment approaches that use greater discretion and judgment.
At the moment, machine learning likely only drives a minor subset of quant funds’ trades.
Quant funds manage on the order of $1 trillion in assets, out of total assets under management
(AUM) invested in mutual funds globally in excess of $40 trillion. The market share of quant
funds has not changed drastically in the years since the crisis, but between 2013 and 2016 the
proportion of trades carried out by quant funds, on one measure, approximately doubled from
13% to 27%.53 In turn, some portion of the trading is based on machine learning. It is hard to
quantify precisely what proportion use machine learning for several reasons:
• Firms are hesitant to share proprietary information.
• When firms share information on their use of machine learning, there is not always a
standard definition or understanding of what is included within the concept.
• Some investments or trades may be made on a discretionary basis but informed (to
varying extents) by the use of machine learning.54
One contact in the industry estimates that ‘pure’ AI and machine learning players have about
$10 billion in assets under management, but that this figure is growing rapidly.55
In addition to the use by fund managers, specialist firms are making available to asset
managers machine learning tools to gain insight from the vast volume of news and market
research available. In other cases, asset managers are themselves building indicators, using AI
capability supplied by third parties. One general issue is that useful trading signals derived
from AI and machine learning strategies may follow a decay function over time, as data are
more widely used and hence become less valuable for gaining an edge over other investors. 56

3.4 AI and machine learning in regulatory compliance and supervision

AI and machine learning techniques are being used by regulated institutions for regulatory
compliance, and by authorities for supervision. RegTech is often regarded as the subset of
FinTech that focus on facilitating regulatory compliance more efficiently and effectively than
existing capabilities.57 The total RegTech market is expected to reach $6.45 billion by 2020,

53
Tabb Group, referenced in Gregory Zuckerman and Bradley Hope (2017), “The Quants Run Wall Street Now,” Wall
Street Journal, 22 May.
54
For some quant funds, machine learning tools inform investment strategies that are implemented by a person. Other firms
provide information generated by machine learning to asset managers. For example, one firm’s machine learning engine
shows how asset prices have behaved historically in response to market events. In other cases, it appears that firms are
using machine learning systems to manage portfolios and to execute trades automatically. One firm runs an automated
fund using an evolutionary computation approach, using a large network of central processing units to randomly generate
trillions of trading “genes;” from which the system selects and “breeds” the best-performing 0.01%.
55
This is based on bilateral discussions by the FSB secretariat with an investor focused on AI and machine learning funds.
56
See for example Luke Smolinski (2017), “Wolfe aims to shake up research with AI push,” Risk.net, 23 May.
57
FCA (2015), “Call for Input: Supporting the development and adoption of RegTech,” November; Institute of International
Finance (2015), “Regtech: exploring solutions for regulatory challenges,” October.

2
growing at a compound annual growth rate (CAGR) of 76%. 58 SupTech is the use of these
technologies by public sector regulators and supervisors. Within SupTech, the objective of AI
and machine learning applications is to enhance efficiency and effectiveness of supervision
and surveillance. While there can be overlap in the terms,59 the two applications are discussed
here separately. Some of the examples below are from the academic community. While not
yet being applied by regulatory or supervisory bodies, they represent potential applications in
this sector. The use cases are grouped by the function for which they are used, namely
regulatory compliance; regulatory reporting and data quality; monetary policy and systemic
risk analysis; and surveillance and fraud detection.

3.4.1 RegTech: applications by financial institutions for regulatory compliance

For analysing unstructured data, RegTech can use machine learning combined with NLP.
Besides being applied to the monitoring of behaviour and communication of traders for
transparency and market conduct, machine learning together with NLP can interpret data
inputs such as e-mails, spoken word, instant messaging, documents, and metadata. This in
turn begs the issue of the boundaries for the employee surveillance policy. Some regulated
institutions are experimenting with cases seeking to enhance their ability to comply with
product suitability requirements.
NLP could be used by asset management firms to cope with new regulations. In the EU,
investment managers have to comply with specific requirements in the Markets in Financial
Instruments Directive (MiFID II), the Undertakings for Collective Investments in
Transferrable Securities (UCITS) Directive, and the Alternative Investment Fund Managers
Directive (AIFMD). Firms could potentially leverage NLP and other machine learning tools
to interpret these regulations into a common language. They could then analyse and codify the
rules for automation into the integrated risk and reporting systems to help firms comply with
the regulations. This could bring down the cost, effort and time needed to interpret and
implement new and updated regulations for fund managers.
Knowing the identity of customers (‘know your customer’ or KYC) is another area where AI
and machine learning are applied to address one of the biggest pain points in the financial
industry, both with regards to user experience and regulator expectations. The KYC process is
often costly, laborious, and highly duplicative across many services and institutions. Machine
learning is increasingly used in remote KYC of financial services firms to perform identity
and background pre-checks. It is predominantly used in two ways: (1) evaluating whether
images in identifying documents match one another, and (2) calculating risk scores on which
firms determine which individuals or applications need to receive additional scrutiny.
Machine learning-based risk scores are also used in ongoing periodic checks based on public
and other data sources, such as police registers of offenders and social media services. Use of
these sources may enable risk and trust to be assessed quickly and often cheaply. Firms can
use risk scores on the probability of customers raising “red flags” on KYC checks to help
make decisions on whether to proceed with the time and expense of a full background
check.

58
Frost & Sullivan (2017), “Global Forecast of RegTech in Financial Services to 2020,” March.
59
For instance, some authors use RegTech to refer to applications used by regulators, as well. See Douglas W. Arner, Jànos
Barberis and Ross P. Buckley (2017), “FinTech, RegTech and the Reconceptualization of Financial Regulation,”

2
Northwestern Journal of International Law and Business 37(3): 371-413.

2
Nonetheless, concerns about their accuracy have kept some financial services from
incorporating these tools.

3.4.2 Uses for macroprudential surveillance and data quality assurance

AI and machine learning methods may help to improve macroprudential surveillance by
automating macroprudential analysis and data quality assurance. A series of new reporting
requirements across jurisdictions has led to a greater volume and frequency of reported data,
as well as greater resources required from financial institutions to complete reporting on time.
In some cases (for example, transactions data in MiFID, AIFMD templates, etc.), the volume
of data received can be challenging for the authorities receiving the data, such that it cannot
be used to its full potential using traditional methods. Moreover, substantial errors, blank
fields, and other data quality issues are often more prevalent in new datasets, and additional
checks and data quality assurance are needed. Machine learning can help improve data
quality, for example, by automatically identifying anomalies (potential errors) to flag them to
the statistician and/or the data-providing source. This may allow for both lower-cost and
higher- quality reporting and more efficient and effective data processing and macroprudential
surveillance of data by authorities.60
Similarly, AI and machine learning could help trade repositories (TRs) tackle data quality
issues, increasing the value of TR data to authorities and the public. Authorities report that
overcoming data quality issues continues to be a key challenge to making full use of TR
data.61 Application of machine learning techniques may help TRs – for over-the-counter
(OTC) derivatives or (where applicable) other types of transactions, such as exchange-traded
derivatives or securities financing transactions – improve data quality. Specifically,
appropriately trained machine learning algorithms could help identify data gaps, data
inconsistencies, and fat-finger errors, as well as match likely pairs of transactions and/or
interpolate missing data. The same techniques can be used by authorities, themselves. In this
context, the Autorité des marchés financiers du Québec reports that it has successfully tested
in its FinTech Laboratory a supervised machine learning algorithm able to recognise distinct
categories from unstructured free text fields in OTC derivatives data, such as the floating leg
of swaps. Implementation of alerts based on this algorithm is underway to automatically
detect transactions that are not compliant with mandatory clearing requirements. 62

3.4.3 SupTech: uses and potential uses by central banks and prudential authorities
Machine learning can be applied to systemic risk identification and risk propagation channels.
Specifically, NLP tools may help authorities to detect, measure, predict, and anticipate,
among other things, market volatility, liquidity risks, financial stress, housing prices, and
unemployment.63 In a recent Banca d’Italia (BdI) study, still in progress, textual sentiment

60
In addition to the applications of AI and machine learning, there are a number of potential applications of distributed
ledger technology (DLT), cloud computing and digital identity to regulatory reporting. These RegTech applications are
beyond the scope of this paper. See IIF (2017) for more detail.
61
See e.g. FSB (2017), Review of OTC derivatives market reforms: Effectiveness and broader effects of the reforms, at p. 28
62
AMF (2017), “AMF creates Fintech lab and signs partnership with R3,” press release, April.
63
David Bholat, Stephen Hansen, Pedro Santos, and Cheryl Schonhardt-Bailey (2015), Text mining for central banks, Bank
of England CCBS Handbook No. 33.

2
derived from Twitter posts is used as a proxy for the time-varying retail depositors’ trust in
banks. The indicator is used to challenge the predictions of a banks’ retail funding model, and
to try to capture possible threats to financial stability deriving from an increase of public
distrust in the banking system. Furthermore, at the BdI, in order to extract the most relevant
information available on the web, newspaper articles are processed through a suitable NLP
pipeline that evaluates their sentiment. In another study, academics developed a model using
computational linguistics and probabilistic approaches to uncover semantics of natural
language in mandatory US bank disclosures. The model found risks as early as 2005 related to
interest rates, mortgages, real estate, capital requirements, rating agencies and marketable
securities.64 Other studies are able to predict and anticipate market outcomes and economic
conditions, including volatility65 and growth.66
Use of machine learning combined with NLP can be used to identify patterns for further
attention from supervisors in large and complex data. Machine learning can also be used with
NLP to link trading databases to other information on market participants. This could include,
for example, the ability to integrate and compare trading activity information with behavioural
data like communications and to compare normal trading scenarios with those that may have
substantial deviations, triggering the need for further analysis.67
Central banks can use AI to assist with monetary policy assessments. A 2015 survey of
central banks’ use of and interest in big data reported, among other things, that central banks
expected a growing use of big data for macroeconomic and financial stability purposes. The
most prevalent expected use was for economic forecasting, in particular for economic
indicators such as inflation and prices. For instance, 39% of central banks expect to ‘nowcast,’
or predict in real time, retail home prices using big data. AI can be used to forecast
unemployment, GDP, industrial production, retail sales, tourism activity, and the business
cycle (for example, with sentiment indicators and nowcasting techniques).68, 69
Recent research highlights how these methods could be used. Researchers at Columbia
University have recently combined newly developed machine learning approaches with
observational studies to enable public authorities and market participants to: (i) ‘score’ policy
choices and link them to indicators of financial sector performance; (ii) simulate the impact of
policies under varying economic and political conditions; and (iii) detect the rate of change of

64
See Kathleen Weiss Hanley and Gerard Hoberg (2016), “Dynamic Interpretation of Emerging Systemic Risks,” working
paper, October.
65
See Harry Mamaysky and Paul Glasserman (2016), “Does Unusual News Forecast Market Stress?” Columbia Business
School Research Paper No. 15-70, April.
66
See Samuel Fraiberger (2016), “News Sentiment and Cross-Country Fluctuations,” February.
67
See Bart van Liebergen, “Machine Learning: A Revolution in Risk Management and Compliance?” The CAPCO Institute
Journal of Financial Transformation, 2017. See also, Singapore Monetary Authority development of algorithms to detect
and identify trading accounts suspected of syndicated activity, Ravi Menon, Managing Director, MAS, “Financial
Regulation – The Forward Agenda,” March 20, 2017.
68
See Irving Fisher Committee on Central Bank Statistics (2015), “Central banks’ use of and interest in “big data,”” October.
69
A growing body of existing research done by academics and others suggests that machine learning tools do in fact make it
possible to better detect, measure, predict, anticipate, and even nowcast market outcomes. See the following for more
examples: Cindy K. Soo (2013), “Quantifying Animal Spirits: News Media and Sentiment in the Housing Market,”
University of Pennsylvania The Wharton School, January; Hal Varian and Hyunyoung Choi (2009), “Predicting the
Present with Google Trends” Google Research Blog, April.

2
market innovation by comparing trends of policy efficacy over time.70 With the aim of
studying the redistributive effects of fiscal policy over different municipalities, a study from
the BdI employs a dynamic factor model and utilises a dataset containing variables from
different sectors of the economy. In order to select the statistically most relevant independent
variables they use automatic regression variable selection.71 At the Office of Financial
Research (OFR), researchers are evaluating the potential for machine learning tools to identify
new financial innovations receiving more attention from market participants in financial
publications. OFR researchers have also used machine learning to extract sentiment and key
topics from financial publications in order to evaluate the relationship between news,
attention, and financial stability.

3.4.4 Uses by market regulators for surveillance and fraud detection

Some regulators are using AI for fraud and AML/CFT detection. The Australian Securities
and Investments Commission (ASIC) has been exploring the quality of results and potential
use of NLP technology to identify and extract entities of interest from evidentiary documents.
ASIC is using NLP and other technology to visualise and explore the extracted entities and
their relationships. In order to fight criminal activities carried out through the banking system
(such as money laundering), BdI collects detailed information on bank transfers and correlates
this information with information from newspaper articles. The correlation involves both
structured and unstructured data for file sizes of more than 50 gigabytes. In the same vein, the
Monetary Authority of Singapore (MAS) is exploring the use of AI and machine learning in
the analysis of suspicious transactions to identify those transactions that warrant further
attention, allowing supervisors to focus their resources on higher risk transactions.
Investigating suspicious transactions is time consuming and often suffers from a high rate of
false positives, due to defensive filings by regulated entities. Machine learning is being used
to identify complex patterns and highlight the suspicious transactions that are potentially more
serious and warrant closer investigation. Coupled with machine learning methods to analyse
the granular data from transactions, client profiles, and a variety of unstructured data, machine
learning is being explored to uncover non-linear relationships among different attributes and
entities, and to detect potentially complicated behaviour patterns of money laundering and the
financing of terrorism not directly observable through suspicious transactions filings from
individual entities.
Market regulators can also use these techniques for disclosure and risk assessment. The US
Securities and Exchange Commission (SEC) staff leverages “big data” to develop text
analytics and machine learning algorithms to detect possible fraud and misconduct. Certain
risk assessment tools are beginning to move into the AI space.72 For instance, the SEC staff
uses

70
Sharyn O’Halloran, Sameer Maskey, Geraldine McAllister, David K. Park and Kaiping Chen (2015), “Big Data and the
Regulation of Financial Markets,” IEEE/ACM International conference on Advances in Social Networks Analysis and
Mining.
71
Monica Andini, Emanuele Ciani, Guido de Blasio, Alessio D’Ignazio, and Viola Salvestrini (2017), “Targeting policy-
compliers with Machine Learning: An application to a tax rebate program in Italy,” Banca d’Italia working paper,
forthcoming.
72
The SEC staff has noted that it extracts words and phrases from narrative disclosures in forms and filings and using
human written rules to define patterns in document to systematically measure and assess how emerging growth
companies are availing themselves of JOBS Act provisions. See Scott Bauguess (2017), “The Role of Big Data, Machine

3
Learning and

3
machine learning to identify patterns in the text of SEC filings. With supervised learning,
these patterns can be compared to past examination outcomes to find risks in investment
manager filings. The SEC staff notes that these techniques are five times better than random at
finding language that merits a referral to enforcement. While the results can generate false
positives that can be explained by non-nefarious actions and intent, these nonetheless provide
increasingly important signals to prioritise examination.73 For investment advisers, the SEC
staff compiles structured and unstructured data. Unsupervised learning algorithms are used to
identify unique or outlier reporting behaviours – including both topic modelling and tonality
analysis.74 The output from this first stage is then combined with past examination outcomes
and fed into a second-stage, machine learning algorithm to predict the presence of
idiosyncratic risks at each investment advisor.75 In Australia, ASIC has also used machine
learning software to identify misleading marketing in a particular sub-sector, such as
unlicensed accountants in the provision of financial advice.76

4. Micro-financial analysis
From a micro-financial point of view, the application of AI and machine learning to financial
services may have an important impact on financial markets, institutions and consumers. 77 In
this section, potential changes to incentives and behaviour and how they may affect financial
stability, for better or worse, are considered.

4.1 Possible effects of AI and machine learning on financial markets

Since AI and machine learning have the potential to substantially enhance the efficiency of
information processing, thereby reducing information asymmetries, applications of AI and
machine learning have the potential to strengthen the information function of the financial
system.78 The mechanisms whereby this improvement may occur include:79
a) AI and machine learning may enable certain market participants to collect and analyse
information on a greater scale. In particular, these tools may help market participants to
understand the relationship between the formulation of market prices and various factors,

AI in Assessing Risks: a Regulatory Perspective,” Speech by Acting Director and Acting Chief Economist, Division of
Economics and Risk Analysis, OpRisk North America, June 21.
73
See Bauguess (2017). See also Gerard Hoberg and Craig M. Lewis (2015), “Do Fraudulent Firms Produce Abnormal
Disclosure?” Working paper.
74
Topic modelling lets the data define the themes of each filing. Tonality analysis gauges the negativity of a filing by
counting terms with a negative connotation. See Bauguess (2017).
75
See Bauguess (2017).
76
See Greg Medcraft (2017), “The Fourth Industrial Revolution: Impact on Financial Services and markets,” speech, March.
77
See FSB FinTech Issues Group (2017), p. 14; 18-19 for a description of comparable impact that FinTech has generally.
78
For a seminal reference on the functions of the financial system, see Robert Merton and Zvi Bodie (2005), “Design of
financial systems: Towards a synthesis of function and structure,” Journal of Investment Management 3(1): 1–23.
79
The Bank of Japan held an AI conference on April 13, 2017, and the market views described in this chapter are in line
with those expressed in this conference.

3
 such as in sentiment analysis. This could reduce information asymmetries and thus
contribute to the efficiency and stability of markets.80
b) AI and machine learning may lower market participants’ trading costs. Moreover, AI and
machine learning may enable them to adjust their trading and investment strategies in
accordance with a changing environment in a swift manner, thus improving price
discovery and reducing overall transaction costs in the system.

Nonetheless, if many market participants come to use similar AI and machine learning
programmes in areas such as credit scoring or financial market activities, the consequent
correlated risks may entail financial stability risks. If machine learning-based traders
outperform others, this could in the future result in many more traders adopting similar
machine learning strategies (even if this may also reduce the profitability of such strategies).
While there is no evidence of this occurring to date, this could become relevant with greater
adoption of such trading strategies. As with any herding behaviour in the market, this has the
potential to amplify financial shocks. Moreover, advanced optimisation techniques and
predictable patterns in the behaviour of automated trading strategies could be used by insiders
or by cybercriminals to manipulate market prices.81

4.2 Possible effects of AI and machine learning on financial institutions

AI and machine learning have the potential to enhance the efficiency and profitability of
financial institutions, while reducing their costs and risks, through various channels. Greater
profitability could aid the build-up of buffers and ultimately benefit system-wide stability:
a) AI and machine learning may enhance machine-based processing of various operations in
financial institutions, thus increasing revenues and reducing costs. For example, if AI and
machine learning help to identify customers’ needs and better target or tailor products to
profitable customers, financial institutions could more efficiently allocate resources toward
serving those customers that account for substantial fees or have the potential for future
growth. Automating routine business processes may allow for lower operating costs.
b) AI and machine learning can be used for risk management through earlier and more
accurate estimation of risks. For example, to the extent that AI and machine learning
enable decision- making based on past correlations among prices of various assets,
financial institutions could better manage these risks. Tools that mitigate tail risks could be
especially beneficial for the overall system. Also, AI and machine learning could be used
for anticipating and detecting fraud, suspicious transactions, default, and the risk of cyber-
attacks, which could result in better risk management. But AI and machine learning based
tools might also miss new types of risks and events because they could potentially
‘overtrain’ on past events. While AI and machine learning tools hold potential to improve
risk management, the recent deployment of these strategies means that they remain
untested at addressing risk under shifting financial conditions.

80
See IOSCO (2017), p. 28.
81
Rachel Wolcott (2017), “‘Hacking the algo:’ when automated traders are victims, not villains,” Thomson Reuters
Regulatory Intelligence, August.

3
c) The data intensity and open-source character of research in AI and machine learning may
encourage collaboration between financial institutions and other industries, such as e-
commerce and sharing economy businesses.
Nonetheless, use of AI and machine learning risks creating ‘black boxes’ in decision-making
that could create complicated issues, especially during tail events. In particular, it may be
difficult for human users at financial institutions – and for regulators – to grasp how
decisions, such as those for trading and investment, have been formulated. 82 Moreover, the
communication mechanism used by such tools may be incomprehensible to humans, thus
posing monitoring challenges for the human operators of such solutions.83 If in doubt, users of
such AI and machine learning tools may simultaneously pull their ‘kill switches,’ that is
manually turn off systems. After such incidents, users may only turn systems on again if other
users do so in a coordinated fashion across the market. This could thus add to existing risks of
system-wide stress and the need for appropriate circuit-breakers.
In addition, if AI and machine learning based decisions cause losses to financial
intermediaries across the financial system, there may be a lack of clarity around
responsibility.84 For example, if a specific AI and machine learning application developed by
a third party resulted in large losses, is the institution that conducted the trading solely
responsible for the losses? Or would regulators or other parties be able to pursue potential
claims against the application developer? Could more widespread use of AI and machine
learning, including by non-traditional market players, impact the nature of supervision?
Furthermore, there are open questions about (identifying) potential collusion among trading
applications that rely on deep learning. Specifically, if algorithms interact in ways that would
be considered collusion if done by human agents, then as with human agents, proof of intent
may be an issue. In this light, there may be a number of legal uncertainties (see annex A).
Finally, the lack of transparency around applications may be problematic for both institutions
and regulators when it may not be possible to understand how undesired events occurred and
when steps may need to be taken to prevent a recurrence.
Any uncertainty in the governance structure in the use of AI and machine learning might
increase the risks to financial institutions. 85 If each investor makes their investment without
fully understanding the applications and his or her possible losses in tail events, the aggregate
risks could be underestimated. In addition, any uncertainty in the governance structure could
substantially increase the costs for allocating losses, including the possible costs of litigation.
In this regard, financial institutions applying AI and machine learning to their businesses need
to establish well-designed governance and maintain auditability.

82
For an article concisely describing the problems of black boxes in AI decision-making, see Will Knight (2017), “The
Dark Secret at the Heart of AI,” MIT Technology Review, April.
83
For example, the recent publicity around the Facebook AI agents illustrates this possibility. See Andrew Griffin (2017),
“Facebook’s AI creating its own language is more normal than people think, researchers say,” The Independent, 3
August.
84
Several regulators argue that final responsibility always lies at the regulated entity, who should perform robust due
diligence for all contracted services. In many jurisdictions, financial entities may contract services from third-party
providers but remain responsible for compliance with relevant rules. See BCBS (2017), “Implications of fintech
developments for banks and bank supervisors - consultative document,” August.
85
For a central banker’s speech illustrating the issue of AI and its governance, see Haruhiko Kuroda (2017), “ AI and the
Frontiers of Finance,” speech by the Governor of the Bank of Japan at the Conference on “AI and Financial
Services/Financial Markets,” Tokyo, April.
3
Finally, there may be important third-party dependencies. In the development of AI and
machine learning to date there is a high reliance on a relatively small number of third-party
technological developers and service providers. This third-party reliance could be relevant for
market participants and financial institutions in the future. For instance, if a major provider of
AI and machine learning tools were to become insolvent or suffer an operational risk event,
this could lead to operational disruptions at a large number of financial institutions at the same
time. These risks may become more important in the future if AI and machine learning are
used for ‘mission-critical’ applications of financial institutions.

4.3 Possible effects of AI and machine learning on consumers and investors

If AI and machine learning reduce the costs and enhance the efficiency of financial services,
consumers could obtain a number of benefits.
a) Consumers and investors could enjoy lower fees and borrowing costs if AI and machine
learning reduce the costs for various financial services.
b) Consumers and investors could have wider access to financial services. For example,
applications of AI for robo-advice might facilitate people’s use of various asset markets for
their investments. Moreover, AI and machine learning, through advanced credit scoring for
FinTech lending, might make wider sources of funds available to consumers and small and
medium enterprises (SMEs).
c) AI and machine learning could facilitate more ‘customised’ and ‘personalised’ financial
services through big data analytics. For example, AI and machine learning might facilitate
the analysis of big data, thus clarifying the characteristics of each consumer and/or investor
and allowing firms to design well-targeted services. Nonetheless, the use of consumers’
data may entail issues of data privacy and information security. 86 Moreover, since AI and
machine learning analytics could analyse the characteristics of each customer through
public data, it would be necessary to consider how the output of customer analysis should
be protected, while protecting the anonymity of each consumer and facilitating the safe and
efficient use of big data for better services. In addition, establishing well-designed
governance structures for financial service providers using AI and machine learning would
be important for consumer and investor protection purposes.
Avoiding discrimination in credit scoring, credit provision, and insurance is also an important
topic. Even where data on sensitive characteristics such as race, religion, gender, etc. are not
collected, AI and machine learning algorithms may create outcomes that implicitly correlate
with those indicators, for example, based on geography or other characteristics of individuals.
There is ongoing research on how to address and mitigate these biases. This is a key area in
the broader discussion on AI ethics (see annex B).

86
Since AI and machine learning analytics could analyse the characteristics of each customer through public data, it would
be necessary to consider how the output of customer analyses and protecting the anonymity of each consumer and
facilitating the safe and efficient use of big data for better services. In addition, establishing well-designed governance
structures for financial service providers using AI and machine learning would be important for consumer protection
purposes. On issues of data privacy and information security, see Haruhiko Kuroda (2016), “Information Technology and
Financial Services: The Central Bank’s perspective,” Remarks by Governor Kuroda at the FinTech Forum, August 23.

3
4.4 Current regulatory considerations regarding the use of AI and machine learning
Because AI and machine learning applications are relatively new, there are no known
dedicated international standards in this area. Yet in light of some of the potential risks
identified above, a few efforts by international standards-setters and similar international fora
of regulators deserve note. For example, several international standards-setters have
considered risks associated with algorithmic trading, as it has become a pervasive feature of
markets that may, among other things, amplify systemic risk. Examples include the following:
- The International Organization of Securities Commissions (IOSCO) reported on the
impact of new technologies including algorithmic trading on market surveillance, and
made recommendations to consider, including for data collection and cross-border
cooperation.87
- The Senior Supervisors’ Group (SSG), a forum for senior representatives of
supervisory authorities from around the world, issued principles for supervisors to
consider when assessing practices and key controls over algorithmic trading activities
at banks.88
Some national regulators note that, from a supervisory perspective, firms developing
algorithmic models based on AI and machine learning should have a robust development
process in place. They need to ensure that possible risks are considered at every stage of the
development process. This is particularly important in order to avoid market abuse and
prevent the strategy from contributing to, or causing, disorderly market behaviour.89 This
requirement is part of MiFID II, which will come into force in the first quarter of 2018 in
Europe. There are similar requirements for algorithms imposed on certain regulated entities
by a US securities self-regulatory organisation.90
Similarly, the Basel Committee on Banking Supervision (BCBS) notes that a sound
development process should be consistent with the firm’s internal policies and procedures and
deliver a product that not only meets the goals of the users, but is also consistent with the risk
appetite and behavioural expectations of the firm. In order to support new model choices,
firms should be able to demonstrate developmental evidence of theoretical construction;
behavioural characteristics and key assumptions; types and use of input data; numerical
analysis routines and specified mathematical calculations; and code writing language and
protocols (to replicate the model). Finally, it notes that firms should establish checks and
balances at each stage of the development process.91

87
IOSCO (2011), “Regulatory Issues Raised by the Impact of Technological Changes on Market Integrity and Efficiency,”
July; IOSCO (2013), “Technological Challenges to Effective Market Surveillance Issues and Regulatory Tools,” April
88
Senior Supervisors’ Group (2015), “Algorithmic Trading Briefing Note,” April.
89
These statements were made in discussions at the two workshops and in internal discussions in the FIN.
90
See FINRA (2015), “FINRA Rule 3110 (Supervision),” June.
91
See BCBS (2011), “Principles for the Sound Management of Operational Risk,” June.

3
5. Macro-financial analysis
Widespread adoption of AI and machine learning could impact the financial system in a
number of ways, depending on the nature of the application. From the perspective of
economic growth, the application of AI and machine learning to financial services has
potential to enhance the efficiency of the economy and to contribute to growth through the
following mechanisms:92
(a) Enhancing the efficiency of financial services: more efficient risk management of
individual banks’ loan portfolio and insurers’ liabilities may benefit the aggregate system.
AI and machine learning could help process information on the fundamental value of
assets, thus allocating funds to investors and projects more effectively. Moreover, if AI
and machine learning increase the speed and reduce the costs of payment and settlement
transactions, for example by executing trades at times when there are available
counterparties with corresponding demand, this may stimulate transactions for real
economic activities.
(b) Facilitating collaboration and realising new ‘economies of scope:’ Were AI and machine
learning to facilitate collaboration between financial services and various industries, such
as e-commerce and ‘sharing economy’ industries, this could realise new economies of
scope and foster greater economic growth. For example, customer analysis based on
transaction data attached to payment and settlement activities (for example, “who buys
what, when, and where?”) would encourage cooperation between e-commerce and
financial services.
(c) Stimulating investments in AI and machine learning related areas: Many firms, including
non-financial businesses, appear eager to apply AI and machine learning to their business.
The growth in investments in AI and machine learning-related R&D can directly
contribute to economy-wide investment and thus stimulate economic growth.
From a macro-financial viewpoint, the short- to medium-term effects of the adoption of AI
and machine learning on financial structure and markets could be more mixed. There are a
number of potential effects on the systemic importance of market participants, the degree of
concentration, and market vulnerabilities, which are elaborated below.

5.1 Market concentration and systemic importance of institutions

AI and machine learning may affect the type and degree of concentration in financial markets
in certain circumstances. For instance, the emergence of a relatively small number of
advanced third-party providers in AI and machine learning could increase concentration of
some functions in the financial system. Similarly, access to big data could be a source of
systemic importance, especially if firms are able to leverage their proprietary sources of big
data to obtain substantial economies of scope. Finally, the most innovative technologies may
be mainly affordable to large companies because the development of uses requires significant
investments (for acquiring and maintaining the infrastructure and the skilled workers).
For the possible impact of AI and machine learning on banks’ systemic importance, there are
a number of key scenarios. If AI and machine learning ‘unbundle’ traditional banking
services

3
92
For a central bank’s view focusing on the impacts of AI on the economy more broadly, see Carolyn Wilkins (2017),
“Blame It on the Machines?” speech to the Toronto Region Board of Trade, Toronto, Ontario, 18 April.

3
and entice new firms to offer financial services, this might reduce the systemic importance of
individual large universal banks. These banks could focus on a more narrow set of activities,
rather than continuing to offer universal services.93 However, taken as a group, universal
banks’ vulnerability to systemic shocks may grow if they increasingly depend on similar
algorithms or data streams. On the other hand, if a large bank, which already has public trust,
successfully adopts AI and machine learning so as to strengthen its market power, its systemic
importance could increase. Whether other market participants provide similar services on
competitive terms may also be affected by market entry costs and regulation. Thus, it is
difficult to assess whether AI and machine learning would generally increase or decrease the
degree of concentration.

5.2 Potential Market Vulnerabilities

Use of AI and machine learning for trading could impact the amount and degree of
‘directional’ trading. Under benign assumptions, the divergent development of trading
applications by a wide range of market players could benefit financial stability. For example,
if machine learning- powered robo-advisors give more customised advice to individuals, their
investment activities may become more tailored to individual preferences and perhaps less
correlated with other trading strategies. By reducing the barriers to entry for retail consumers
to invest, these applications could also expand the investor base in capital markets. Similarly,
the use of AI and machine learning for new and uncorrelated trading strategies by hedge funds
could also result in greater diversity in market movements. More efficient processing of
information could help to reduce price misalignments earlier and hence mitigate the build-up
of macro-financial price imbalances.
On the other hand, new trading algorithms based on machine learning may be less predictable
than current rule-based applications and may interact in unexpected ways. To the extent that
firms using AI or machine learning techniques can generate higher returns or lower trading
costs, it is likely that incentives for adoption will increase. In the absence of data on the extent
of market-wide use, market movements may be ascribed to AI and machine learning models,
and interpretation of market shocks may be hampered. Finally, high frequency trading (HFT)
applications of AI and machine learning could be new sources of vulnerabilities. If a similar
investment strategy based on AI and machine learning is widely used in HFT, it might
increase market volatility through large sales or purchases executed almost simultaneously. 94
Regarding leverage, liquidity, and maturity transformation, the adoption of AI and machine
learning by financial market participants such as hedge funds and market makers may also
have both positive and negative impacts. AI and machine learning could increase liquidity in
financial markets through enhanced speed and efficiency of trading activities. AI and machine
learning could be used to detect excessive risks and overly-complicated transactions and to
design more effective hedging strategies for risk management by individual financial

93
See Hiroshi Nakaso (2016), “FinTech – Its Impacts on Finance, Economies and Central Banking,” speech, November 18.
94
See, e.g., FSB FinTech Issues Group (2017), p.46: “we have found no empirical evidence so far on convergence of robo-
advisors’ algorithms or portfolios.”

3
institutions.95 To the extent these tools enable the growth of new credit platforms to directly
connect lenders and borrowers (broadly called FinTech credit),96 this could reduce reliance on
bank loans, reduce banks’ leverage, and achieve a more diversified risk-sharing structure in
the overall financial system. On the other hand, to the extent that market participants use AI
and machine learning in order to minimise capital or margins or maximise expected returns on
capital (within the constraints of regulations, and without paying due attention to risks), the
use of AI and machine learning may increase risks. Specifically, it may allow for much tighter
liquidity buffers, higher leverage, and faster maturity transformation than in cases where AI
and machine learning had not been used for such optimisation.

5.3 Networks and interconnectedness

Applications of AI and machine learning may enhance the interconnectedness of financial
markets and institutions in unexpected ways. Institutions’ ability to make use of big data from
new sources may lead to greater dependencies on previously unrelated macroeconomic
variables and financial market prices, including from various non-financial corporate sectors
(e-commerce, sharing economy, etc.). As institutions find algorithms that generate
uncorrelated profits or returns, there is a risk these will be exploited on a sufficiently wide
scale that correlations actually increase. These potentially unforeseen interconnections will
only become clear as technologies are actually adopted.
More generally, greater interconnectedness in the financial system may help to share risks and
act as a shock absorber up to a point. Yet the same factors could spread the impact of extreme
shocks.97 If a critical segment of financial institutions rely on the same data sources and
algorithmic strategies, then under certain market conditions a shock to those data sources – or
a new strategy exploiting a widely-adopted algorithmic strategy – could affect that segment as
if it were a single node. This may occur even if, on the surface, the segment is made up of
tens, hundreds, or even thousands of legally independent financial institutions. As a result,
collective adoption of AI and machine learning tools may introduce new risks.

5.4 Other implications of AI and machine learning applications

AI and machine learning applications in insurance markets could reduce the degree of moral
hazard and adverse selection – but could also undermine the risk pooling function of
insurance. Moral hazard and adverse selection are inherent problems in insurance.
Nonetheless, if AI and machine learning are used to continuously adjust insurance fees in
accordance with changing behaviour of the policyholders, this may reduce moral hazard. If AI
and machine learning are utilised to offer customised insurance policies reflecting detailed
characteristics of each person, it may also decrease adverse selection. On the other hand, these
uses may pose various new challenges. For example, the more accurate pricing of risk may
lead to higher premiums for riskier consumers (such as in health insurance for individuals
with a genetic predisposition to

95
See the debates in the panel discussion entitled “FinTech and the Transformation of Financial Services” held in the
International Monetary Fund on April 19, 2017.
96
See CGFS and FSB (2017).
97
Andrew Haldane (2009), “Rethinking the financial network,” speech at the Financial Student Association, April.

4
certain diseases) and could even price some individuals out of the market. Even if innovative
insurance pricing models are based on large data sets and numerous variables, algorithms can
entail biases that can lead to non-desirable discrimination and even reinforce human
prejudices. This warrants a societal discussion on the desired extent of risk sharing, how the
algorithms are conceived, and which information is admissible.98
Meanwhile, AI and machine learning can continue to be a useful tool both for financial
institutions (RegTech) and supervisors (SupTech). Many of the uses described in section 3.4
could result in improvements in risk management, compliance, and systemic risk monitoring,
while potentially reducing regulatory burdens. Yet, if a similar type of AI and machine
learning is used without appropriately ‘training’ it or introducing feedback, reliance on such
systems may introduce new risks. For example, if AI and machine learning models are used in
stress testing without sufficiently long and diverse time series or sufficient feedback from
actual stress events, there is a risk that users may not spot institution-specific and systemic
risks in time. These risks may be pronounced especially if AI and machine learning are used
without a full understanding of the underlying methods and limitations.
Furthermore, as the current regulatory framework is not designed with the use of such tools in
mind, some regulatory practices may need to be revised for the benefits of AI and machine
learning techniques to be fully harnessed. For example, in MiFID II, where an obligation is
placed on the firm to submit a report when a reportable event occurs, regulatory compliance is
expected of the firm at all times. If AI and machine learning tools are used to deem if a
particular activity is reportable or not, mistakes would still result in regulatory action, even if
the tools can identify what information the regulators truly needs in order to reduce the risk of
market disruption. In this regard, combining AI and machine learning with human judgment
and other available analytical tools and methods may be more effective, particularly to
facilitate causal analysis.99 More generally, the greater adoption of AI, machine learning, and
other technological advances in finance may benefit also from more of a ‘systems’
perspective in financial regulation to contribute to financial stability in an increasingly
complex system.100
If optimisation solutions are adopted primarily by the private sector but not the public sector,
there may be a risk that some individuals or firms may use them more successfully to ‘game’
regulatory rules or conduct regulatory arbitrage.

6. Conclusions and implications for financial stability

The use of AI and machine learning technology is changing the provision of some financial
services. While data on the extent of adoption in various markets is quite limited, dialogue
with market participants suggests that some segments of the financial system are actively
employing AI and machine learning. These applications are thus currently more widely used
than other

98
See IAIS (2017).
99
Susan Athey (2017), “Beyond prediction: Using big data for policy problems,” Science 355(6324): 483-485.
100
Andrei A. Kirilenko and Andrew W. Lo (2013), “Moore’s Law versus Murphy’s Law: Algorithmic Trading and Its
Discontents,” Journal of Economic Perspectives, 27(2) 51–72; Susan Athey (2017), “Beyond prediction: Using big data

4
for policy problems,” Science 355(6324): 483-485.

4
key FinTech innovations, such as distributed ledger technology or smart contracts. In
particular, fraud detection, capital optimisation, and portfolio management applications appear
to be growing rapidly. Most market participants expect that AI and machine learning will be
adopted further. Because of this, it is important to start thinking about the financial stability
implications now rather than after the potential implications have been realised.101 The
analysis is necessarily partial and will benefit from greater understanding of use cases over
time. Moreover, many of the changes will not result in a material change to financial stability
and hence fall outside the scope of this report.
The use of AI and machine learning in financial services may bring key benefits for financial
stability in the form of efficiencies in the provision of financial services and regulatory and
systemic risk surveillance. The more efficient processing of information on credit risks and
lower-cost customer interaction may contribute to a more efficient financial system. The
internal (back-office) applications of AI and machine learning could improve risk
management, fraud detection, and compliance with regulatory requirements, potentially at
lower cost. In portfolio management, the more efficient processing of information from AI
and machine learning applications could help to boost the efficiency and resilience of
financial markets – reducing price misalignments earlier and (under benign assumptions)
reducing crowded trades. Finally, with use cases by regulators and supervisors, there is
potential to increase supervisory effectiveness and perform better systemic risk analysis in
financial markets.
At the same time, network effects and scalability of new technologies may in the future give
rise to additional third-party dependencies. This could in turn lead to the emergence of new
systemically important players. AI and machine learning services are increasingly being
offered by a few large technology firms. Like in other platform-based markets, there may be
value in financial institutions using similar third-party providers given these providers’
reputation, scale, and interoperability. There is the potential for natural monopolies or
oligopolies. These competition issues – relevant enough from the perspective of economic
efficiency – could be translated into financial stability risks if and when such technology firms
have a large market share in specific financial market segments. These third-party
dependencies and interconnections could have systemic effects if such a large firm were to
face a major disruption or insolvency.
Many current providers of AI and machine learning tools in financial services may fall
outside the regulatory perimeter or may not be familiar with applicable law and regulation.
Where financial institutions rely on third-party providers of AI and machine learning services
for critical functions, and rules on outsourcing may not be in place or not be understood, these
servicers and providers may not be subject to supervision and oversight. Similarly, if
providers of such tools begin providing financial services to institutional or retail clients, this
could entail financial activities taking place outside of the regulatory perimeter.
The lack of interpretability or ‘auditability’ of AI and machine learning methods has the
potential to contribute to macro-level risk if not appropriately supervised by microprudential
supervisors. Many of the models that result from the use of AI or machine learning techniques
are difficult or impossible to interpret. The lack of interpretability may be overlooked in
various situations, including, for example, if the model’s performance exceeds that
of more

4
101
See CGFS and FSB (2017); FSB (2017).

4
interpretable models. Yet the lack of interpretability will make it even more difficult to
determine potential effects beyond the firms’ balance sheet, for example during a systemic
shock. Notably, many AI and machine learning developed models are being ‘trained’ in a
period of low volatility. As such, the models may not suggest optimal actions in a significant
economic downturn or in a financial crisis, or the models may not suggest appropriate
management of long-term risks.
Should there be widespread use of opaque models, it would likely result in unintended
consequences. For example, if multiple firms develop trading strategies using AI and machine
learning models but do not understand the models because of their complexity, it would be
very difficult for both firms and supervisors to predict how actions directed by models will
affect markets. When the models’ actions interact in the marketplace, it is quite possible that
unintended, and possibly negative, consequences could result for financial markets. Similar
unintended consequences may occur in applications aimed at credit scoring, capital
optimisation, or cyber threat detection, where the build-up of risks may occur slowly.
As with the use of any new product or service, there are important issues around the
appropriate risk management and oversight of AI and machine learning. In discussions with
FSB members for this report, industry representatives noted the challenges posed by
conducting audits effectively, including sufficient skills in-house to understand and supervise
AI and machine learning models. Beyond the staff operating these applications, key functions
such as risk management and internal audit and the administrative management and
supervisory body should be fit for controlling and managing the use of applications. Yet, the
scarcity of resources with the required skills and knowledge can be an issue. 102 On the
supervisory side, auditing of models may require skills and expertise that supervisory
institutions may not currently have. Some supervisors note a need to examine specifications
developed in the scheduling and staging process of model development and to assess the
governance structure around various stages of the model after its launch.103
Assessing AI and machine learning applications for risks, including adherence to any relevant
protocols regarding data privacy, conduct risks, and cybersecurity, is important at this stage. It
is important that progress in AI and machine learning applications is accompanied with
further progress in the interpretation of algorithms’ outputs and decisions. Increased
complexities of models may strain the abilities of developers and users to fully explain,
and/or, in some instances, understand how they work. Efforts to improve the interpretability
of AI and machine learning may be important conditions not only for risk management as
noted above, but also for greater trust from the general public as well as regulators and
supervisors in critical financial services.
The uses of AI and machine learning should continue to be monitored. As the underlying
technologies develop further, there is potential for more widespread use, beyond the use cases
discussed in this report. It will be important to continue monitoring these innovations and to
update this assessment in the future.

102
FSB (2017), pp. 31-32.
103
This statement derives from discussions at the two workshops and in internal discussions in the drafting team.

4
Glossary
Algorithm: a set of computational rules to be followed to solve a mathematical problem. More
recently, the term has been adopted to refer to a process to be followed, often by a computer.
Artificial intelligence: the theory and development of computer systems able to perform
tasks that traditionally have required human intelligence.
Augmented intelligence: augmentation of human capabilities with technology, for instance by
providing a human user with additional information or analysis for decision-making.
Big data: a generic term that designates the massive volume of data that is generated by the
increasing use of digital tools and information systems.
Chatbots: virtual assistance programmes that interact with users in natural language.
Cluster analysis: A statistical technique whereby data or objects are classified into groups
(clusters) that are similar to one another but different from data or objects in other clusters.
Deep learning: a subset of machine learning, this refers to a method that uses algorithms
inspired by the structure and function of the brain, called artificial neural networks.
FinTech: technologically enabled financial innovation that could result in new business
models, applications, processes, or products with an associated material effect on financial
markets and institutions and the provision of financial services.
InsurTech: the application of FinTech for insurance markets.
Internet of things: the inter-networking of physical devices, vehicles, buildings, and other
items embedded with electronics, software, sensors, actuators, and network connectivity that
enable these objects to collect and exchange data and send, receive, and execute commands.
Machine learning: a method of designing a sequence of actions to solve a problem that
optimise automatically through experience and with limited or no human intervention.
Margin valuation adjustment: a method to determine the funding cost of the initial margin
posted for a derivatives transaction.
Natural Language Processing (NLP): An interdisciplinary field of computer science,
artificial intelligence, and computation linguistics that focuses on programming computers
and algorithms to parse, process, and understand human language.
Open source: a designation for a computer programme in which underlying source code is
freely available for redistribution and modification.
RegTech: any range of applications of FinTech for regulatory and compliance requirements
and reporting by regulated financial institutions. This can also refer to firms that offer such
applications, and in some cases can encompass SupTech (see below).
Reinforcement learning: a subset of machine learning in which an algorithm is fed an
unlabelled set of data, chooses an action for each data point, and receives feedback (perhaps
from a human) that helps the algorithm learn.
Robo-advisors: applications that combine digital interfaces and algorithms, and can also
include machine learning, in order to provide services ranging from automated financial
recommendations to contract brokering to portfolio management to their clients, without or

4
with very limited human intervention. Such advisors may be standalone firms and platforms,
or can be in-house applications of incumbent financial institutions.
Social trading: a range of trading platforms that allow users to compare trading strategies or
copy the trading strategy of other investors. The latter is often referred to as ‘copy trading’ or
‘mirror investing.’
SupTech: applications of FinTech by supervisory authorities.
Supervised learning: a subset of machine learning in which an algorithm is fed a set of
‘training’ data that contains labels on the observations.
Terabyte: a unit of data storage, equal to one trillion (1012) bytes, or 1,000 gigabytes.
Tonality analysis: a method to gauge the negativity of a piece of text by counting terms with
a negative connotation.
Topic modelling: a method of unsupervised learning lets the data define key themes in text.
Unsupervised learning: a subset of machine learning in which the data provided to the
algorithm does not contain labels.
Zettabyte: a unit of data storage, equal to one sextillion (1021) bytes, one trillion gigabytes, or
one billion terabytes.

4
Annex A: Legal issues around AI and machine learning

AI and machine learning present a range of legal issues relating to privacy and data
protection, consumer protection, anti-discrimination and liability issues, and cross-border
issues.
The issues around data privacy relate to the ability to access the data being processed by AI
and machine learning tools. While big data are widely used to generate profits, they can only
do so with technology that converts the data into relevant services. 1 In financial services, AI
and machine learning applications usually depend on access to, and use of, large amounts of
data in a ‘life cycle’ that includes data collection, data compilation and consolidation, data
mining and analytics.2 The applicability of laws and regulations will be generally fact-driven
and jurisdiction-specific. But certain legal issues are being commonly evaluated in the context
of the use of AI and machine learning with big data, including the applicability of data
ownership rights and data privacy protections and cross-border flows of data. Regulatory
authorities are accelerating efforts to understand the implications for the financial sector. 3
Data ownership rights and protections are being revised in several jurisdictions. Among
OECD members, many have privacy protection laws, and the OECD has guidelines on the
protection of privacy and cross-border uses.4 The European Union (EU) recently enacted a
General Data Protection Regulation (GDPR), due to come into force in 2018. Especially
relevant with respect to the use of AI and machine learning are Article 11, which provides a
right to “an explanation of the decision reached after [algorithmic] assessment,” and allied
articles providing for similar disclosures.5 Other key articles relating to AI and machine
learning are Article 9, which prohibits the processing of “special [sensitive] categories of
personal data” as defined; Article 22, which provides for a data subject’s qualified right not to
be subject to a decision with legal or significant consequences based solely on automated
processing; and Article 24, which provides that decisions shall not be based on special
categories of personal data.6
Ownership of intellectual property in personal data and data protections may be of particular
relevance when a data subject wishes to move their profile to a different data processor, or
when a regulator steps in to transfer a service to another provider. In both cases, having
intellectual property arrangements that conform with regulatory and client expectations may
be challenging given the complexity of the subject area, including cross-border issues (see
below).
Meanwhile, some authorities are exploring if consumers should have the ability to understand
complex modelling techniques for credit systems.7 New tools developed to improve
interpretability of AI and machine learning models can aid firms and policymakers.8
Consumer protection, anti-discrimination and liability issues are areas of emerging focus and
future work, including on the use of big data analytics in risk management and
macroprudential surveillance. A variety of consumer protection laws already apply to big data
practices that might include the use of AI and machine learning, such as fair credit reporting,
equal opportunity laws, and fair trade practices. Yet not all of these regulations may address
AI and machine learning techniques using large quantities of data in digital form.
Anti-discrimination laws in particular may be relevant to AI and machine learning techniques.
Machine learning models may result in discriminatory practices that may be unlawful, even

4
where characteristics such as racial or gender information is not input directly; the
development of non-discriminatory data mining techniques is an active but unsolved area of
research.9

4
Regarding legal liability, there may be questions on the allocation of responsibility among
suppliers, operators and users of AI and machine learning systems – for example the
responsibility of a manufacturer or distributor of a financial product that is based on third
party data input devices or algorithms. 10 There are difficult liability issues, including the
extent to which humans may be entitled to rely on expert systems in a wide range of settings.
Such liability issues will become increasingly important as artificial agents perform a broader
range of tasks currently performed by humans, with the potential for mistakes and for legal
disputes around damages.11
Finally, the growth of AI and machine learning applications could lead to cross-border issues.
Currently, the development of these technologies in finance is concentrated in a small number
of countries, while adoption may occur at financial institutions around the world. Regulators
should keep in mind that cross-border supervision, cooperation and investigation and other
regulatory issues may be expected to arise with AI and machine learning applications active
across jurisdictions.

1
Nobuchika Mori (2017), “Will FinTech create shared values?” speech at Columbia Business School conference, May.
2
Defined in EIOPA (2017), “Opinion of the Occupational Pensions Stakeholder Group on JC Big Data,” EIOPA-OPSG-
17- 06 15, March, pp. 6-7. See also U.S. Federal Trade Commission (2016), “Big Data: A Tool for Inclusion or
Exclusion,” January, p. 3.
3
See EIOPA (2017); U.S. Federal Register (2017), Vol. 82, No.33, and Bureau of Consumer Financial Protection: Docket
No. CFPB Notice and Request for Information Regarding Use of Alternative Data and Modelling Techniques in the
Credit Process, February 21, 2017 (“CFPB RFI”); European Banking Authority (2017), “Report on innovative uses of
consumer data by financial institutions, June. See also FSB FinTech Issues Group (2017), p. 19.
4
OECD (2013), “Guidelines on the Protection of Privacy and Transborder Flows of Personal Data,” July.
5
For instance, Articles 13, 14, and 15 require disclosure of the existence of automated decision-making, including
profiling, referred to in Article 22(1) and (4) and, at least in those cases, “meaningful information about the logic
involved,” as well as the significance and the envisaged consequences of such processing for the data subject.
6
Note that Articles 9, 22 and 24 are all subject to exceptions. See Sandra Wachter, Brent Mittelstadt, and Luciano Floridi
(2017), “Why a Right to Explanation of Automated Decision-Making Does Not Exist in the General Data Protection
Regulation,” International Data Privacy Law, Forthcoming; and Bryce Goodman and Seth Flaxman (2016), “European
Union regulations on algorithmic decision-making and a ‘right to explanation,’” paper presented at 2016 ICML
Workshop on Human Interpretability in Machine Learning (WHI 2016), New York. Wachter et al. argue that these
provisions confer no right to an ex-post explanation of decisions, though ex-post explanations may be crafted through
jurisprudence or EDPB work. Goodman and Flaxman on the other hand argue the law will also effectively create a “right
to explanation,” whereby a user can ask for an explanation of an algorithmic decision that was made about them.
7
Michael Gordon and Vaughn Stewart (2017), “Insights on Alternative Data use on Credit Scoring,” CPFB Law360, May.
8
See Pang Wei Koh and Percy Liang (2017), “Understanding Black-box Predictions via Influence Functions,” Proceedings
of the 34th International Conference on Machine Learning, Sydney; Marco Tulio Ribiero, Sameer Singh and Carlos
Guestrin (2016), “’Why Should I Trust You?’ Explaining the Predictions of Any Classifier,” arXiv:1602.04938v3; and
Fast Forward Labs (2017), “New Research on Interpretability,” August.
9
See Bettina Berendt and Sören Preibusch (2014), “Better decision support through exploratory discrimination-aware data
mining: foundations and empirical evidence,” Artificial Intelligence and Law 22 (2): 175-209; Indrė Žliobaitė (2017),
“Measuring discrimination in algorithmic decision making,” Data Mining and Knowledge Discovery 31(4): 1060–1089;
and Bruno Lepri, Jacopo Staiano, David Sangokoya, Emmanuel Letouze and Nuria Oliver (2016), “The Tyranny of Data?
The Bright and Dark Sides of Data-Driven Decision-Making for Social Good,” working paper, December.
10
See EIOPA (2017), pp. 6-7.
11
Laurence White and Samir Chopra (2011), A Legal Theory for Autonomous Artificial Agents, University of Michigan
Press, chapter 4.

5
Annex B: AI ethics

The rapid development of AI has inspired both hope and concern about rapidly growing
sophistication and capabilities. These discussions extend far beyond AI in financial services
and have inspired new research in philosophy on ‘machine ethics,’ which concerns itself with
ethical norms in the behaviour of artificial agents. 1 As artificial agents take on responsibilities
in areas, such as executing financial transactions, driving cars, and controlling complex
systems of devices, there are concerns about how to ensure ethical behaviour in the interests
of users. Recently, Bostrom (2014) has discussed the implications of ‘superintelligence,’ or
AI systems whose capabilities surpass those of humans, potentially resulting in new
challenges and unintended consequences.2 In 2015, a diverse group of technology and science
luminaries including Stephen Hawking, Bill Gates, and Elon Musk signed an open letter
urging more research into the benefits and risks of AI and another on the dangers of
autonomous weapons.3
One issue is that AI and machine learning may reinforce biases. Some commentators point to
the potential of big data analytics to entrench existing biases in college applications, job
selection, prison sentencing, and credit provision.4 For example, unsupervised learning
algorithms may show fewer high-level job vacancies to female applicants, and sentencing
algorithms may be harsher for ethnic minorities. Yet depending on programming choices,
they may also present opportunities to reduce such discrimination. 5 Other writers note that the
use of machine learning to filter news and social contacts or programming choices which
disadvantage some languages may have far-reaching long-run consequences.6
There are a number of initiatives to further research on AI ethics and the ethical use of AI. For
example, in September 2016, a group of large tech firms (Amazon, Facebook, Google, IBM
and Microsoft) founded the Partnership on Artificial Intelligence, which aims to “develop and
share best practices on AI,” “advance public understanding” and “identify and foster… AI for
socially beneficial purposes.” Apple, eBay, SalesForce, Sony, the ACLU, Human Rights
Watch, UNICEF, and a number of other organisations have since joined the partnership, and it
has launched research around thematic pillars including “AI, labour and the economy” and
“safety-critical AI”.7 Concurrently, the European parliament has debated the legal status of
autonomous agents.8 Finally, machine ethics remains a field of active research, which may yet
yield further insights for broader public policy.

1
Michael Anderson and Susan Leigh Anderson (eds., 2011), Machine Ethics, Cambridge University Press;
Wendell Wallach and Colin Allen (2008), Moral Machines: Teaching Robots Right from Wrong. Oxford
University Press.
2
Nick Bostrom (2014), Superintelligence: Paths, Dangers, Strategies, Oxford: Oxford University Press.
3
Future of Life Institute (2015), “Research Priorities for Robust and Beneficial Artificial Intelligence: an Open
Letter,” January; --- (2015), “Autonomous Weapons: an Open Letter from AI & Robotics Researchers,” July.
4
Cathy O’Neil (2016), Weapons of Math Destruction: How Big Data Increases Inequality and Threatens
Democracy, London: Allen Lane.
5
Executive Office of the President (2016), “Big Data: A Report on Algorithmic Systems, Opportunity, and Civil
Rights,” Washington, D.C., May.
6
Robert Munro (2015), “The threat from AI is real, but everyone has it wrong,” Operational Database
Management Systems, August.
7
Partnership on AI (2017), “Partnership on AI Strengthens Its Network of Partners and Announces First
Initiatives,” May.
8
European Parliament Committee on Legal Affairs (2017), “Draft Report with recommendations to the
Commission on Civil Law Rules on Robotics,” January.

5
Annex C: Drafting team members
Members of the FSB Financial Innovation Network (FIN) contributed to the drafting of this
report. In many cases, they were assisted by additional colleagues from respective institutions.

Chairman of the FIN John Schindler

Associate Director, Financial Stability
Board of Governors of the Federal Reserve System

Brazil Alexandre de Souza Moraes

Head of Division, Financial System Monitoring Division
Central Bank of Brazil

Hong Kong Shu-Pui Li

Executive Director, Financial Infrastructure Department
Hong Kong Monetary Authority

Italy Giuseppe Bruno

Economics and Statistics Directorate
Bank of Italy

Juri Marcucci
Economics and Statistics Directorate
Bank of Italy

Mirko Moscatelli
Economics and Statistics Directorate
Bank of Italy

Japan Hiromi Yamaoka

Director General, Payment and Settlement Systems Department
Bank of Japan

Mexico Alan Elizondo

General Director, Financial System Affairs
Bank of Mexico

Netherlands Mirel ter Braak

Senior Policy Advisor
Netherlands Authority for Financial Markets

Singapore Damien Pang

Head, Technology Infrastructure Office
Monetary Authority of Singapore

Li Xuchun
Head, Supervisory Technology Office
Monetary Authority of Singapore

United Kingdom Henrike Mueller

Global Affairs Manager, International Division
Financial Conduct Authority
Makoto Seta
Senior Associate, Markets Policy, Strategy & Competition Division
Financial Conduct Authority
5
 Sophie Thurner
Associate, Markets Policy, Strategy & Competition Division
Financial Conduct Authority

United States Beth Caviness

Officer, Markets Group
Federal Reserve Bank of New York

Jonathan Landesman
Data Scientist, Research and Statistics
Federal Reserve Bank of New York

Elizabeth Jacobs
Advisor, Office of Financial Research
Department of the Treasury

Ayeh Bandeh-Ahmadi
Advisor, Office of Financial Research
Department of the Treasury

Irina Leonova
Senior Policy Analyst, Division of Risk Management Supervision
Federal Deposit Insurance Corporation

ESMA Patrick Armstrong

Senior Advisor
European Securities and Markets Authority

Alexander Harris
Risk Officer
European Securities and Markets Authority

FSB Secretariat Jon Frost

Member of Secretariat
Financial Stability Board

Laurence White
Member of Secretariat
Financial Stability Board

Other contributors to this report included:

Hong Kong Bénédicte N. Nolens

Vice Chair of IOSCO Committee on Emerging Risks (CER)
Senior Director, Risk & Strategy
Securities and Futures Commission

Ron Chiong
Associate Director, Licensing, Intermediaries
Securities and Futures Commission