Notes from the Vault
Larry D. Wall
May 2018

Will machine learning and artificial intelligence have a transformative impact on the financial system, the economy, and even society? If so, are our regulatory and monetary policy frameworks ready for this brave new world? The Atlanta Fed's 2018 Financial Markets Conference, Machines Learning Finance. Will They Change the Game? addressed these and related topics associated with machine learning (ML) and artificial intelligence (AI). This Notes from the Vault reviews some of the highlights from the conference related to ML/AI and their use in the financial system. The conference also had a panel discussing potential implications of AI/ML for monetary policy, which I review in a companion macroblog post.

How do machines learn finance?
The concept of artificially intelligent robots has existed and been a part of our culture throughout the lifetime of most people living today. As a result, for those not working in AI/ML there can easily be some confusion between what is currently possible and what is still science fiction.1 The first conference panel, "How Do Machines Learn Finance?" sought to clarify how ML works, describe what it can and cannot do, and give examples of how it is being currently applied.

Princeton professor Ryan P. Adams began the discussion with a presentation titled "What Is Machine Learning." Adams observed there are three types of machine learning: supervised learning used for prediction, unsupervised learning used to find structure in data, and reinforcement learning used for decision making. With supervised learning, each observation has a label or value, and the machine uses input data to build a function that approximates that value. Supervised machine learning is currently being applied to identifying cancer cells and making movie recommendations, among other uses. Adams noted that one popular form of supervised learning is deep learning, which he said is "just adaptive basis function regression." In contrast to supervised learning, in unsupervised learning the input data do not have a known output label or value, but rather the machine seeks to identify the structure in the data. Adams gave several examples where unsupervised learning has been used, including finding communities in social networks and modeling sports data.

Reinforcement learning attracted considerable attention in recent years with AlphaGo's success in beating champion Go players. Given the complexity of Go, the success of reinforcement learning could be taken to indicate machines have become as or more intelligent than humans. However, Adams noted that training machines to play games like Go required substantial bespoke efforts by humans, and efforts to get machines to transfer learning across different games (such as old Atari games) required the machine to gain a "huge amount of experience." Reinforcement learning could also be applied to robotics to get machines to figure out what they need to do, but Adams noted that for low-level problems, reinforcement learning is not currently competitive with traditional control approaches. On the bigger question of whether artificial intelligence is on the verge of overtaking human intelligence (in other words, strong AI), Adams said it is highly unlikely without many technical advances.

The use of "AI in the Enterprise" was the topic of a presentation by Columbia University professor John P. Cunningham. He observed firms could apply AI in four ways to add value to their business: strategic intelligence, capital efficiencies, risk mitigation, and optimization of value-centric processes. However, he said that the technical resources in AI are being consolidated in the "AI majors," a handful of large U.S. and Chinese technology firms. Given the high cost of breaking into this group, he suggests other businesses take a different approach, centered on finding a value-centric opportunity that would allow them to leverage their large proprietary data sets and to upskill the firms' existing software engineering resources.

Gideon Mann, the head of data science at Bloomberg, talked about natural language processing in his presentation, "The New Alchemy: Turning Words into Signals." Mann started by explaining how documents being written for humans are converted into data used by machines. Bloomberg uses this technology in a supervised learning setup to train machines to read news stories and determine whether they have positive, neutral, or negative news about a firm's stock. The machine can then provide traders in financial markets with a news feed that identifies stories that have the potential to move market prices and/or volumes.

Machines learning regulation (and vice versa)
Machine learning and artificial intelligence raise both regulatory issues specific to finance and broader social, legal, and ethical issues. The financial regulatory issues were raised both in comments by Federal Reserve Board vice chairman for Supervision Randal Quarles and in a panel on "Machines Learning Regulation (and Vice Versa)."

Quarles discussed developments in machine learning and some of their implications for financial regulation in an armchair chat with Federal Reserve Bank of Atlanta president Raphael Bostic. As a part of that discussion, Quarles observed that some concerns, such as interacting with clients, are already covered in the existing regulatory framework and that the Federal Reserve would "like to be certain that the traditional protections have been complied with." However, in terms of developing new regulations, at this stage he is reluctant to impose "bounds" on what banks can do. Quarles would prefer firms establish circuit breakers that could stop activities that produce bad outcomes beyond some milestones.

The panel looked at the use of machine learning from a variety of perspectives. Scott Bauguess, a deputy director at the Securities and Exchange Commission (SEC), talked about how his agency is using ML to help decide which financial advisers should be subject to a review and to identify potential fraud. One important ML issue raised by Bauguess is that of the data used to train the models. He observed the SEC often needed to engage in considerable efforts to clean the data so it would be usable by ML algorithms. Bauguess also pointed out that although the SEC has a large amount of cross-sectional data, at most it has about 10 years of time-series data, but for some purposes, it would prefer to have 70 to 100 years of data.

Both the potential benefits and risks associated with the application of ML in finance were highlighted by two examples given by John Schindler, an associate director at the Federal Reserve Board. As an example of the benefits, Schindler described the use of ML to price insurance on the delivery of consumer goods ordered electronically. The prices set for this insurance are based on a variety of factors, such as the products being ordered, and happens almost instantly as the buyer goes to the transactions page. The pricing of such insurance in real time would be almost impossible to provide without ML. As an example of the risks, he discussed a trader who was using ML to make what had so far proved to be profitable investment decisions. The problem was the trader did not understand how the model worked and hence could not adequately evaluate the risks being taken.

The potential for ML to help financial services firms comply with regulation was the subject of the presentation "Cognitive RegTech Solutions" by William Lang, a managing director at the Promontory Financial Group. Lang observed that cognitive systems can augment human expertise and existing risk analytics to provide enhanced decision making. He also noted ML has the capacity in some areas to turn what had been soft information that is not readily transmitted, such as the opinions of informed analysts, into hard information that can be more readily used by decision makers higher up in the organization. However, echoing Bauguess, Lang said that doing so requires lots of good data. Indeed, he noted that even if we had 500 years of data we would still complain, because historical financial data are not always informative about future events.

Machines learning investments
At first glance, the potential for ML to improve investment decisions would seem obvious, given that financial markets produce large volumes of data that could be analyzed. However, the conference's panel on the use of ML in investing highlighted a number of problems. Rishi Narang, the founding principal of T2AM, observed that although ML has proved useful in analyzing unstructured data, such as the transcripts of firm's earnings calls, ML has not proven so helpful in identifying alpha (investments earning positive risk-adjusted returns) or in risk management.2 One of the problems is that in practice there is not so much data. ML techniques typically assume the process generating the data is stable over time. However, Narang discussed that this stability does not exist with financial data, so data drawn from one period may not be very predictive of future performance.3 This problem is compounded by the way orders interact with one another. A trading strategy may be highly profitable when it is first implemented, but those profits decline over time as other traders start implementing the same strategy.

One area where algorithms dominate is in the area of high-frequency trading, in which decisions need to be made within a fraction of a second, far faster than humans could respond. However, Christina Qi, a partner at Domeyard, observed that high-frequency trading relies almost exclusively on algorithms programmed by humans and not on investment decisions by ML algorithms. The problem with ML is it is too complex and slow. Even if ML is better at identifying profitable high-frequency trades, by the time the ML algorithm identifies the opportunity the market has moved on and the trade may no longer be profitable. Also in terms of speed, Qi made an important point about how markets evolve. She observed that one of the biggest strengths of her firm five years ago was that it was one of the fastest investors. However, other firms have caught up and now speed does not provide a profitable competitive advantage.

Andrei Kirilenko, a professor at Imperial College London, noted that differences in latency—the delay between when data is first available and when different high-frequency trading algorithms respond—has contributed to flash crashes. He also made an interesting observation about the potential for ML to be used by regulators to identify market manipulation. He noted that a group of MIT students applied ML to some financial market data and were able to identify new instances of market manipulation. This example illustrates an important point; just because some illegal market activity cannot be currently detected by regulators does not mean that the perpetrator is safe. Regulators ability to apply ML to identify illegal activity is likely to increase over time, while the incriminating data may remain available for a very long time.

Legal and ethical structures to manage machine learning
The legal and ethical structures needed to allow ML to provide significant gains while reducing the costs were the subject of a research paper presentation and a keynote speech. The research paper, "Regulating Artificial Intelligence Systems: Risks, Challenges, Competencies, and Strategies," was presented by Matt Scherer, an associate at Littler Mendelson. Scherer observed that the largely informal legal structures of the preindustrial society with decentralized production transitioned over time into a much more formalized structure, as society became industrialized with centralized production. These changes were needed to cope with a variety of new risks that arose from industrialization, such as mass-produced defective products and environmental threats. He then observed that the existing forms of regulation designed to deal with the problems of industrialization have some shortcomings when applied to managing the risk associated with artificial intelligence. Scherer concluded by discussing the strengths and weaknesses of relying on legislatures, regulators, and the courts to address the looming problems created by increased reliance on artificial intelligence.

Gregory Scopino, special counsel at the Commodity Futures Trading Commission, discussed the paper from the perspective of a regulatory body that has already had to address some of the challenges created by AI/ML. He gave an example of a software program that told people when to buy and sell futures contracts. The decision of the courts was that if the software performed the functions of a commodity trading adviser, then it should be subject to regulation as such. Scopino observed that to the extent financial markets are increasingly replacing human decision making with AI/ML, then the financial regulators are "already on their way to becoming full-blown AI regulators."

Although the legal structure establishes society's minimum expectations, the standards for ethical behavior are higher. The keynote presentation by John Havens, the executive director of the IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems, discussed the importance of ethics in AI/ML. He began by observing the importance of individuals understanding and living according to their values if they want those values incorporated into the behavior of machines. He then addressed two important ethical issues. First, he argued that individuals should be at the center of control over their personal data rather than technology companies. Second, he noted the potential for AI to result in a large number of job displacements and increased inequality of income distribution. As a consequence, Havens emphasized that our measure of economic performance has to extend beyond just measuring gross domestic product (GDP) and include people's quality of life.

Digitization of money and finance: challenges and opportunities
Along with all of the developments in AI/ML, another potentially important development in financial technology is that of crypto-assets, such as Bitcoin. Tao Zhang, deputy managing director at the International Monetary Fund (IMF), discussed the benefits, risks, appropriate regulatory response, and IMF actions related to such crypto-assets. Zhang observed that such crypto-assets could provide a variety of benefits, including improved payments and a new asset class whose returns have (thus far) been uncorrelated with other asset returns. However, he also noted that such crypto-assets can also be used for criminal activity. Zhang suggested four key points in terms of the regulatory response: (1) "regulators need to complement their focus on entities with increasing attention to activities," (2) "rules and standards will need to be developed to ensure the integrity of data, algorithms, and platforms," (3) "policy options could be considered to support open networks," and (4) "legal principles need to be modernized."4

Conclusion
Advances in technology and the collection of large volumes of data have allowed machine learning to play an increasingly important role in the financial system and the broader economy. The Atlanta Fed's recent Financial Markets Conference highlighted some of the strengths and weaknesses of ML as a tool for use by financial firms and their supervisors. The discussion also addressed some of the legal and ethical challenges posed by AI.

One theme that emerged from the conference is both the value and limitations of a large quantity of time-series and cross-sectional data. Many ML techniques produce better results, given a large quantity of data drawn from a process that is stable over time and across firms. This poses a challenge, however, for many financial applications where not only do we have a limited time series but also where the process changes over time. Such changes in the process can occur due to structural changes in the economy and structure of markets but also as a response to the use of ML by some market participants.

Larry D. Wall is executive director of the Center for Financial Innovation and Stability at the Atlanta Fed. The author thanks Nikolay Gospodinov and Mark Jensen for helpful comments. The view expressed here are the author's and not necessarily those of the Federal Reserve Bank of Atlanta or the Federal Reserve System. If you wish to comment on this post, please email atl.nftv.mailbox@atl.frb.org.

_______________________________________

1 A post by Harvard graduate student Rockwell Anyoha discusses that artificially intelligent robots have been a part of science fiction since the first half of the 20th century.

2 It should be noted some hedge funds and asset managers are using ML to combine recommendations from multiple models to produce higher-risk-adjusted returns (a process Rishi Narang referred to as "alpha mixing"). For example, see Richard Craib's (founder of Nuerai) description of his firm's approach.

3 In more technical terms, financial data are not necessarily stationary (that is, the distribution of these data is not constant but rather is time-varying).

4 Zhang also noted regulators may want to take steps to promote the transparency and integrity of initial coin offerings (ICOs). For more on ICOs, see parts one and two of my two-part series in Notes from the Vault.