The Role of Big Data, Machine Learning, and AI in Assessing Risks: a Regulatory Perspective
Scott W. Bauguess, Acting Director and Acting Chief Economist, DERA
Champagne Keynote Address:
OpRisk North America 2017, New York, New York
June 21, 2017
Thank you, Alexander [Campbell] for the introduction.
Thanks also to Genevieve Furtado and the other conference organizers for the invitation to speak here today, at the 19th Annual Operational Risk North America Conference. I understand that this is the Champagne Keynote address. Given that title, I feel obligated as an economist to share with you the reported last words of John Maynard Keynes – the father of modern macroeconomics: “I should have drunk more champagne.” I hope my words here today do not inspire a similar sentiment. And finally, I must remind you that the views that I express today are my own and do not necessarily reflect the views of the Commission or its staff.[1]
My remarks this afternoon will center on a technology topic that is encroaching on many aspects of our lives and increasingly so within financial markets: Artificial Intelligence. Perhaps better known by its two-letter acronym “AI,” artificial intelligence has been the fodder of science fiction writing for decades. But the technology underlying AI research has recently found applications in the financial sector – in a movement that falls under the banner of “Fintech.” And the same underlying technology [machine learning and AI] is fueling the spinoff field of “Regtech,” to make compliance and regulatory-related activities easier, faster, and more efficient.
This is the first time that I have addressed the emergence of AI in one of my talks. But I have spoken previously on the two core elements that are allowing the world to wonder about its future: big data and machine learning.[2] Like many of your institutions, the Commission has made recent and rapid advancements with analytic programs that harness the power of big data. They are driving our surveillance programs and allowing innovations in our market risk assessment initiatives. And the thoughts I’m about to share reflect my view on the promises – and also the limitations – of machine learning, big data, and AI in market regulation.
Perhaps a good place to begin is with a brief summary of where we were, at the Commission, 2 years ago. I remember well, because it was then that I was invited to give a talk at Columbia University on the role of machine learning at the SEC. I accepted the invitation with perhaps less forethought than I should have had. I say this because I soon found myself googling the definition of machine learning. And the answers that Google returned—and I say answers in plural, because there seem to be many ways to define it—became the first slide of that presentation.[3]
The Science of Machine Learning and the Rise of Artificial Intelligence
Most definitions of machine learning begin with the premise that machines can somehow learn. And the central tenets of machine learning, and the artificial intelligence it implies, have been around for more than a half a century. Perhaps the best known, early application was in 1959, when Arthur Samuel, an IBM scientist, published a solution to the game of checkers. For the first time, a computer could play checkers against a human and win.[4] This is now also possible with the board game “Go,” which has been around for 2,500 years and is purported to be more complicated and strategic than Chess. Twenty years ago, it was widely believed that a computer could never defeat a human in a game of “Go.” This belief was shattered in 2016, when AlphaGo, a computer program, took down an 18-time world champion in a best-of-seven match.[5] The score: 4 to 1.
Other recent advancements in the area of language translation are equally, if not more, impressive. Today, if the best response to my question on the definition of machine learning is in Japanese, Google can translate the answer to English with an amazing degree of clarity and accuracy. Pull out your smart phone and try it. Translate machine learning into Japanese. Copy and paste the result into your browser search function. Copy and paste the lead paragraph of the first Japanese language result back into Google Translate. The English language translation will blow your mind. What would otherwise take a lifetime of learning to accomplish comes back in just a few seconds.
The underlying science is both remarkable and beyond the scope of this talk.[6] (Not to mention my ability to fully explain it.) But it is not too difficult to understand that the recent advancements in machine learning are shaping how AI is evolving. Early AI attempts used computers to mimic human behavior through rules-based methods, which applied logic-based algorithms that tell a computer to “do this if you observe that.” Today, logic-based machine learning is being replaced with a data-up approach. And by data-up, I mean programming a computer to learn directly from the data it ingests. Using this approach, answers to problems are achieved through recognition of patterns and common associations in the data. And they don’t rely on a programmer to understand why they exist. Inference, a prerequisite to a rule, is not required. Instead, tiny little voting machines, powered by neural networks, survey past quantifiable behaviors and compete on the best possible responses to new situations.
If you want a tangible example of this, think no further than your most recent online shopping experience. Upon the purchase of party hats, your preferred retailer is likely to inform you that other shoppers also purchased birthday candles. Perhaps you need them too? Behind this recommendation is a computer algorithm that analyzes the historical purchasing patterns from you and other shoppers. From this, it then predicts future purchasing-pair decisions. The algorithm doesn’t care why the associations exist. It doesn’t matter if the predictions don’t make intuitive sense. The algorithm just cares about the accuracy of the prediction. And the algorithm is continually updating the predictions as new data arrives and new associations emerge.
This data-driven approach is far easier to apply and is proving in many cases to be more accurate than the previous logic-based approaches to machine learning. But how does it help a market regulator to know that purchasers of protein powder may also need running shoes?
The simple, and perhaps obvious, answer is that regulators can benefit from understanding the likely outcomes of investor behaviors. The harder truth is that applying machine learning methods is not always simple. Outcomes are often unobservable. Fraud, for example, is what social scientists call a latent variable. You don’t see it until it’s found. So, it is more challenging for machine learning algorithms to make accurate predictions of possible fraud than shopping decisions, where retailers have access to full transaction histories—that is, complete outcomes for each action. The same is true for translating languages; there is an extremely large corpus of language-pair translations for an algorithm to study and mimic.
Two years ago, tackling these types of issues at the Commission was still on the horizon. But a lot of progress has been made since then, and machine learning is now integrated into several risk assessment programs—sometimes in ways we didn’t then envision. I’m about to share with you some of these experiences. But let me preview now, that while the human brain will continue to lose ground to machines, I don’t believe it will ever be decommissioned with respect to the regulation of our financial markets.
The Rise of Machine Learning at the Commission
Let me start by giving you some background on staff’s initial foray into the fringes of machine learning, which began shortly after the onset of the financial crisis. That is when we first experimented with simple text analytic methods. This included the use of simple word counts and something called regular expressions, which is a way to machine-identify structured phrases in text-based documents. In one of our first tests, we examined corporate issuer filings to determine whether we could have foreseen some of the risks posed by the rise and use of credit default swaps [CDS] contracts leading up to the financial crisis. We did this by using text analytic methods to machine-measure the frequency with which these contracts were mentioned in filings by corporate issuers. We then examined the trends across time and across corporate issuers to learn whether any signal of impending risk emerged that could have been used as an early warning.
This was a rather crude proof-of-concept. And it didn’t work exactly as intended. But it did demonstrate that text analytic methods could be readily applied to SEC filings. Our analysis showed that the first mention of CDS contracts in a Form 10-K was by three banks in 1998. By 2004, more than 100 corporate issuers had mentioned their use. But the big increase in CDS disclosures came in 2009. This was, of course, after the crisis was in full swing. And identification of those issues by the press wasn’t much earlier. We analyzed headlines, lead paragraphs, and the full text of articles in major news outlets over the years leading up to the financial crisis and found that robust discussions of CDS topics did not occur until 2008. During that year, we found a ten-fold increase in CDS articles relative to the prior year.
Use of Natural Language Processing
Even if the rise in CDS disclosure trends had predated the crisis, we still would have needed to know to look for it. You can’t run an analysis on an emerging risk unless you know that it is emerging. So this limitation provided motivation for the next phase of our natural language processing efforts. This is when we began applying topic modeling methods, such as latent dirichlet allocation to registrant disclosures and other types of text documents. LDA, as the method is also known, measures the probability of words within documents and across documents, in order to define the unique topics that they represent.[7] This is what the data scientist community calls “unsupervised learning.” You don’t have to know anything about the content of the documents. No subject matter expertise is needed. LDA extracts insights from the documents, themselves using the data-up approach to define common themes – these are the topics – and report on where, and to what extent, they appear in each document.
One of our early topic modeling experiments analyzed the information in the tips, complaints, and referrals (also referred to as TCRs) received by the SEC. The goal was to learn whether we could classify themes directly from the data itself and in a way that would enable more efficient triaging of TCRs. In another experiment, DERA – the Division of Economic and Risk Analysis – research staff examined whether machine learning could digitally identify abnormal disclosures by corporate issuers charged with wrongdoing. DERA research staff found that when firms were the subject of financial reporting-related enforcement actions, they made less use of an LDA-identified topic related to performance discussion. This result is consistent with issuers charged with misconduct playing down real risks and concerns in their financial disclosure.[8]
These machine learning methods are now widely applied across the Commission. Topic modeling and other cluster analysis techniques are producing groups of “like” documents and disclosures that identify both common and outlier behaviors among market participants. These analyses can quickly and easily identify latent trends in large amounts of unstructured financial information, some of which may warrant further scrutiny by our enforcement or examination staff.
Moreover, working with our enforcement and examination colleagues, DERA staff is able to leverage knowledge from these collaborations to train the machine learning algorithms. This is referred to as “supervised” machine learning. These algorithms incorporate human direction and judgement to help interpret machine learning outputs. For example, human findings from registrant examinations can be used to “train” an algorithm to understand what pattern, trend, or language in the underlying examination data may indicate possible fraud or misconduct. More broadly, we use unsupervised algorithms to detect patterns and anomalies in the data, using nothing but the data, and then use supervised learning algorithms that allow us to inject our knowledge into the process; that is, supervised learning “maps” the found patterns to specific, user-defined labels. From a fraud detection perspective, these successive algorithms can be applied to new data as it is generated, for example from new SEC filings. When new data arrives, the trained “machine” predicts the current likelihood of possible fraud on the basis of what it learned constituted possible fraud from past data.
An Example of Machine Learning To Detect Potential Investment Adviser Misconduct
Let me give you a concrete example in the context of the investment adviser space. DERA staff currently ingests a large corpus of structured and unstructured data from regulatory filings of investment advisers into a Hadoop computational cluster. This is one of the big data computing environments we use at the Commission, which allows for the distributed processing of very large data files. Then DERA’s modeling staff takes over with a two-stage approach. In the first, they apply unsupervised learning algorithms to identify unique or outlier reporting behaviors. This includes both topic modeling and tonality analysis. Topic modeling lets the data define the themes of each filing. Tonality analysis gauges the negativity of a filing by counting the appearance of certain financial terms that have negative connotations.[9] The output from the 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 adviser.
The results are impressive. Back-testing analyses show that the algorithms are five times better than random at identifying language in investment adviser regulatory filings that could merit a referral to enforcement. But the results can also generate false positives or, more colloquially, false alarms. In particular, identification of a heightened risk of misconduct or SEC rule violation often can be explained by non-nefarious actions and intent. Because we are aware of this possibility, expert staff knows to critically examine and evaluate the output of these models. But given the demonstrated ability of these machine learning algorithms to guide staff to high risk areas, they are becoming an increasingly important factor in the prioritization of examinations. This enables the deployment of limited resources to areas of the market that are most susceptible to possible violative conduct.