In a recent interview, the computer scientist Andrew Ng said, "Just as electricity transformed almost everything 100 years ago, today I actually have a hard time thinking of an industry that I don't think AI [artificial intelligence] will transform in the next several years." Whether AI effects such widespread change so soon remains to be seen, but the financial services industry is clearly in the early stages of being transformed—with implications not only for market participants but also for financial supervision.
Some of the implications of this transformation were discussed in a panel at a recent workshop titled "Financial Regulation: Fit for the Future?" The event was hosted by the Atlanta Fed and cosponsored by the Center for the Economic Analysis of Risk at Georgia State University (you can see more on the workshop here and here). The presentations included an overview of some of AI's implications for financial supervision and regulation, a discussion of some AI-related issues from a supervisory perspective, and some discussion of the application of AI to loan evaluation.
As a part of the panel titled "Financial Regulation: Fit for New Technologies?," I gave a presentation based on a paper I wrote that explains AI and discusses some of its implications for bank supervision and regulation. In the paper, I point out that AI is capable of very good pattern recognition—one of its major strengths. The ability to recognize patterns has a variety of applications including credit risk measurement, fraud detection, investment decisions and order execution, and regulatory compliance.
Conversely, I observed that machine learning (ML), the more popular part of AI, has some important weaknesses. In particular, ML can be considered a form of statistics and thus suffers from the same limitations as statistics. For example, ML can provide information only about phenomena already present in the data. Another limitation is that although machine learning can identify correlations in the data, it cannot prove the existence of causality.
This combination of strengths and weaknesses implies that ML might provide new insights about the working of the financial system to supervisors, who can use other information to evaluate these insights. However, ML's inability to attribute causality suggests that machine learning cannot be naively applied to the writing of binding regulations.
John O'Keefe from the Federal Deposit Insurance Corporation (FDIC) focused on some particular challenges and opportunities raised by AI for banking supervision. Among the challenges O'Keefe discussed is how supervisors should give guidance on and evaluate the application of ML models by banks, given the speed of developments in this area.
On the other hand, O'Keefe observed that ML could assist supervisors in performing certain tasks, such as off-site identification of insider abuse and bank fraud, a topic he explores in a paper with Chiwon Yom, also at the FDIC. The paper explores two ML techniques: neural networks and Benford's Digit Analysis. The premise underlying Benford's Digit Analysis is that the digits resulting from a nonrandom number selection may differ significantly from expected frequency distributions. Thus, if a bank is committing fraud, the accounting numbers it reports may deviate significantly from what would otherwise be expected. Their preliminary analysis found that Benford's Digit Analysis could help bank supervisors identify fraudulent banks.
Financial firms have been increasingly employing ML in their business areas, including consumer lending, according to the third participant in the panel, Julapa Jagtiani from the Philadelphia Fed. One consequence of this use of ML is that it has allowed both traditional banks and nonbank fintech firms to become important providers of loans to both consumers and small businesses in markets in which they do not have a physical presence.
Potentially, ML also more effectively measures a borrower's credit risk than a consumer credit rating (such as a FICO score) alone allows. In a paper with Catharine Lemieux from the Chicago Fed, Jagtiani explores the credit ratings produced by the Lending Club, an online lender that that has become the largest lender for personal unsecured installment loans in the United States. They find that the correlation between FICO scores and Lending Club rating grades has steadily declined from around 80 percent in 2007 to a little over 35 percent in 2015.
It appears that the Lending Club is increasingly taking advantage of alternative data sources and ML algorithms to evaluate credit risk. As a result, the Lending Club can more accurately price a loan's risk than a simple FICO score-based model would allow. Taken together, the presenters made clear that AI is likely to also transform many aspects of the financial sector.