As machine learning continues to make deeper inroads in banking and financial markets, a debate has broken out over its potential impact on the core of the business: lending and credit underwriting. Can machine learning really help tech-savvy lenders understand their borrowers better than they already do?
The skeptics’ side of the argument rests on a seemingly common-sense theory that was expressed in a major report released in July by BaFin, the main German financial regulatory agency. (Officially known as the Federal Financial Supervisory Authority, BaFin combines oversight functions of the S.E.C., Federal Reserve, and Comptroller of the Currency.)
The report, titled “Big data meets artificial intelligence,” was an effort to anticipate which parts of the global economy machine learning will change most dramatically. Overall, the report’s authors foresee vast potential for the new technology to revolutionize many financial markets, but one place where they surprisingly saw little potential for change was in lending.
The authors, who include both regulators and outside consultants, caution that lenders might be able to wring minor improvements out of their existing risk assessment models, but any benefit is likely to be limited because banks are already such experts on assessing credit risks: “Extensive analytical and data-based optimization measures have been introduced in exactly this area over the past few decades which means that the potential for improvement might be more limited here than in the case of process automation and optimization.”
While that logic may make intuitive sense to some in the financial industry who feel they are already making the best decisions possible when it comes to lending, innovators such as ZestFinance are proving that there are all sorts of insights still left for lenders to discover about their would-be borrowers by using the right kinds of machine learning models.
Clients who use our ZAML software tools to build more predictive machine learning credit models have been able to decrease charge-offs by an average of 30%. Those who target an improvement in approval rates (keeping risk steady) have typically seen increases of 15% by using machine learning in their lending decisions.
The key is spotting borrowers with the ability to repay the loan but without the traditional credit history to qualify. That was the case with Prestige Financial Services, a lender that specializes in financing car buyers with relatively low credit scores. By using ZAML, Prestige was able to identify applicants who lacked a traditional credit history but deserved credit just the same.
The conventional wisdom has it that banks and credit agencies have voluminous credit data on nearly everyone. That’s not quite right: roughly 46 million Americans lack a standard, adequate credit file. And, while credit scores are still considered the gold standard for lending decisions, these metrics may not be telling us the same thing that they have historically because people are so focused on manipulating their score to get them up. One economist told Bloomberg that these score-fluffing strategies are “mucking with their relationship to the underlying credit risk.”
The work we’re doing with a growing range of clients is changing the conversation about the primacy of any one number on someone’s creditworthiness and the impact of machine learning on credit decisions. The models we work on with banks and lenders often increase the variables used by 10x or 100x, producing more nuance and signal that help banks make large improvements in reducing charge-offs and increasing their lending base.
Just because humans have spent decades analyzing myriad aspects of credit risk is no reason to assume there is nothing left to discover. Time and again, inventions such as the microscope, radar, and ultrasound imaging have allowed us to peel away assumed limits to our perception.
Humans are quite good at developing strategies and rules for selecting the right data for a model, and telling important stories based on the results. But it may be wiser — and humbler — to acknowledge that people won’t always make the best possible decision on their own given the massive number of insights we’re overlooking.
A recent article in the Atlantic summarized this new, more optimistic hypothesis this way: “The connections that feel uncanny are actually the outliers, because they are the ones we [humans] notice. What about all the rest that go unobserved, linking behaviors in ways individuals haven’t even thought to think? Those are the links that machine learning promises to bind.”