A month ago I was among the 5,000 attendees at the LendIt Fintech trade show in San Francisco. LendIt is one of the larger fintech events around, making it a good time to take the industry’s pulse. Throughout the two days, I heard a lot about fintech’s future: the adoption of blockchain, improving customer experience, using artificial intelligence (AI) to power all sorts of lending activities. But the biggest theme reverberating across the Moscone Center’s busy aisles was data. Consumers are generating rising volumes of it while companies are collecting more of it to find their business edge.
One important trend is the rise of “consumer-permissioned” credit scores that encourage consumers to give credit bureaus a peek at their checking and savings accounts, or link up their cell phone and utility payment history on the assumption that more data will help them improve their credit scores. This isn't true for everyone: For some people who may have had a temporary slip of their financial footing, scores could go down.
There’s no question that more data is better. All the work we've done at Zest over a decade has proven that using more data can improve credit access for the tens of millions with limited or no credit histories or whose credit scores sit under 680. A lack of information handcuffs them in the loan process. But widespread use of techniques to lift people’s credit scores comes with plenty of caveats, especially this late in the economic growth cycle. Market watchers are reminding banks and lenders to look past traditional measures that always rise along with the economic tide.
The bigger issue is that these consumer opt-in data services are a half-measure. That new data still most likely feeds into a traditional underwriting model, probably a logistic regression, which maxes out at 30 or 40 variables. Lenders can make far more accurate predictions by loading all of that new data (along with the mountain of data they already have) into a machine learning model that uses hundreds of variables to make trillions of calculations to produce a better decision. Anything else is tinkering around the edges.
My other big takeaway from LendIt is a growing concern around the current economic slowdown. It’s a reasonable worry given the consensus that loan activity has slowed in the past year and the global business environment seems to be in a synchronized deceleration. In credit, the riskiest loans tend to get booked in the period we are in right now, because most people improve their risk profiles in strong economic times, plus customers tend to want loans for increasingly speculative reasons. Making the situation more uncertain is that it’s been a decade since the industry has had to navigate a recession. Quite simply, many models on both bank and fintech sides haven’t been stress tested. Are we making loans to the people to whom we want to make loans? In the past, some underwriters have tended to use fewer data points as the economic booms extended themselves. That’s where smarter and greater use of data can mitigate harm. No model or modeling technique is recession-proof, but ML models can at least act as early-warning systems for economic swings. And that’s an important start.
