The traditional FICO score is a 30-year-old algorithm built on a narrow set of credit-bureau variables. Payday and short-term lenders rejected its limits long ago and built their own machine learning systems — often gradient-boosted decision trees or neural nets — that ingest hundreds of features and produce an approval decision in seconds. The models are powerful, the variables are sometimes surprising, and the accuracy claims deserve scrutiny.
The variables that actually move the needle
Income and employment matter, but they aren’t usually the top features. In published model papers and patent filings from companies like LendUp, ZestFinance, and Elevate, the most predictive variables tend to be behavioral and transactional: how often a borrower’s checking account balance dips below $50, the number of NSF (non-sufficient funds) fees in the last 90 days, the recency of the last cash advance, and the time of day the loan application is submitted. Late-night applications correlate with higher default rates. So does using a mobile device with a low battery — a proxy, researchers think, for general life disorganization. Bank-transaction data, accessed via Plaid or similar APIs, has become the dominant data source, often outweighing credit-bureau information.
The accuracy debate
Lenders advertise AUC (area under the curve) scores of 0.75 to 0.85 for their default models — meaningfully better than the 0.65 to 0.70 typical of FICO on this population. Independent academic replications, when researchers can access the data, generally confirm the lift, but with two important caveats. First, the lift is concentrated in identifying obvious bad bets; the marginal cases — borrowers who might or might not default — are still hard to predict regardless of model sophistication. Second, the models are optimized for lender profitability, not borrower welfare. A model that approves a loan it expects to roll over five times before default is profitable for the lender and ruinous for the borrower, and it scores well by the lender’s metric.
Fairness and proxy variables
The Consumer Financial Protection Bureau and several state regulators have raised concerns that machine learning models, even when explicitly excluding race, can use proxies that produce disparate impact. ZIP codes, smartphone OS, and shopping patterns can all correlate with protected class membership. The models don’t “know” race, but they predict outcomes that track racial lines because the underlying training data reflects historical inequities in income and access to credit. Vendors push back that their models are more inclusive than FICO — and on raw approval rates for thin-file borrowers, that’s often true — but the inclusion comes with high APRs that critics argue capture rather than alleviate the underlying disadvantage.
The bottom line
Payday lending’s machine learning models are genuinely more accurate than legacy credit scoring on the populations they serve. They’re also tuned to a business model — short-term, high-fee credit with frequent rollovers — that many economists argue is structurally harmful. Improving the math doesn’t fix the product. Until regulation catches up to the algorithmic frontier, the most important question isn’t whether the model can predict default, but whether the loan it approves should exist in the first place.
Leave a Reply