0Real lending data from an open research dataset
An explanation that caught the model
of 1,000 declines that paying down other debts would have reversed, because the model read debt the wrong way
- What was observed
- A credit policy fitted on 2010 and 2011 loans decided 53,320 applications from 2012 and declined 4,625. For 1,000 of the declines, the smallest single change that would have led to approval was sought: a smaller loan, lower card balances, less other debt or more income, each within set limits.
- What a generic approach says
- A list of the model's top reasons, ranked by weight, would show debt as a minor factor and never reveal its direction.
- What the engine read
- Every change found was confirmed by running the decision again: 1,432 of 1,432. Lower card balances would have reversed 654 declines, a smaller loan 442 and more income 336. Paying down other debts reversed none: in the fitted model, a higher debt to income ratio slightly lowered the predicted risk.
- What it meant for the decision
- The fault was found before it reached a customer: a model should never treat more debt as safer. The fix is to constrain the model's direction; the policy's own cap on debt to income still applied. This was a model we fitted for the study, not a lender's.