Most consumer credit models are trained to predict default — not to maximize profit. These two objectives diverge in ways that quietly cost lenders millions. Here’s the diagnostic every model audit should run.

The proxy problem no one talks about

Binary classification has been the standard in consumer credit underwriting since the 1940s. Train a model to separate good loans from bad. Measure it with AUC, KS, Gini. Deploy. The approach works, but it’s optimizing a proxy for the thing lenders actually care about.

The true objective is portfolio profitability. The proxy is default probability. In many portfolios, these objectives diverge materially and the evidence is sitting in your own loan data right now.

Same binary label, very different economics

Consider two borrowers on a $15,000 personal loan at 19% APR over 60 months:

Borrower A is correctly flagged. Borrower B survives the 24-month observation window, gets labeled “good,” and the model trains on that outcome, learning to approve more borrowers like them. The lender loses $7,100 when month 36 arrives. But the model declared victory at month 24.

The observation window problem: The binary model’s loss function covers 24 months. The lender’s loss function covers the full loan term. For 36- and 60-month products, these windows are materially different.

The diagnostic: run this on your portfolio today

You don’t need to rebuild your modeling stack to find out if this is happening. Segment your approved loan book by model score deciles. Then, compute realized net profit per dollar originated (total payments minus net charge-off) for each decile.

If the relationship is monotonically increasing, your binary model is a reasonable proxy. If it’s non-monotonic — if mid-score deciles are outperforming your top-score deciles — you have direct evidence the proxy is failing. The improvement from a profit-based model will be material.

What auditors should flag: Under SR 11-7, if a model’s stated purpose is portfolio profitability but its objective function is binary cross-entropy, that is a model purpose mismatch. Most credit model validations check AUC and calibration — almost none check objective function alignment.

The solution: predict economic outcome directly

Rather than predicting default probability, a net cash flow (NCF) model predicts how much money each borrower will generate over the loan life (principal and interest received, minus net charge-off). This fuses the revenue stream and credit loss into a single continuous dollar value.

The NCF model will have lower AUC on default classification than the binary model. That’s expected and correct. Lower AUC, higher profit — academic research has documented this tradeoff since 2008.

The rank-order differences show up most clearly in specific segments:

• Large-loan subprime: undervalued by binary models; large revenue base before potential default
• Small-loan near-prime: overvalued; limited revenue ceiling regardless of repayment
• Revolving card revolvers: significantly undervalued; years of interest income not captured by default probability alone