Your profit model isn’t broken. It just can’t see the risk that will eventually break you.
Most credit models are answering the wrong question.
Binary classification, whether this borrower defaults or not, has been the standard framework in consumer lending for decades. The natural upgrade is a profit model: predict expected net cash flow instead of default probability, optimize for revenue rather than risk avoidance, and approve people who have high positive expected value.
It’s the right direction. Yet it’s incomplete in a way that can destroy a portfolio.
The Problem With Expected Value
A profit model predicts the average outcome. But averages hide everything that matters:
- the width of the distribution
- the depth of the downside
- the borrowers who look fine on paper until they don’t
In credit, the tail is what kills you, and the mean never shows you the tail.
Consider two borrowers with identical expected profit. One has a tight distribution where outcomes cluster around the mean. The other has a wide distribution with strong upside, but a deep left tail. In benign conditions, the profit model treats them identically. In a stress scenario (like a recession, an unemployment spike, or a rate shock), the wide-distribution borrower produces catastrophic losses the model never anticipated.
The profit model built you a portfolio that looked great for years. Then a single stress quarter took back a significant share of it.
The model wasn’t wrong. It just never told you what it didn’t know.
The Revolver Problem
For credit card portfolios, this plays out in a specific and well-documented pattern.
A profit model correctly identifies revolvers, borrowers who carry balances and pay interest, as high-value. Their expected profit is strong. The model approves them aggressively.
What the model misses: revolvers are disproportionately exposed in stress scenarios. They’re already carrying debt. A job loss hits them harder. When they default, they default at near-maximum utilization. The loss severity is high precisely because the revenue was high.
The portfolio that outperformed for three years becomes the portfolio with the worst stress performance. The profit model selected for it systematically.
What the Model Doesn’t See
The profit model approved the portfolio that looks best on average. It had no way to see the portfolio that could destroy you in a bad year.
That gap — between expected value and tail risk, between mean performance and stress performance — is where credit portfolios quietly accumulate danger. It doesn’t show up in your monthly performance reports. It doesn’t trigger your model monitoring alerts. It builds in the shape of the distribution, invisible until the environment changes and the tail becomes real.
The question isn’t whether your profit model is well-built. It probably is. The question is what it’s not designed to see and whether anyone has looked there.
