Paused
Calibration in Credit Risk Scoring
A loan default model served behind an API, built to ask whether the score is honest about its own uncertainty rather than whether it is accurate.
- Status
- Paused
- Last updated
- February 2026
01
So far
A well ranked model can be badly calibrated, and for a lending decision calibration is what matters, because the probability is the thing being acted on. Isotonic calibration on a held out split closed most of the gap. Paused rather than finished: the remaining work is measuring calibration drift over time, which needs a longer window of production data than I have.
02
Stack
- scikit-learn
- Jupyter
- FastAPI