Concept drift in credit and fraud models
Distribution shift, concept drift and non-stationary time series appear on 371 recent papers, almost all in AI journals, and only 27 of them concern credit, fraud, banking or markets. Credit and fraud models decay as borrowers, fraudsters and the economy change. How fast, how should it be detected, and when should a model be retrained rather than recalibrated?
What prompted this
Since October 2025, 371 papers carry a concept-drift, non-stationarity, distribution-shift, covariate-shift or dataset-shift keyword (347 in AI journals): "distribution shift" on 170, "concept drift" on 65, "non-stationary time series" on 42, "non-stationarity" on 22. Only 27 of the 371 also carry a credit, finance, fraud, bank, loan, stock, market, economics or trading keyword.
Questions
- In credit scoring models deployed through the pandemic and the rate rises, how much did discrimination and calibration decay per year, and did the decay come from borrower mix or from the relationship between features and default?
- Which drift detectors raise the alarm early enough to matter and rarely enough to be trusted?
- When fraudsters adapt to a model, is the drift adversarial in a way that standard detectors miss?
Data that could help
Loan performance data over 2018–2025 (LendingClub, mortgage loan-level data), fraud transaction data with time stamps, and model monitoring logs from a lender.
Next step
Train a credit model on 2018–2019 loans and track its monthly discrimination and calibration through 2025, attributing the decay to feature drift and relationship drift.