Feature drift in factor models
How to read PSI and KS drift statistics on factor exposures once a model is live, and what a stable population stability index actually lets you conclude.
Editorial catalogue
Free articles on evaluating model drift after deployment
Every entry below is a free editorial and informational article. Read them, link to them, and write to us when a question about your own model goes past what an article can cover.
Article areas we maintain
Training-serving skew
The gap between the distribution a model was trained on and the one it meets in production, why it grows, and how to measure it without retraining the model first.
Concept drift versus data drift
Telling apart a model whose inputs have moved from a model whose input-to-output relationship has moved — because the recommended fix is different for each.
Slice-level error decomposition
Breaking a degraded top-line metric into sector, horizon, and volatility-regime slices so an average that looks fine is not hiding a segment that has broken.
Recalibration, retrain, or retire
Deciding drift thresholds ahead of time — the magnitude that warrants a reweight, the one that calls for retraining, and the one that takes the model offline until the cause is clear.
Shadow-mode monitoring
Running a candidate model alongside the live one to compare drift behaviour before any decision to promote, and how long that window should be.
What this catalogue is not
These articles are not signals, not trade ideas, and not personalised advice. They describe how practitioners evaluate drift; they do not tell you what to hold or when to act.