Glossary
Concept drift
A change over time in the relationship between a model's inputs and the outcome, so that a model fitted in the past loses accuracy.
Concept drift occurs when the link between features and outcome changes. Related data drift is a change in the distribution of the inputs alone. Both are common in markets, where behaviour, costs and regimes change.
Drift is tracked by comparing recent input distributions with the training period (for example with the population stability index) and by watching live accuracy and calibration.
Example: a feature whose five equal-sized bins now hold 10%, 15%, 20%, 25% and 30% of observations has a population stability index of about 0.135 against the original 20% each.