SECAI Core · phase 15 of 15
Model drift
A phenomenon wherein the statistical properties of the target variable or features of a machine learning model change over time, leading to a reduction in model performance.
The Explain card
- Plain English
- Model drift is the decline in a model's performance over time because the statistical relationships it learned, between its features and its target, no longer hold in the real world.
- Example
- A fraud model learned that late-night foreign transactions signal fraud. After the business expands internationally, that pattern becomes normal behaviour, and the model starts flagging legitimate customers while missing newer fraud types.
- Why it matters
- A drifting model gives false confidence. Defenders need scheduled performance checks and retraining triggers, because a model that was accurate at launch can become a liability.
- Hook
- The model did not change, the world did.
Where it sits in the deck
Phase 15: MLOps, Continuous Delivery, and Operational Resilience
Close the loop — operational practices, drift management, and delivery discipline that keep deployed AI systems healthy, current, and continuously improving.
