What Is Model Drift? Causes, Detection, and How to Fix It

This post details model drift, distinguishing between data drift, concept drift, prediction drift, training-serving skew, and upstream drift. It emphasizes that most drift issues stem from upstream data quality or pipeline changes rather than the model itself. Practical detection methods are described, including storing training baselines, using Population Stability Index (PSI) for numerical and categorical features, and setting alert thresholds based on business impact and use-case risk. It also highlights limitations of input monitoring and suggests user query patterns as an early signal. The post provides a table mapping symptoms to root causes and first responses, stressing the importance of fixing data issues before retraining and discussing when retraining is appropriate.