Your Model Isn’t Done: Understanding and Fixing Model Drift

Model drift is the gradual or sudden deterioration of a production ML model's performance over time. It falls into two categories: data drift (input feature distributions change) and concept drift (the relationship between features and target changes). Real-world examples include unit changes in measurement systems and demographic shifts in hospital populations. The best defense is regular monitoring of performance metrics (AUC, precision, recall) over time, combined with feature missingness and distribution tracking. Fixes range from correcting data pipelines to restore original formats, to fully retraining the model on new data when the underlying population has shifted.