Model Monitoring: Data Drift and Concept Drift

A model's accuracy is measured at training time, but the world it predicts keeps changing, and a model that was excellent at launch can silently decay into uselessness months later without a single line of code changing. Model monitoring exists to detect this decay, watching the live inputs and predictions for signs that the data or the relationship the model learned has shifted. The two central phenomena are data drift, when the distribution of incoming features moves away from the training distribution, and concept drift, when the relationship between features and the target itself changes. The problem is acute and uniquely ML: ordinary software fails loudly with errors, but a degrading model keeps returning confident predictions that are quietly wrong, so without monitoring you discover the failure through falling business metrics or customer complaints long after the damage began. Monitoring closes this gap by making degradation observable and actionable, turning silent decay into an alert that triggers investigation or retraining before predictions become harmful.