The loss of ML model performance over time is known as model drift. This means that the model begins to generate predictions with reduced accuracy over time.

Monitoring for drift is an essential part of ML observability, which is the practice of monitoring,

troubleshooting, and explaining an ML model throughout its lifecycle. Monitoring helps teams quickly identify issues during production that have a detrimental impact on your model’s performance, especially if the model has either a delayed or possibly no ground truth (i.e. the target for training/validating, which is the reality you want your ML model to achieve).