Model Drift

Model drift is the degradation of a deployed machine learning model's predictive accuracy over time, caused by changes in the statistical properties of real-world data relative to the data the model was trained on. Two primary forms are data drift (input distributions shift) and concept drift (the relationship between inputs and outputs changes, such as when customer preferences evolve or market dynamics shift).

In commerce AI systems, model drift is an operational reality that requires active monitoring and response plans. A demand forecasting model trained before a major market shift, or a recommendation model that has not seen new product categories, will produce increasingly unreliable outputs even if it scored well at deployment. Left undetected, model drift causes silent failures — business decisions based on stale predictions, revenue loss from poor recommendations, or degraded customer experience — making drift detection a core component of any responsible AI operations program.