What is Model Drift? - Definition & Meaning

Definition

Model drift (or ML drift) is the gradual or sudden decline of model performance in production because input data or the relationship between input and output changes. The model trained on historical data no longer matches current reality.

Technical explanation

Two main types: data drift (input distribution changes) and concept drift (relationship between input and target changes). Detection methods: statistical tests (KS, PSI), feature distribution monitoring, performance metrics (accuracy, AUC) over time, and reference data comparison. Solutions: periodic retraining, online learning, and human-in-the-loop review. Drift can be gradual (seasonal patterns) or abrupt (COVID, policy change). Monitoring and automated retraining are part of MLOps.

How AVARC Solutions applies this

AVARC Solutions implements drift monitoring for production models. We track feature distributions, model performance, and business metrics, and trigger retraining or human review when significant drift is detected.

Practical examples

• A fraud model where transaction patterns change after a new payment option, reducing model effectiveness.
• A recommendation system where product catalog and customer behavior change seasonally.
• A churn model where the definition of "churn" was internally changed, making labels inconsistent.

Related terms