COMPEL Glossary / concept-drift
Concept Drift
Concept drift occurs when the underlying relationship between input data and the outcome being predicted changes over time.
What this means in practice
Unlike data drift (where inputs change), concept drift means the real-world rules have changed. For example, a customer churn model trained during normal times may fail during a recession because the factors driving churn have fundamentally shifted. Concept drift is particularly challenging because the model may receive inputs that look normal (no data drift detected) while producing increasingly inaccurate predictions because the mapping from inputs to outcomes has changed. Addressing concept drift requires retraining the model on recent data that reflects the new reality. In the COMPEL framework, concept drift is classified as a model risk in the risk taxonomy and monitored through both outcome tracking and prediction distribution analysis.
Why it matters
Concept drift is particularly dangerous because models may receive normal-looking inputs while producing increasingly inaccurate predictions — the real-world rules have changed but the model has not adapted. Unlike data drift where statistical tests can detect input changes, concept drift can be invisible until business outcomes deteriorate significantly. Organizations that do not monitor for concept drift risk operating on stale models that give false confidence.
How COMPEL uses it
Concept drift is classified as a model risk in the COMPEL risk taxonomy, monitored through both outcome tracking and prediction distribution analysis. During Calibrate, concept drift monitoring capability is assessed as a Technology pillar maturity indicator. The Model stage designs drift detection systems with appropriate thresholds and response procedures. The Produce stage implements continuous monitoring, and the Evaluate stage reviews whether drift detection is catching degradation before business impact.
Related Terms
Other glossary terms mentioned in this entry's definition and context.