COMPEL Glossary / drift-detection
Drift Detection
Drift detection is automated monitoring that identifies when the statistical properties of input data or model outputs have shifted significantly from baseline measurements established during model training or initial deployment.
What this means in practice
Drift can manifest as data drift (changes in input distributions), concept drift (changes in the underlying relationships being modeled), or prediction drift (changes in the distribution of model outputs). Effective drift detection requires establishing baseline distributions during deployment, continuously comparing production data against these baselines, and defining thresholds that trigger alerts and response procedures. Drift detection prevents the slow, silent degradation that causes AI systems to become less accurate over time without anyone noticing until business outcomes deteriorate. In COMPEL, drift detection capability is a key differentiator in the MLOps maturity domain.
Why it matters
Drift detection prevents the slow, silent degradation that causes AI systems to become less accurate over time without anyone noticing until business outcomes have already deteriorated. Without automated drift monitoring, organizations discover performance problems only through downstream impacts like declining revenue or rising customer complaints. Early detection enables proactive retraining before business value is lost.
How COMPEL uses it
Drift detection capability is a key differentiator in the MLOps maturity domain (Domain 7) assessed during Calibrate. During Model, drift detection thresholds and response procedures are designed as part of the operational monitoring infrastructure. The Produce stage implements automated drift monitoring, and the Evaluate stage reviews drift alert effectiveness and response times as operational resilience metrics.
Related Terms
Other glossary terms mentioned in this entry's definition and context.