What is Model Monitoring & Drift

Definition

Model monitoring and drift detection is the ongoing tracking of a deployed model's performance and inputs to catch degradation, particularly drift, where the live data or relationships diverge from training conditions, so models can be retrained or corrected before predictions become unreliable.

« Back to Glossary Index

Detects when live performance degrades from its trained baseline
Catches data and concept drift as real-world inputs change over time
Triggers retraining before stale models cause costly wrong predictions
Maintains trust by proving deployed models still perform as expected

Real World Example

A credit-scoring model is monitored in production; when consumer behaviour shifts during an economic change, drift detection flags rising input divergence and falling accuracy, prompting a retrain before bad decisions accumulate.

FAQs

What is model drift?

Drift is when the live data distribution (data drift) or the input-output relationship (concept drift) diverges from training, degrading model accuracy.

How is drift detected?

By monitoring input distributions and prediction performance over time and alerting when they deviate beyond thresholds from the baseline.

What happens when drift is detected?

Teams investigate and typically retrain or recalibrate the model on fresh data to restore reliable performance.