Definition

A lagging indicator is a metric that confirms the impact of changes after they have already occurred.

  • They measure outcomes, not early signals.
  • In ML monitoring, these are model performance metrics (e.g., loss, AUC, accuracy, calibration).

Characteristics

  • Reactive → only show issues once they’ve happened.
  • Direct measure of end results (model performance, business KPIs).
  • Often used for validation and confirmation.

Examples in Machine Learning

  1. Model Performance Metrics
    • Accuracy, Precision, Recall, F1, AUC.
    • Log loss, cross-entropy.
    • Calibration error.
  2. Business KPIs (after the fact)
    • Click-through rate (CTR) decreases.
    • Fraud losses increase.
    • Customer churn rate rises.
  3. Monitoring Context
    • You notice a drop in AUC → drift already harmed predictions.
    • Loss spikes after new feature distribution changes.

Why They Matter

  • They confirm whether leading indicators (drift, input quality issues) had real impact.
  • Critical for assessing model health and business outcomes.
  • Used in go/no-go retraining decisions.

Example

  • Fraud detection model:
    • Leading indicator: Feature distribution for transaction_type drifts.
    • Lagging indicator: Model AUC drops from 0.87 → 0.72 (fraud detection is worse).

Here, the lagging indicator proves the model is now underperforming.

Summary
Lagging indicators = outcome metrics that confirm the actual impact on model performance or business results.
They are reactive and usually follow leading indicators.