Model health

checking…
U
user #0 · admin

Model health

Rolling MAPE, prediction-interval coverage, and drift score per series (LLD §6.5). High drift in >50 % of active series triggers a drift.widespread log entry but no auto-refit — a human should look.

Asset Parameter Kind MAPE Coverage Drift Window end
loading…

MAPE: mean absolute % error of the forecast median vs. actuals. Can spike for series whose actuals hover near zero — that's a known artefact, not a real regression. Coverage: fraction of actuals that fell inside the [q10, q90] band; should be ≈ 0.80 if the model is well calibrated. Drift: composite score, >MAPE_DRIFT_THRESHOLD means refit recommended.

Model health — how to read this page

What this page shows

How well the forecasting model is doing, per series. Every 5 minutes the service compares the model's predictions to what actually happened and writes a score. This table shows the latest score for each active series.

The three numbers

What happens automatically

If more than half of all active series are drifting at the same time, that's almost always an upstream problem (sensor outage, clock skew). The service logs a drift.widespread warning but deliberately doesn't auto-retrain — a human should look first.

Filtering

Use the dropdowns at the top to narrow to a zone, division, station or single asset — each choice limits the next, and every level is optional. Once an asset is picked you can also focus on one parameter. Clear drops all filters.

How to use this page

Sort by drift to find series whose model has gone stale. Sort by coverage (low → high) to find the series most often producing suspicious-looking readings.