Monitoring snapshots compare completed forecasts against observed actuals. The current release is not a live stream. It computes a status view from the forecast and actual paths available at run time.
Metrics
For each model or combined forecast, the monitor reports:
- forecast error
- absolute error
- squared error
- rolling bias
- rolling MAE
- rolling RMSE
- deterioration ratio against baseline RMSE
The window size is configurable and capped by the sample size.
Alerts
Alerts can be based on rolling RMSE or latest absolute error. The threshold is a multiplier. The output labels each alert as watch or alert depending on the rule and observed value.
Health Status
Health rows summarize the latest state for each input. A healthy status does not mean the model is structurally correct. It means the recent error pattern does not breach the selected threshold under the chosen window.
Limits
- Scheduled refresh is separate work.
- Websocket or server-sent event delivery is separate work.
- External feed transport is separate work.
- Monitoring quality depends on timely actuals and consistent series definitions.