hesela.devPrediction cadence
Prediction cadence is the schedule or interval at which a prediction system produces successive predictions for the same monitored entity.
Cadence is distinct from telemetry sampling, the historical input window and the future prediction horizon. D-FAD received latency summaries every 20 minutes but scored a rolling 24-hour input window every four hours. Its qualifying alarm-to-replacement window was four hours to 18 days. These are separate clocks, not a recommended monitoring schedule. [1]
The original evaluation counted unresolved month-end alarms as false positives. Incomplete follow-up does not establish a full event-free horizon; the synthetic counting example explicitly assumes complete observation instead. Repeated scores are also not distinct-drive notifications. [1] [2]
Relations
- related
- prediction-horizon, false-positive-rate, precision, recall
Limitations
- Repeated scores are not automatically separate incident notifications or disk replacements; record the aggregation policy and the unit used by each metric. [2]