Rolling backtests
Move forecast origins forward through time. Each run uses only history available before that origin, including all additional input metrics. Match windows and horizons to actual operating decisions.
Read resource prediction performance
Accuracy concerns resources such as CPU, memory, and GPU. Replica counts are derived from resource demand and constraints; distinguish them from actual and Ready replicas.
| Measure | Meaning and use |
|---|---|
| WAPE | Total absolute error relative to total absolute actual values over the same window. |
| Peak underestimation | The risk of predicting less than actual high-water demand. |
| Resource accuracy and score | Read resource performance, completeness, and protection factors together. Scores do not override execution bounds. |
| Actual evaluation | Link observations to predictions after their target times arrive. |
Scores, rankings, and consensus
The rankings page shows current performance and historical changes. Rank helps compare candidates; look-ahead decisions combine available forecasts and confidence rather than simply following the top-ranked method.
Inspect training and run details for window, version, parameters, duration, and reasons. Recheck performance after configuration changes.
Wait for actuals and keep evaluating
Future target times do not yet have actual observations; missing evaluations are not zero error. Continue reviewing resource behavior and outcomes as inputs and workload patterns change.
Read a module effect report
Select a module and time window in forecasting results to review resource error, peak underestimation, scale-up timeliness, excess and missing capacity, reversals, degradation reasons, and algorithm adoption. Insufficient or unmeasurable data is identified rather than treated as zero error.
Actual Pod-hours and capacity SLO
The execution and capacity section compares observed Ready replicas with replicas required by measured load. It reports actual, needed, excess, and missing Pod-hours and minutes of insufficient capacity. Data gaps are excluded from compared time.
Capacity SLO is time without a capacity shortfall divided by compared time. Its default target is 99%, configured with forecast.slo_capacity_target_percent. It is not a latency or error-rate SLO; excess Pod-hours are not proven monetary savings. Values require at least 12 compared moments covering at least half the requested window.
Preview configuration changes
Open change preview from the effect report to compare saved and proposed settings on the same history. Review shortfall and surplus windows, risk error, replica-hours, changes, and the largest differences. Compare methods and parameters, safety margin, per-Pod capacity, replica bounds, and short-term correction.
Preview is read-only: it saves no recommendations, changes no modules, and performs no scaling. It compares forecast-derived replicas without fully replaying cooldowns, steps, look-ahead, or safeguards. Model services use pretrained models rather than background-tuned copies. Interpret results within this scope.