Cluster-level AutoFit¶
AutoFit runs backtests: it hides historical periods, forecasts as if it were at that point in time, and compares the result with actuals. In cluster mode, it selects a winning configuration for each material-cluster and location-cluster combination.
AutoFit does not create the official Demand Plan. It creates a configuration that can be used by the official run.
The AutoFit run owns candidate models, backtests, and the selection metric. The cluster-level configuration owns aggregation, split, unit, historical windows, and participation in the official run. They share the same profile and cluster pair, but they are not the same record.
The profile, clusters, historical window, and AutoFit use must be reviewed together.
Prerequisites¶
Confirm reconciled history, clusters, the planning calendar, planning frequency, minimum training length, and an appropriate selection metric.
Essential settings¶
| Setting | Decision |
|---|---|
| History end date | Last closed bucket available |
| Training window | Historical periods available to each candidate |
| Validation horizon | Periods forecast in each backtest |
| Backtest origins | Historical cutoffs evaluated |
| Candidate models | Statistical families allowed |
| Selection metric | Rule for comparing candidates |
| Evaluation level | Cluster combination where error is consolidated |
Percentage errors may be unstable when actuals are zero or very small. For intermittent demand, also inspect absolute error, bias, and operational usefulness.
Run and review¶
- Create an AutoFit model and associate its execution profile.
- Select cluster-level mode.
- Configure history, horizon, origins, candidates, and metric.
- Start and monitor the process.
- Review the winner for each cluster pair.
- Check stability, bias, and disaggregation.
- Approve the completed model for use by the Enterprise profile.
When an approved AutoFit model is promoted as the profile default, the Enterprise execution resolves the winner for each cluster pair. The remaining calculation rules still come from that pair's cluster-level configuration.
Lowest error is not the whole decision
A statistical winner may respond poorly to launches, stockouts, campaigns, or structural changes. AutoFit reduces repetitive work; it does not remove governance.
Next step¶
Also see the execution profile, forecast cluster, and cluster-level calculation fields.