Configuring the forecast
Horizon, candidate models, selection metric, reconciliation, backtests, and cold-start relations.
The Parameters step names the scenario and tells the engine how to evaluate models. The engine tests the models you pick and keeps the best one per DFU.
Horizon and granularity
Monthly horizon is how many months ahead to forecast. It must be a positive integer and shorter than the number of historical months in the file.
Granularity is monthly. Daily forecasting is marked coming soon and cannot be selected.
Candidate models
Choose at least one model. More models make the run longer. Prefer the recommended set; machine-learning models mainly help when known-future regressors exist.
The picker groups models using the Analysis diagnosis:
| Group | Models | When it is recommended |
|---|---|---|
| Recommended for your data | Auto ETS, Auto ARIMA, Auto Theta, Seasonal naive, plus Croston optimized and/or LightGBM when the diagnosis calls for them | Pre-selected from intermittency and future regressors |
| Baselines | Naive, Historical average, Window average | Simple references to compare against |
| Intermittent (Croston) | Croston optimized, Croston classic | When the file has intermittent series |
| Machine learning | GBM, LightGBM, XGBoost | When known-future regressors are present |
Select recommended restores that diagnosis-based set. You can still add or remove models by hand.
Selection metric
The metric decides which candidate wins per DFU. Lower scores are better.
| Metric | What it measures |
|---|---|
| WMAPE (default) | Volume-weighted percentage error — larger series weigh more |
| MASE | Error on the series’ own scale — comparable across items of different volume |
| MAE | Mean absolute error, in the same unit as demand |
| RMSE | Penalizes large misses; useful when peaks matter more |
| MAPE | Simple percentage error — small series can distort the result |
Reconciliation
Reconciliation makes totals consistent across hierarchy levels (for example, SKUs adding up to their category).
| Option | Role |
|---|---|
| MinT (default) | Minimum trace reconciliation across the hierarchy |
| Bottom-up | Aggregate lower-level forecasts upward |
| Top-down | Distribute higher-level forecasts downward |
| No reconciliation | Leave each level independent |
Backtest windows
Backtests replay the forecast over past periods to measure error. You can set 1–12 windows; the default is 4. More windows give a more robust evaluation and a longer run.
Cold-start relations
New or short-history DFUs can borrow structure from an existing series. Add a target DFU and a source DFU. Identifiers autocomplete from the file.
- Each target can have only one relation.
- Target and source must be different DFUs.
If you skip relations, the engine can still fall back to similar products, hierarchy share, or a group mean. Explicit relations are the method you control.
Save a draft, or Generate forecast to start Processing.