The question.
Operational forecasts fail when evaluation ignores hierarchy, promotions, stockouts, cold starts, or the cost of being wrong.
The study uses rolling time windows and business-specific error costs. Complex models must improve on strong seasonal and intermittent-demand baselines.
The proposed approach.
- Define forecast horizons and what was known at each cutoff
- Compare seasonal, boosting, and pretrained-model baselines
- Reconcile forecasts across product and regional hierarchies
- Evaluate prediction intervals and downstream planning costs
Target architecture.
Historical demandTime-aware featuresForecast candidatesHierarchy reconciliationUncertaintyPlanning views
The difficult parts.
- Missing demand caused by stockouts
- Sparse series and product cold starts
- Changing seasonality and promotion effects
How to evaluate it.
- Rolling-origin errors by horizon and segment
- Prediction-interval coverage and calibration
- Inventory or staffing cost under simulated decisions
Planned deliverables.
- A rolling-evaluation forecasting benchmark
- An uncertainty-aware planning report
- A baseline and model-selection study
Design principles.
Show uncertainty where a decision depends on it.
A simpler forecast is preferable when it supports the decision equally well.