The operating problem
A forecast chart can look convincing while relying on information that was unavailable at the time of the decision. Aggregate accuracy can also hide persistent overforecasting or products silently missing from the evaluation. Buying decisions need evidence that is tied to a particular horizon and product universe.
What we built
In an apparel forecasting implementation, we built rolling-origin evaluation: each historical prediction uses only data available before its forecast date and is compared with the periods that follow. Candidate methods are evaluated against straightforward recent-demand baselines instead of being judged by complexity.
The reports separate error magnitude, directional bias, and product coverage. Related consumer-products work adds checks around source freshness, training cutoffs, launch availability, and consistency between demand output and inventory projection. These checks make it possible to investigate whether a surprising purchase recommendation originates in data, the forecast, or the supply calculation.
What the work enables
The delivered evaluation process gives the team evidence for selecting or revising a forecasting method. It can reveal that a simpler approach is more useful, or that an apparent improvement comes from changing the evaluated assortment. No client-specific accuracy figures or proprietary model parameters are disclosed here.
Evaluate a forecast using the information available at the decision date, a stable comparison universe, and a meaningful baseline.