Hybrid foundational + classical forecasting
FORGE framework: combining time-series foundation models with classical statistical baselines for production retail forecasting.
The problem
Retail demand forecasting trades off accuracy against interpretability. Foundation models offer accuracy gains but can't be calibrated, audited, or trusted in production decision-making contexts. Classical models are interpretable but plateau on accuracy.
Why standard tooling didn't fit
Picking one model family forces a tradeoff. Production teams need a system that can use the foundation model's accuracy where it matters and the classical model's interpretability where it matters more, then combine them coherently.
What we built
The system does not commit to one model family. It runs a time-series foundation model and a classical statistical baseline over the same history, then decides per series which of the two to trust. That judgement is driven by how much history exists and how seasonal or sparse the signal is. Confidence is derived inside each family and then reconciled, so the published forecast carries its own interval rather than a borrowed one. Series are watched for drift and retrained once they stop behaving like the data the model was fitted on. Output is shaped for the systems that consume it, so a forecast can drive replenishment and scenario work without being reinterpreted on the way.
Outcomes
- Measurable wMAPE improvement vs. classical-only baseline at SKU × week, with confidence intervals retained per forecast.
- Built on research published at AAAI and KDD venues.
- Productionized inside Horizon for retail customers.