ReasonCast and the Case for Restraint in Demand Forecasting

On August 15, a team of researchers (Yang, Xu, Wang et al.) posted ReasonCast on arXiv — a demand forecasting framework built around one question most systems never ask: does this particular promotion or holiday actually change the forecast, and by how much?

The problem with "always-on" context

Most "context-aware" forecasters fold every promo flag, holiday, or news event into a generic text embedding and feed it to the model uniformly. ReasonCast instead runs a lightweight reasoning step per event: it estimates relevance, direction, shape, and intensity, then applies that as an additive correction (for trend shifts) or a multiplicative correction (for level shifts) — only when the model judges the event actually matters.

Old way vs. new way

Old way: treat context as always-on — more signal in, hopefully more accuracy out.

New way: treat reasoning as a gate — only intervene when an event is judged to actually matter.

The result

On holiday-heavy categories it cut WMAPE by 3.29 points, and 1.25 points during mega-sale periods. But the more telling number is this: when that same reasoning was forced onto stable, unremarkable periods, error went up 1.68 points. Some of the gain here isn't from smarter context — it's from knowing when to leave the forecast alone.

The open question

That's an uncomfortable finding for anyone who's been told to "just add more features." It suggests part of the value in event-aware forecasting is really about restraint, not richness — and restraint is hard to train, harder to audit, and easy to get subtly wrong at scale. Is the future of demand forecasting more sophisticated event modeling, or better-calibrated systems that know when to switch the extra signal off?

Our take

It's a tension we navigate constantly building inventory and demand forecasting systems for clients: every SKU-holiday combination looks like a reason to intervene, and the discipline is knowing which ones actually are.