A paper on arXiv, 23 May 2026, tested time-series foundation models across four operational domains, including demand forecasting. The takeaway for anyone sitting on real data: foundation models hold up well in the cold-start scenario and on transferable periodic structures. Translated: they give you a usable forecast even on a new product, or one with thin history, without waiting months of sales. Supervised methods stay more accurate where the process is physically constrained and history is abundant.
The authors propose a Complexity Router: it assigns each series to the best model based on empirical features, and that cuts inference costs. Not a single model, a choice per series. For anyone running a mixed catalog, mature products and launches, it's the difference between a forecast that runs and one that costs too much for what it's worth.
Why this matters for anyone building enterprise AI: if you have no history, don't wait, start from the foundation model and keep the supervised one where the data exists.