Time series foundation models (TSFMs) are pretrained on series from diverse domains, yet demand series constitute only a tiny fraction. Demand data exhibit short histories, frequent zeros, censoring due to stock‑outs, and exogenous events that the series itself does not record. To address these gaps we introduce EXAONE Demand, built around a demand‑specific corpus and a demand‑aware adapter. The corpus aggregates 73 sources, yielding 11.3 M series and 48.4 B observations, and a synthetic generator supplies behaviours that open demand data under‑represent. The adapter attaches low‑rank branches to a frozen general‑domain backbone, one branch for each of the four demand classes—smooth, intermittent, erratic, and lumpy—and a router that reads eight scale‑free statistics of the input series to weight the branches. Two versions are trained: one on real‑world plus synthetic demand, and another on synthetic data alone. On 22 held‑out datasets both versions outperform 36 existing TSFMs, and incorporating real‑world demand yields additional gains over synthetic‑only training.
Review