ReasonCast: Agentic Demand Forecasting with Selective Semantic Reasoning
| Source: arXiv AI
Tags: demand forecasting, LLM, time series, e-commerce, reinforcement learning, supply chain
ReasonCast reduces demand forecast error by 3.29 WMAPE points on holiday-sensitive categories by teaching LLMs to selectively intervene with structured event fields (direction, shape, amplitude, peak) — while suppressing unnecessary intervention on stable periods to avoid a 1.68-point penalty.
Details
Demand forecasting requires combining historical patterns with forward-looking event knowledge (promotions, holidays, price changes). Most text-enhanced forecasting methods inject event context uniformly, regardless of whether it actually affects demand for specific products. ReasonCast introduces selective semantic intervention: an agent first examines the event context, the baseline forecast, and forecast uncertainty to decide whether textual reasoning is needed at all. When it intervenes, it uses structured fields — event relevance, demand direction, temporal shape, amplitude, and peak intensity — rather than free-form text. These fields interact with temporal components of a time-series foundation model via additive (trend/shape) and multiplicative (level shift) paths. Training uses a three-stage curriculum: schema SFT to establish the field structure, semantic-field RL to calibrate directional and shape judgments, and forecast-utility RL to align reasoning with actual forecast improvement. On holiday-sensitive categories, WMAPE drops by 3.29 points. Crucially, suppressing unnecessary intervention avoids the 1.68-point penalty that indiscriminate intervention causes on stable-sales periods.