PaperlayerView exact revision on arXiv

Source-linked AI summary

ReasonCast: Agentic Demand Forecasting with Selective Semantic Reasoning

Ziyue Yang, Chaolin Xu, Yijing Wang, Tiankai Gu, Hui Yang, Yanhong Lin, Kaiyuan Liu, Fei Xiao

arXiv:2608.15291v1cs.AI

TL;DR

Demand forecasting must combine historical dynamics with future event context without uniformly injecting irrelevant semantics. ReasonCast selectively routes structured semantic interventions into a numerical forecaster, reducing WMAPE on event-sensitive categories while preserving performance when intervention is unnecessary.

  • Problem

    Forecasting methods lack a principled way to determine whether, where, and how much semantic event context should alter a strong numerical forecast.

  • Method

    ReasonCast uses an agent, instance-wise gating, structured event interactions, and additive or multiplicative corrections to selectively intervene on temporal forecasts.

  • Results

    7.8%, 5.6%, and 1.4% WMAPE reductions occur on holiday-sensitive categories, mega-sale-sensitive categories, and M5 event windows, respectively.

  • Takeaways & Limitations

    Forecast-aligned routing exploits event-driven shifts while preserving the numerical backbone when semantic intervention is unnecessary.

  • Takeaways & Limitations

    Event effects vary across products, regions, and seasons, requiring recalibration under distribution shift.

Abstract

from arXiv · show

Demand forecasting increasingly requires combining two complementary sources of information: historical sales reveal recurring numerical dynamics, while future promotions, holidays, price changes, and platform interventions provide forward-looking knowledge. Existing text-enhanced forecasting methods often encode such context into generic representations and fuse it uniformly with time-series features, without explicitly distinguishing which semantic effects are forecast-relevant or how they should modify future dynamics. We introduce ReasonCast, a structured semantic intervention framework that translates event knowledge into forecast-specific operations. An agent examines the event context, the no-text forecast, and its uncertainty to determine whether textual reasoning is needed. Rather than injecting free-form text, ReasonCast represents event knowledge through structured fields describing event relevance, demand direction, temporal shape, amplitude, and peak intensity. These fields interact selectively with temporal components of a time-series foundation model. An additive path corrects local trends and temporal shapes, while a multiplicative path captures event-driven level shifts. ReasonCast introduces a forecast-grounded post-training curriculum. Schema SFT establishes semantic fields; semantic-field RL calibrates direction, shape, amplitude, and peak judgments; and forecast-utility RL evaluates semantic interventions through a frozen forecaster, aligning reasoning outputs with marginal forecast improvement. ReasonCast lowers WMAPE by 3.29, 1.25, and 0.47 percentage points on holiday-sensitive categories, mega-sale-sensitive categories, and M5 event windows, respectively. On stable-sales periods, indiscriminate semantic intervention increases WMAPE by 1.68 percentage points, whereas suppressing unnecessary intervention preserves the numerical backbone.

1 Introduction

ReasonCast treats event-enhanced demand forecasting as selective semantic intervention, routing each instance according to whether context adds forecast-relevant information. It injects structured event semantics through gated additive and multiplicative paths while preserving the numerical forecaster when intervention is skipped.

  • Motivation: Historical sales capture recurring dynamics, but future promotions, holidays, prices, interventions, and product events can create shifts unavailable from past sales alone.
  • Motivation: Semantic value depends on whether text contributes information beyond numerical history, and unfiltered event evidence can degrade forecasting.
  • Selective semantic intervention: ReasonCast routes each instance to Skip, low-cost Basic semantics, or tool-augmented Tool semantics using event context, the no-text forecast, and its entropy.Holidays and mega-sales are required to activate semantic reasoning, while a pretrained TSFM remains responsible for numerical prediction.
  • Structured intervention: Instance-wise gating combines additive paths for local trends and shapes with multiplicative paths for level changes caused by major events.Semantic tokens are queried only by event-related temporal components, and skipped instances exactly recover the no-text computation.
  • Training and evaluation: Forecast-aligned policy optimization rewards routing and semantic outputs according to marginal improvement over the numerical backbone.The framework uses alignment pretraining to ground event semantics and evaluates performance across normal, event-driven, and corrupted-context periods.

2 Related Work

Related work advances numerical forecasting through specialized architectures and time-series foundation models, while text-enhanced and agentic methods add future-event information through language-based representations, evidence selection, or adaptive fusion. ReasonCast distinguishes itself by retaining the TSFM as forecast owner and selectively routing instances to structured semantic interventions rather than generating sales values.

  • Numerical forecasting: Numerical forecasting has progressed from probabilistic and covariate-aware models to specialized linear, patch-based, inverted, and multiscale architectures.Time-series foundation models extend this progression by transferring recurring structure across datasets through tokenization, decoder-style generation, or heterogeneous-series pretraining.
  • Numerical forecasting: Time-series foundation models are strong numerical specialists but cannot infer future events absent from their inputs.This limitation motivates incorporating external event knowledge rather than relying on historical numerical inputs alone.
  • Language models for forecasting: Pretrained language models have been repurposed as numerical forecasters or representation learners, but language components do not consistently explain the resulting gains.ReasonCast therefore retains a TSFM as forecast owner and assigns the LLM to reason over event knowledge rather than generate sales values.
  • Text-enhanced forecasting: Text-enhanced forecasting uses endogenous series descriptions or exogenous evidence such as news, events, and domain conditions.Multimodal gains depend on backbone capacity, alignment, data scale, and whether text adds predictive information.
  • Agentic forecasting: ReasonCast routes each instance among no intervention, low-cost reasoning, and tool-augmented reasoning, with structured semantics modifying a dedicated TSFM only when selected.This differs from prior approaches such as validation-feedback evidence selection and retrieval-based adaptive fusion.

3 Problem Formulation

ReasonCast formulates forecasting as predicting the next H = 7 observations from an observed demand history and pre-forecast event context. A pretrained Chronos-2 model supplies the no-text forecast, while holidays and mega-sales require semantic processing rather than Skip routing.

  • Forecast setup: The task predicts the next H = 7 observations from an observed demand history of length L and pre-forecast event records c_i.The event context is available before the forecast origin.
  • Forecast setup: A pretrained Chronos-2 model f_θ provides the no-text forecast used as the numerical baseline.This forecast is produced before semantic intervention.
  • Semantic routing: Holidays and mega-sales enforce a_i ≠ Skip; Skip yields T_i = ∅, whereas other cases generate semantic tokens with π_sem.The formulation distinguishes events requiring semantic processing from instances routed to Skip.

4 ReasonCast

ReasonCast selectively reasons over event context, aligns structured semantics with temporal representations, and applies gated additive and multiplicative interventions to forecasting. Its three-stage training curriculum calibrates semantic fields and optimizes forecast utility against a frozen forecaster while preserving exact backbone behavior when reasoning is skipped.

  • Training curriculum: The curriculum combines schema SFT, semantic-field RL, fusion-corpus regeneration with alignment and dual-path fusion, and forecast-utility RL against a frozen forecaster.Forecast-utility RL makes Skip compete directly with semantic intervention, while Tool use must justify its additional cost.
  • Selective reasoning: ReasonCast routes each instance to Skip, Basic, or Tool, using at least Basic for hard-triggered holidays and mega-sales.The agent decides whether and how much to reason while the time-series foundation model retains numerical forecasting.
  • Structured semantics: Structured outputs encode event relevance, dominant event, demand direction, temporal shape, amplitude, peak intensity, and a forecast-trend narrative.SFT establishes the schema, while forecast-aligned optimization calibrates routing and semantic utility.
  • Selective intervention: Only the event component of temporal tokens queries text through an orthogonal event basis, followed by masked, gated fusion before the Chronos-2 prediction head.The interface explicitly avoids claiming that the semantic update remains confined to the event component after attention.
  • Dual-path fusion: Additive latent residuals target fine-grained temporal pattern changes, while a complementary multiplicative correction captures semantic level shifts.An instance-wise gate uses semantic features, recent-demand volatility, no-text forecast entropy, and event descriptors to modulate the additive residual.
  • Backbone preservation: Skipped instances with unavailable text recover the Chronos-2 backbone prediction exactly through routing and an explicit mask.This guarantee applies to the fusion interface and does not establish causal or globally identifiable event subspaces.

5 Experiments

Experiments show that ReasonCast improves event-period forecasting while preserving stable-period performance through selective routing and adaptive semantic intervention. Ablations and post-training analyses attribute gains to structured correction paths, entropy-conditioned gating, and forecast-utility alignment.

  • Main results: ReasonCast achieves the lowest WMAPE on all six proprietary event slices and M5 event windows, with the largest margins during holidays.It also attains the best WMAPE and second-best MAE on mega-sale-insensitive categories.
  • Routing and stable periods: Always-on semantics degrade the Chronos2-small backbone from 7.91%/185.15 to 9.59%/224.43, whereas agent routing exactly recovers the backbone.Event windows require at least Basic reasoning, while stable periods benefit from the exact no-text Skip path.
  • Selective fusion: Removing orthogonal decomposition–alignment or adaptive gating increases WMAPE, while additive correction helps both event regimes and multiplicative correction is especially important for large holiday level shifts.The ablations support adaptive aggregate control and complementary additive–multiplicative corrections.
  • Reasoner post-training: SFT establishes parseability and improves direction and shape; semantic-field RL calibrates direction, amplitude, and peak, while forecast-utility RL targets Event WMAPE and NTR without sacrificing semantic fidelity.The staged curriculum separates well-formed semantics from interventions that improve the final forecast.
  • Adaptive gating: Adaptive gating outperforms a global α in every entropy tier, with larger gains at medium and high entropy as semantic correction strength tracks no-text forecast uncertainty.The comparison isolates continuous gate control after a non-skip route.
  • Semantic sensitivity: Incorrect semantics increase WMAPE in 13 of 15 windows, with mean and median increases of 2.67 and 0.76 percentage points and a worst-case increase of 25.33 points.The stress test shows forecasts respond to semantic content rather than merely text presence, although gating does not guarantee per-window immunity.

6 Discussion

ReasonCast supports treating event-enhanced forecasting as conditional intervention rather than uniform multimodal fusion. Its gains depend on selective routing, entropy-adaptive control, structured temporal interactions, and forecast-grounded post-training, while deployment must address unnecessary intervention and distribution shift.

  • Conditional intervention: 7.8%, 5.6%, and 1.4% WMAPE reductions occur on holiday-sensitive categories, mega-sale-sensitive categories, and M5 event windows, respectively, versus the strongest non-ReasonCast baselines.The clearest gains arise when future context is difficult to recover from historical demand alone.
  • Conditional intervention: Discrete routing and continuous gating decide when semantic reasoning is invoked and how strongly it influences forecasts, while orthogonal decomposition restricts interaction to event-related temporal components.Removing orthogonal decomposition and alignment degrades all evaluated slices.
  • Conditional intervention: Adaptive gating outperforms global-strength control across all forecast-entropy tiers, assigning greater semantic influence when the numerical forecast is less reliable.The additive path improves event-sensitive forecasts but can harm stable demand when intervention is forced.
  • Stable-sales behavior: The stable-sales analysis forces every instance onto the Basic route to diagnose unnecessary fusion, whereas Skip can disable both correction paths and exactly recover the numerical backbone.Thus, the stable-sales column is not end-to-end ReasonCast performance.
  • Post-training: Schema SFT establishes parseable outputs, semantic-field RL improves direction, amplitude, and peak-intensity judgments, and forecast-utility RL scores candidates using predictive improvement, harm, and tool cost.The distinction matters because well-formed event descriptions are not necessarily useful interventions.
  • Deployment challenges: Deployment requires recalibration under distribution shift, validation of external signals with uncertainty-aware reasoning, and management of the trade-off introduced by production-scale tool-augmented reasoning.Event effects vary across products, regions, and seasons, while signals differ in timeliness, completeness, and reliability.

7 Conclusion · A Detailed Window-Level Results

ReasonCast routes event semantics through selective, structured interventions rather than treating them as always-on forecasting inputs, with training stages aligned to downstream predictive value. The appendix provides unaggregated seven-day forecast-window results using consistent items, origins, and horizons without window-specific model selection.

  • 7 Conclusion: ReasonCast treats event semantics as a routed intervention instead of an always-on forecasting input.The framework supports selective use of semantic information during forecasting.
  • 7 Conclusion: The agent selects among no intervention, low-cost structured reasoning, and tool-augmented reasoning.These are the framework’s available reasoning paths for handling event context.
  • 7 Conclusion: Continuous gating controls intervention strength after the agent chooses not to skip reasoning.Gating adjusts how strongly the selected semantic intervention affects forecasting.
  • 7 Conclusion: Schema SFT, semantic-field RL, and forecast-utility RL progressively align structured reasoning with downstream predictive value.The curriculum combines schema establishment, semantic-field calibration, and forecast-grounded utility evaluation.
  • A Detailed Window-Level Results: The appendix reports unaggregated seven-day forecast windows underlying the main-text averages.The reported windows are the detailed results behind the aggregated averages.
  • A Detailed Window-Level Results: Each appendix cell contains WMAPE and MAE for exactly the same items, forecast origin, and horizon.This preserves comparability across the detailed window-level results.
  • A Detailed Window-Level Results: No window-specific model selection is performed in the detailed results.The appendix therefore does not select models separately for individual forecast windows.
  • A Detailed Window-Level Results: Chronos2-small is abbreviated as C2-small, and the time-series backbone is abbreviated as TS backb.These abbreviations are used for compact presentation in the appendix.

A.1 Commerce Dataset

The commerce dataset comprises 15 chronological windows spanning mega-sale, holiday, and stable-sales periods. The hardest cases are Spring Festival–Valentine windows, where ReasonCast substantially lowers WMAPE on February 12.

  • Dataset composition: The dataset contains 15 commerce windows: three mega-sale, ten holiday, and two stable-sales windows.Tables 6 and 7 present identical rows and evaluation protocols; their split is purely presentational.
  • Challenging windows: The two hardest cases are Spring Festival–Valentine windows beginning on February 12 and February 15.These windows are identified as the most difficult cases in the commerce evaluation.
  • Forecast results: 108.75% to 83.85%: ReasonCast reduces WMAPE on the February 12 window versus the strongest non-ReasonCast comparator.The cited comparison reports the strongest non-ReasonCast comparator at 108.75% and ReasonCast at 83.85%.

A.2 M5 Dataset

ReasonCast is evaluated across all 14 M5 event windows, with each holiday assessed at two forecast origins. It remains consistently competitive across major holidays, though it does not dominate every origin.

  • Evaluation coverage: All 14 M5 event windows are reported, with each holiday evaluated at two forecast origins to avoid dependence on one event-day alignment.Results are presented in Tables 8 and 9.
  • Results: ReasonCast remains consistently competitive across Halloween, New Year, Super Bowl, Easter, and Mother’s Day forecast origins.The M5 benchmark exhibits smaller absolute differences and a different error profile from the commerce data.
  • Results: ReasonCast does not dominate every M5 forecast origin, indicating that its competitiveness is not universal across event windows.

A.3 Semantic Corruption Stress Test … B.2 Structured Fields and Representative Rationale

Semantic inversion substantially harms forecasting, with especially severe failures in event-sensitive windows. ReasonCast separates routing from semantic interpretation and uses structured fields, optional tool augmentation, and forecast-utility optimization to control interventions.

  • A.3 Semantic Corruption Stress Test: Incorrect semantics increase WMAPE in 13 of 15 windows, by 2.67 percentage points on average.The test inverts six control fields and the final trend narrative while holding the time-series input, checkpoint, and inference configuration fixed.
  • A.3 Semantic Corruption Stress Test: The median increase is 0.76 percentage points, while the February 12 Spring Festival–Valentine window increases by 25.33 points.The right-skewed distribution shows that aggregate averages can obscure both typical harm and severe event-sensitive failures.
  • A.3 Semantic Corruption Stress Test: Gating attenuates many incorrect interventions but cannot guarantee that every corrupted description worsens every finite test slice.Two small negative deltas remain in the controlled inversion test.
  • B.1 Agent Policy and Basic Semantic Interpreter Prompt: The agent policy selects Skip, Basic, or Tool using event context, the no-text forecast, and predictive entropy.Skip preserves the numerical backbone; Basic uses available context, while Tool adds retrieved event evidence or temporal statistics before semantic interpretation.
  • B.1 Agent Policy and Basic Semantic Interpreter Prompt: The Basic semantic interpreter prompt is invoked only after a non-skip decision, with scale judgments relative to the most recent seven-day mean.The Basic route supplies original context, whereas the Tool route supplies tool-augmented context.
  • B.1 Agent Policy and Basic Semantic Interpreter Prompt: Schema SFT and semantic-field RL train valid, consistent, and calibrated fields, while forecast-utility RL optimizes routing, tool use, and intervention decisions against frozen-forecaster utility.Forecast-utility RL includes explicit penalties for negative transfer and tool cost.
  • B.2 Structured Fields and Representative Rationale: Every final policy response begins with one machine-readable routing action; Basic and Tool additionally return six control fields and a short trend narrative.Skip activates the no-text path, while tool calls and observations remain intermediate trajectory elements rather than final-schema fields.
  • B.2 Structured Fields and Representative Rationale: A representative teacher example presents a concise forecasting rationale, with Chinese compressed text accompanied by an English translation for readability.The rationale is not a raw token-by-token trace of the teacher’s internal generation.

B.3 Teacher Distillation and Rationale Compression · B.4 SFT, Semantic-Field RL, and Forecast-Utility RL Configuration

ReasonCast builds its SFT pool through regime-stratified teacher distillation, then compresses rationales while preserving the forecasting decision sequence. Its two GRPO-based RL stages separately calibrate semantic fields and evaluate gated interventions with a frozen fusion model.

  • B.3 Teacher Distillation and Rationale Compression: 11,627 successful teacher generations are filtered by direction, amplitude, and peak consistency, retaining 7,312 candidates.Claude Opus 4.6 provides structured semantic supervision using only the Basic semantic interpreter prompt, without routing or tool interaction.
  • B.3 Teacher Distillation and Rationale Compression: Teacher rationales are translated into Chinese and compressed into a single 200–300-character paragraph while preserving the original forecasting decision order.The preserved sequence covers item and window, event position, baselines and historical statistics or the no-YoY branch, relevance, event selection, semantic fields, and recalibration.
  • B.3 Teacher Distillation and Rationale Compression: 568.99 to 206.61 to 202.77 tokens describes mean rationale length across the original LLM, SFT, and semantic-field RL stages.This corresponds to a 64.36% reduction relative to the original LLM, alongside declining direction, amplitude, and peak errors; the result is cumulative rather than an isolated compression ablation.
  • B.4 SFT, Semantic-Field RL, and Forecast-Utility RL Configuration: Both RL stages use group-relative policy optimization (GRPO), with semantic-field RL targeting forecast-critical semantic fields.The configuration tables record the main optimization settings for SFT and semantic-field RL.
  • B.4 SFT, Semantic-Field RL, and Forecast-Utility RL Configuration: Forecast-utility RL applies a semantic validity gate before evaluating gatepassing candidates with the frozen fusion model.Its main optimization settings are recorded in the forecast-utility RL configuration table.
Loading 2608.15291v1…