[4Yin-B] Poster 7

[4Yin-B-16] Integration of a Text-Based Causal Suggestion System and Time-Series Forecasting Models for Evidence-Based Decision Support

〇Eijiro Mochida¹, Ryuji Hashimoto¹, Takuya Yasuda¹, Yuri Murayama¹, Kiyoshi Izumi¹ (1. The University of Tokyo)

Keywords:

Forecasting Performance Evaluation,Time Series Forecasting,Causal-Chain Presentation System,Large Language Models,Causal Discovery

Evidence-based decision-making is increasingly emphasized in corporate activities and policymaking, highlighting the need for causal analysis to assess how interventions affect outcomes. Recent methods extract causal relationships from text, but their support in time-series data remains underexplored. We propose a framework that evaluates text-derived causal factors through time-series validation and forecasting. First, we map factors from LLM-based causal chains to time-series variables. We then use statistical tests to identify factors and lag structures that improve forecasting beyond the target ’s autoregressive history. Finally, we compare an autoregressive-and-calendar model with an augmented model that adds the selected factors, evaluating predictive accuracy and feature contributions. Results show that for heatstroke ambulance transports in Tokyo, selected factors substantially reduce prediction error, enabling quantitative identification of key predictive drivers. In contrast, for nationwide grocery expenditure, autoregressive and calendar features already achieve high accuracy, and gains from added factors are limited.