Unlocking the Value of Text: Event-Driven Reasoning and Multi-Level Alignment for Time Series Forecasting

Authors: Siyuan Wang, Peng Chen, Yihang Wang, Wanghui Qiu, Chenjuan Guo, Bin Yang, Yang Shu Date: 2026-03-16 Paper ID: openalex:2603.15452

Summary

This paper introduces VoT, a method designed to unlock the value of multimodal text information for time series forecasting, addressing the limitations of current methods that primarily rely on numerical data. VoT employs Event-driven Reasoning, leveraging Large Language Models (LLMs) guided by Historical In-context Learning to process exogenous text. It further implements Multi-level Alignment, integrating endogenous text via Endogenous Text Alignment and fusing frequency components of different predictions using Adaptive Frequency Fusion. Experimental results across ten real-world domains show substantial performance gains, validating the approach’s effectiveness in utilizing complex textual context for improved forecasting accuracy.

Key Contributions

• Proposal of VoT, a novel time series forecasting method that explicitly leverages textual information through Event-driven Reasoning and Multi-level Alignment.
• Introduction of Historical In-context Learning to provide LLMs with relevant past examples for context-aware forecasting reasoning.
• Development of Multi-level Alignment, featuring Endogenous Text Alignment at the representation level and Adaptive Frequency Fusion at the prediction level to maximize text utilization.
• Demonstration of significant performance improvements over existing methods across real-world datasets spanning 10 different domains.

Limitations

The reliance on rich exogenous text and the computational overhead associated with integrating LLMs for event-driven reasoning might limit applicability in resource-constrained or purely numerical forecasting scenarios.

Open Questions & Future Work

• hic-robustness-extreme-noise-sparsity
• aff-non-linear-fusion-potential
• pure-llm-event-driven-forecasting
• eta-alternative-decomposition-strategies

Key Concepts

• Event-driven Reasoning: A mechanism that combines rich exogenous text information with the reasoning capabilities of Large Language Models (LLMs) to enhance time series forecasting.
• Historical In-context Learning: A technique used to guide LLMs in effective reasoning for time series forecasting by retrieving and applying relevant historical examples as in-context guidance.
• Endogenous Text Alignment: A representation-level alignment technique designed to integrate intrinsic, endogenous text information directly with the time series data.
• Adaptive Frequency Fusion: A prediction-level fusion mechanism that combines the frequency components derived from event-driven predictions and purely numerical predictions.

Limitations

The reliance on rich exogenous text and the computational overhead associated with integrating LLMs for event-driven reasoning might limit applicability in resource-constrained or purely numerical forecasting scenarios.

Links

• Abstract
• PDF