ReTaMTSF: Retrieval-Based Multimodal Framework for Multivariate Time Series Forecasting

ICLR 2026 Conference Submission16562 Authors

Keywords: time series forecasting, text augmented, multimodal

TL;DR: A text-retrieval augmented multi-variate time series forecasting paradigm is proposed in this paper.

Abstract:

The rapid advancement of Large Language Models (LLMs) has ushered multivariate time series forecasting (MTSF) into a transformative era through the integration of natural language. Despite effectiveness of recent language-integrated TSF approaches, they originally stem from engineering intuition, lacking theoretical grounding, and entail considerable manual effort. Moreover, given the importance of inter-channel correlations in MTSF task, current MTSF methods either superficially investigate the intrinsic relations among time series channels or rely heavily on expert knowledge to predefine them, both are with limited flexibility. To address these challenges, we provide an information-theoretic analysis of the role of textual information in augmenting TSF and propose ReTaMTSF, an MTSF paradigm that automatically aligns and incorporates exogenous text with time series while adaptively capturing inter-channel correlations. We further introduce ReTaMForecaster, a baseline model for ReTaMTSF, and validate its effectiveness through extensive experiments on multimodal MTSF benchmarks spanning diverse domains. ReTaMForecaster achieves state-of-the-art or second-best performance in more than half of the benchmarks and forecasting horizons, with mean squared error (MSE) reductions of up to $74$\% compared to the best baseline, thereby demonstrating the soundness of the proposed framework with substantial manual effort reduction. Code is available
at https://anonymous.4open.science/r/ReTaMTSF-CC0A/.

Primary Area: learning on time series and dynamical systems

Submission Number: 16562