Exploring Inter-Variate and Long-Term Dependencies to Boost Multivariate Time Series Forecasting

Abstract

Multivariate Time Series Forecasting (MTSF) is a critical task in various domains, and Large Language Models (LLMs) for MTSF have recently received considerable attention. Despite significant progress in large-scale time series models, particularly in fine-tuning pre-trained LLMs for MTSF, there are still limitations with existing works. First, multivariate time series (MTS) are often handled as multiple independent univariate inputs and processed separately across different variates, which neglects the dependencies between the variates. Second, most existing approaches employ Mean Squared Error (MSE) as loss function, which evaluates error at each time point separately, ignoring long-term dependencies. To address these limitations, this paper explores inter-variate and long-term dependencies to boost MTSF performance. We propose a temporal channel adapter to capture inter-variate relationships, and introduce a post-constraint module to model correlations between consecutive time points. Extensive experiments on benchmark datasets show that our method achieves state-of-the-art performance across diverse datasets and prediction horizons.