FACT: Frequency-Aware Channel-Guided Multivariate Time Series Forecasting

ICLR 2026 Conference Submission18042 Authors

Keywords: Time Series Forecasting, Multivariate Time Series, Spatiotemporal Dependencies, Fine-grained Interpretability, Deep Learning

Abstract: Forecasting Multivariate Time Series (MTS) requires capturing complex intra-channel dynamics and evolving inter-channel dependencies. However, existing methods often struggle to disentangle meaningful signals from inter-channel noise and intricate interaction patterns. To address this, we propose a novel framework that operates entirely in the frequency domain, modeling inter-channel relationships at the component level. Our approach first dynamically decomposes each time series into its constituent frequencies. A channel masking mechanism then identifies and isolates the most salient frequency components, simultaneously filtering noise and enhancing computational efficiency. This allows our model to capture time-varying inter-channel dependencies with high fidelity. Furthermore, our learning objective effectively balances accuracy against regularization constraints for both computational efficiency and interpretability. Extensive experiments on diverse, real-world datasets demonstrate that our method achieves competitive performance. Code is available at this repository: https://anonymous.4open.science/r/FACT.

Primary Area: learning on time series and dynamical systems

Submission Number: 18042