TCAN: An Asymmetry Modeling Network for Time Series Forecasting

ICLR 2026 Conference Submission14885 Authors

Keywords: time series, convolutional network, forecasting, asymmetric modeling

TL;DR: This study breaks through the existing research paradigm and proposes a time series convolutional association network from the perspective of information correlation and asymmetric modeling.

Abstract: Most existing time series forecasting methods assume shared statistical consistencies across variables, such as periodicity. This assumption enforces symmetric modeling with shared encoders, yet real-world datasets often reveal distinct primary cycles for different variables. To address this gap, we introduce the Temporal Convolutional Association Block (TCAB), a flexible temporal convolution module that combines the strengths of attention and convolution to enable efficient asymmetric modeling of temporal and causal relationships. TCAB performs patch-wise equivalent sequence modeling by replacing attention score computation with learnable weights while preserving relative positional information. Building on TCAB, we propose the Temporal Convolutional Association Network (TCAN), a framework designed to capture asymmetric long-term dependencies and causal relationships across variables and patches. Extensive experiments on seven real-world datasets demonstrate that TCAN consistently outperforms state-of-the-art methods, validating the effectiveness of TCAB and providing a robust solution for efficient asymmetric modeling in multivariate time series forecasting. The code is available at https://anonymous.4open.science/r/TCAN-8F21.

Primary Area: learning on time series and dynamical systems

Submission Number: 14885