RePatch: Learning Entropy-Guided Patch Structures with Quantized Representations for Time Series Forecasting

Abstract

Multivariate time series forecasting is a fundamental yet challenging task due to non-stationary dynamics, evolving temporal structures, and limited generalization across scenarios. Recent patch-based methods improve long-range modeling by segmenting sequences into local units. Still, they typically rely on fixed or weakly adaptive patching strategies and continuous latent representations, which limit flexibility and pattern reuse. In this paper, we propose RePatch, a two-stage framework that learns adaptive and reusable temporal representations for time series forecasting. During self-supervised pretraining, RePatch introduces a Dynamic Patcher that adaptively segments time series into variable-length patches guided by an entropy-based objective, enabling precise alignment with local temporal dynamics. A Temporal Interaction Module further captures both local and global temporal dependencies and maps the resulting representations into a structured discrete latent space via vector quantization. The learned discrete representations provide a compact and transferable abstraction that naturally supports token-based sequence modeling for downstream forecasting. Extensive experiments across 14 real-world datasets demonstrate that RePatch achieves state-of-the-art performance on most benchmarks and consistently generalizes well across diverse transfer scenarios.