Multivariate time series forecasting remains difficult because temporal dynamics are non-stationary, structures evolve over time, and models often generalize poorly across scenarios. Patch-based approaches strengthen long-range modeling by dividing a sequence into local units, but fixed or weakly adaptive patching and continuous latent representations can restrict flexibility and pattern reuse.

RePatch addresses these limitations with a two-stage framework. During self-supervised pretraining, its Dynamic Patcher uses an entropy-based objective to divide each series into variable-length patches aligned with local temporal dynamics. A Temporal Interaction Module then models both local and global dependencies before vector quantization maps the representations into a structured discrete latent space.

The resulting discrete representations provide a compact and transferable abstraction of temporal patterns and naturally support token-based sequence modeling for downstream forecasting. Experiments on 14 real-world datasets show state-of-the-art results on most benchmarks and consistent generalization across a range of transfer scenarios.