How can we design time-series representations that better capture the underlying dynamics essential for accurate forecasting? Long-term time-series forecasting has recently benefited from patch-based representation learning, where input sequences are ...
How can we design time-series representations that better capture the underlying dynamics essential for accurate forecasting? Long-term time-series forecasting has recently benefited from patch-based representation learning, where input sequences are divided into fixed-length patches and processed by Transformer-style architectures. However, fixed segmentation often fails to reflect the heterogeneous temporal dynamics of real-world signals. Adjacent patches with highly similar behaviors may be arbitrarily separated, while structurally different patterns may be forced into the same patch, disrupting temporal continuity and limiting representation quality. Furthermore, value-level similarity is insufficient to capture coarse structural cues such as periodicity or shape that are more evident in the frequency domain.
In this paper, we propose STITCH, a lightweight refinement framework that adaptively groups semantically coherent patches before they are encoded by downstream forecasting models. STITCH combines time-domain embeddings with i frequency-domain features derived from the Fast Fourier Transform to compute a unified similarity measure between adjacent patches. Leveraging this similarity, STITCH forms variable-length, continuity-preserving segments without introducing any additional learnable parameters, resulting in more stable and expressive patch representations.
Extensive experiments across multiple long-term forecasting benchmarks demonstrate that STITCH consistently improves prediction performance over strong existing baselines in most forecasting horizons. These results show that integrating time–frequency similarity into patch clustering provides a robust enhancement to modern time-series forecasting pipelines