![[2602.16468] HPMixer: Hierarchical Patching for Multivariate Time Series Forecasting](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16468] HPMixer: Hierarchical Patching for Multivariate Time Series Forecasting
Summary
The paper presents HPMixer, a novel approach for multivariate time series forecasting that effectively models periodic patterns and residual dynamics through hierarchical patching.
Why It Matters
As multivariate time series data becomes increasingly prevalent in various fields, improving forecasting accuracy is crucial. HPMixer's innovative approach addresses both periodicity and residuals, offering a competitive edge in predictive performance, which is vital for industries relying on accurate forecasting.
Key Takeaways
- HPMixer integrates decoupled periodicity modeling with multi-scale residual learning.
- The model utilizes a Learnable Stationary Wavelet Transform for stable frequency-domain representations.
- Extensive experiments show HPMixer achieves state-of-the-art performance on standard benchmarks.
Computer Science > Machine Learning arXiv:2602.16468 (cs) [Submitted on 18 Feb 2026] Title:HPMixer: Hierarchical Patching for Multivariate Time Series Forecasting Authors:Jung Min Choi, Vijaya Krishna Yalavarthi, Lars Schmidt-Thieme View a PDF of the paper titled HPMixer: Hierarchical Patching for Multivariate Time Series Forecasting, by Jung Min Choi and 2 other authors View PDF HTML (experimental) Abstract:In long-term multivariate time series forecasting, effectively capturing both periodic patterns and residual dynamics is essential. To address this within standard deep learning benchmark settings, we propose the Hierarchical Patching Mixer (HPMixer), which models periodicity and residuals in a decoupled yet complementary manner. The periodic component utilizes a learnable cycle module [7] enhanced with a nonlinear channel-wise MLP for greater expressiveness. The residual component is processed through a Learnable Stationary Wavelet Transform (LSWT) to extract stable, shift-invariant frequency-domain representations. Subsequently, a channel-mixing encoder models explicit inter-channel dependencies, while a two-level non-overlapping hierarchical patching mechanism captures coarse- and fine-scale residual variations. By integrating decoupled periodicity modeling with structured, multi-scale residual learning, HPMixer provides an effective framework. Extensive experiments on standard multivariate benchmarks demonstrate that HPMixer achieves competitive or state-of-the-art...
