PatchDecomp: Interpretable Patch-Based Time Series Forecasting

Authors: Hiroki Tomioka, Genta Yoshimura Date: 2026-03-04 Paper ID: openalex:2603.03902

Summary

PatchDecomp is a novel neural network architecture for time series forecasting designed to address the trade-off between high accuracy and model interpretability. The core methodology involves segmenting the input time series, including any associated exogenous variables, into discrete subsequences called patches. Predictions are then generated by an aggregation mechanism that explicitly calculates and combines the specific contribution of each individual patch to the final output. Experimental results confirm that PatchDecomp achieves competitive accuracy against leading forecasting models while simultaneously offering clear, visualizable explanations for its predictions based on patch-wise contributions.

Key Contributions

• Introduction of PatchDecomp, a forecasting method that achieves high accuracy through a patch-based decomposition of the input time series.
• Demonstration that PatchDecomp’s predictive performance is comparable to state-of-the-art forecasting methods on multiple benchmarks.
• Development of an interpretable framework where the contribution of each input patch, including exogenous variables, can be clearly attributed to the final forecast.
• Visualization of patch-wise contributions, offering qualitative interpretability alongside quantitative prediction accuracy.

Limitations

The abstract primarily focuses on performance and interpretability without detailing specific limitations regarding forecasting horizon or handling of complex non-linear dynamics relative to other complex models.

Open Questions & Future Work

• static-exogenous-variable-contribution-analysis

Key Concepts

• PatchDecomp: A neural network-based time series forecasting method that achieves high accuracy and interpretability by dividing the input series into patches and aggregating their contributions for prediction.

Datasets

• benchmark datasets

Limitations

The abstract primarily focuses on performance and interpretability without detailing specific limitations regarding forecasting horizon or handling of complex non-linear dynamics relative to other complex models.

Links

• Abstract
• PDF