Patch size influences time series forecasting accuracy and efficiency
Investigating the effects of the patch size in Time Series Foundation Models
Machine LearningArtificial Intelligence
Summary
Choosing the size of patches—chunks of time steps—in time series models affects how well these models predict future data and how much computing power they need. The authors compared two model types and found that smaller patches help one model type work better, while the other performs best with medium-sized patches, especially depending on how far ahead predictions need to go. They also showed that longer context helps some models, but with diminishing returns. Depending on the prediction length and computing limits, different patch sizes and models are better choices.
What this means in practice
- •For data engineers: Optimize time series forecasting deployment by selecting patch sizes and model architectures based on forecast length and resource constraints.
- •For financial analysts: Improve forecasting accuracy for financial data by configuring model patch size according to prediction horizons and architecture choice.
Authors
Jad Yehya, Andrei-Tiberiu Pantea, Aymeric Dieuleveut, Matthieu Kowalski, Thomas Moreau
Abstract
Patch size is a central design choice in patch-based time-series foundation models, controlling temporal resolution, sequence length, and computational cost. We study how this choice interacts with transformer architecture through a controlled comparison of encoder-only and decoder-only forecasting models trained with the same data mixture and optimization budget. Across 97 GIFT-Eval tasks, encoder-only models consistently benefit from smaller patches, whereas decoder-only performance follows a U-shaped trend, with an intermediate patch size performing best in aggregate. Dataset-level results further show that the preferred decoder patch size increases with the forecasting horizon. For encoder-only models, increasing context length improves performance across patch sizes, although the gains diminish beyond a context length of 2048. Finally, training-compute and CPU-latency Pareto analyses reveal complementary regimes: decoder-only models provide efficient low-cost alternatives, while encoder-only models attain the highest overall accuracy. These results show that the choice of patch size and architecture should be made depending on the forecasting horizon and deployment constraints.