TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting

Authors: Sravan Kumar Ankireddy, Nikita Seleznev, Nam H. Nguyen, Yulun Wu, Senthil Kumar, Furong Huang, C. Bayan Bruss Date: 2026-03-11 Paper ID: openalex:2603.11352

Summary

This paper introduces TimeSqueeze, a novel dynamic patching mechanism designed to overcome the tokenization trade-off in Transformer-based time series models, where point-wise embeddings offer fidelity but poor efficiency, and fixed patching sacrifices temporal structure. TimeSqueeze first extracts full-resolution features using a lightweight state-space encoder, then segments the sequence adaptively, creating shorter patches for high-complexity regions and longer patches for smooth segments. This content-aware compression drastically reduces the sequence length fed to the Transformer backbone, leading to up to 20x faster convergence and 8x better data efficiency in pre-training. Empirical results on long-horizon forecasting benchmarks confirm that TimeSqueeze consistently yields superior performance compared to fixed-patching and point-wise baselines.

Key Contributions

• Introduction of TimeSqueeze, a dynamic patching mechanism that adaptively selects patch boundaries based on local signal complexity, resolving the fidelity vs. efficiency trade-off in Transformer tokenization for time series.
• The mechanism uses a lightweight state-space encoder for feature extraction followed by content-aware segmentation, allocating shorter patches to information-dense regions and longer patches to smooth segments.
• Achieved up to 20x faster convergence and 8x higher data efficiency during large-scale pre-training compared to standard point-token baselines.
• Demonstrated consistent outperformance over fixed-size patching and point-wise tokenization methods across various long-horizon forecasting tasks.

Limitations

The paper focuses primarily on pre-training efficiency and long-horizon forecasting; the performance impact on very short sequence tasks or specific anomaly detection scenarios is not detailed.

Open Questions & Future Work

• learning-end-to-end-dynamic-patching-rate
• variable-rate-patching-large-scale-forecasting
• generalizing-dynamic-patching-to-other-backbones
• probabilistic-forecasting-with-dynamic-patching

Key Concepts

• TimeSqueeze Dynamic Patching: A dynamic patching mechanism for time series forecasting that adaptively selects patch boundaries based on local signal complexity to optimize tokenization efficiency for Transformer models.

Datasets

• long-horizon forecasting benchmarks

Limitations

The paper focuses primarily on pre-training efficiency and long-horizon forecasting; the performance impact on very short sequence tasks or specific anomaly detection scenarios is not detailed.

Links

• Abstract
• PDF