Time series models save computation by skipping some steps during forecasting

How Much Computation Does Every Forecast Need? Structured Block Skipping in Two Time-Series Foundation Models

Machine Learning

Summary

Some computer programs that predict future data points, called time-series models, usually work the same way no matter how hard the prediction is. This paper tests whether parts of these programs can be skipped during the prediction without making the results worse. The researchers found that skipping certain parts in two popular models saves computing time with little loss in accuracy on many examples. However, the skipping plans that work well for some datasets do not always work for others, and the models do not yet support changing skipping on new data reliably. This means skipping can reduce work, but it’s not yet ready for practical use in every situation.

What this means in practice

  • For time series engineers: Improve inference speed of time-series forecasting models by skipping computation blocks without losing much accuracy.
  • For cloud service operators: Reduce processing latency and resource use in deployed forecasting services by selectively skipping transformer blocks during inference.

Authors

Max Bourgeat

Abstract

Time-series foundation models use the same network depth for every input, although forecasting windows differ in structure and difficulty. We test whether whole Transformer blocks can be removed at inference time in two pretrained models: Chronos-Bolt and TimesFM 2.5. Our bypasses avoid the original block computation and preserve the dense model exactly when all blocks are enabled. Masks are selected only on calibration windows and evaluated on 1,088 held-out windows from ETTh1, Traffic, Weather, and Exchange. At moderate sparsity, Chronos-Bolt with 7/11 candidate blocks and TimesFM with 16/20 blocks show little change in mean relative forecasting quality across datasets while reducing median A100 batch-1 latency by 11.2% and 13.4%. More aggressive TimesFM skipping accelerates inference further but substantially harms Traffic and Weather. Across both models, different equal-budget masks win on different windows, and a retrospective oracle significantly improves over the best static mask on every dataset. However, dataset-specific masks transfer inconsistently, and Chronos representations do not support reliable unseen-domain routing. These results establish cross-architecture structured redundancy and input-dependent opportunity, but not a deployable dynamic inference method.
View PDFOpen arXiv