A preprint posted on arXiv on 9 October by a researcher at Polytechnique Montréal examined whether time-series forecasting needs the same computation depth for every input. The study evaluated structured masks that bypass Transformer blocks at inference in the pretrained Chronos-Bolt and TimesFM 2.5 models. Masks were chosen only on calibration windows, then tested on 1,088 held-out windows from the ETTh1, Traffic, Weather and Exchange datasets.

The author reports that, at moderate skipping levels, running seven of Chronos-Bolt’s 11 candidate blocks and 16 of TimesFM’s 20 blocks changed mean relative forecasting quality little across the datasets, while reducing median batch-one A100 latency by 11.2% and 13.4%, respectively. More aggressive skipping substantially harmed Traffic and Weather forecasts. The best block mask varied by window, but the study explicitly does not provide a deployable method for reliably choosing one for each unseen input. The findings concern two models, four datasets and the tested hardware setup; the work is a preprint that has not yet undergone peer review.