A time-series foundation model called SATS posted the best reported average performance among compared models on the LSTF zero-shot benchmark, according to an arXiv preprint. Its base version, SATS-B, was reported to improve mean squared error (MSE) by 9.2% versus Moirai-B, 4.2% versus Timer-XL and 27.4% versus Chronos-L. LSTF covered five datasets excluded from SATS’s LOTSA pretraining corpus, with forecast horizons of 96, 192, 336 and 720 steps.
A model built for mixed time scales
The preprint addresses the challenge of pretraining across time-series datasets with different sampling frequencies. SATS breaks each series into non-overlapping patches—chunks of consecutive observations—whose sizes depend on the dataset. Its encoder-only Transformer uses separate feed-forward networks for different scales and aligns mean- and max-pooled representations across resolutions.
Training adds hybrid masking: random patches and contiguous spans are hidden during pretraining, a setup intended to represent both fine-grained patterns and long-range dependencies. The total objective combines the masked prediction loss with the scale-aware alignment loss without adding new learnable parameters.
SATS was pretrained on LOTSA in small and base configurations. The small model used 100,000 training steps with a batch size of 256; the base model used 200,000 steps with a batch size of 128.
Gains across three benchmark suites
On strict zero-shot GIFT-Eval, SATS-B was reported to gain 8.3% in MASE and 8.2% in CRPS against Moirai-L. The evaluation covered 23 datasets; all GIFT-Eval overlaps were removed from LOTSA before SATS was retrained for the test.
SATS-B also led Chronos-BB in CRPS while requiring 35% of that model’s parameters.
The pattern extended to the 29 Monash datasets used for in-distribution forecasting. SATS-B was reported to improve normalized mean absolute error by 6.9% versus Moirai-L while using 22.6% of its parameters. The smaller SATS-S also outperformed Chronos-S using 30.4% of its parameter count.
A separate efficiency score based on zero-shot error and the logarithm of model size favored SATS in the LSTF analysis. SATS-B exceeded Timer-XL by 8.9%, while SATS-S showed a 65.6% efficiency improvement. SATS-B has 70 million parameters and SATS-S has 14 million.
What the component tests found
Comparisons that removed individual components showed lower metrics without SA, the scale-aware alignment module. Removing either contiguous masking or random masking produced larger drops in most ETT and Weather cases, although the reported roles differed by evaluation regime: hybrid masking was associated with long-horizon learning and SA with cross-dataset generalization.
A representation diagnostic measured cosine similarity across resolutions. With SA, similarity ranged from 0.979 to 0.998 for mean-pooled representations and from 0.424 to 0.639 for max-pooled representations. Without SA, the corresponding ranges were -0.054 to 0.293 and 0.721 to 0.889.
Patch choice also mattered in a separate LSTF sensitivity analysis: the prescribed protocol performed best, smaller patches were associated with substantial degradation, and larger patches with only mild loss.
Why the results need caution
The findings come from an arXiv preprint. The supplied analysis says that confidence intervals, significance tests and variability across repeated runs were not reported, so the document does not quantify uncertainty around the percentage differences.
The efficiency measure uses parameter count as a proxy for model capacity. It therefore does not by itself show lower latency, memory use, energy use or deployment cost, and the reported tests do not establish performance beyond the LSTF, GIFT-Eval and Monash benchmarks.
Paper data and sources
Original title: Scale-Aware Pretraining of Time Series Foundation Models via Multi-Patch Token Alignment and Hybrid Masking
Authors: Taihua Chen, Xiang Ma, Yixin Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text