Backtest-gated verifier brings test-time compute to time series forecasting models
Researchers propose BGAA, an inference-time method that scores candidate forecast contexts against a series's own history and averages the survivors, reporting gains across three benchmarks.
In plain words
Test-time compute means spending extra effort when a model answers, by generating several candidates and picking the best. The authors adapt this idea to forecasting, where the model cannot check answers against the future. Their method, BGAA, tests candidate inputs against past data and averages the forecasts that pass.
What was announced
A paper on arXiv introduces Backtest-Gated Augmentation Averaging (BGAA), a test-time compute algorithm for time series forecasting. The authors note that the recipe of generating, verifying and selecting among candidate outputs, which works for language models, has not transferred directly to forecasting, because a forecaster must be causal and cannot use future information to score a candidate. BGAA instead builds a verifier from the series's own history, running rolling-origin backtests to score structured views of the context, discarding views that underperform the untransformed context, and averaging the predictive distributions of the survivors.
According to the abstract, BGAA improved every tested time series foundation model and benchmark combination, cutting MASE by 0.5-1.0% on GIFT-Eval, 1.3-5.0% on fev-bench and 1.7-5.9% on BOOM; a 313M model with the method outperformed the 1B and 2.5B models in its family. The authors also report a gap: on 53 cross-fittable GIFT-Eval configurations, the deployed policy gained only 0.4-1.3% CRPS, versus 1.4-4.1% when choosing the best view per configuration and 17-19% per test window, suggesting window-level view selection remains an open challenge.
What it does not show
- The evidence is the paper's abstract only; the full methodology, experiments and peer review status are not available in the supplied passages.
- All reported improvements and comparisons come from the authors' own evaluation and have not been independently confirmed.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.