AI
A Verifier from the Past: Test-Time Compute for Time Series Foundation Models
Publication: cs.AI updates on arXiv.org
Author: Arjun Choudhry, Aditya Narayanan, Raghav Mallampalli, Po-han Li, Sai Shankar Narasimhan, Raimi Shah, B. Aditya Prakash, Shubhankar Agarwal
Source published: 20/11/2026, 05:00:00 UTC
Briefing created: 20/11/2026, 05:50:43 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 20/11/2026, 05:03:24 UTC.
Briefings
A new algorithm named Backtest-Gated Augmentation Averaging (BGAA) improves time series forecasting by using the series's own history to verify candidate outputs. The study tested 53 cross-fittable configurations on GIFT-Eval, showing that TTC reduces Mean Absolute Scaled Error (MASE) by 0.5-1.0% on GIFT-Eval, 1.3-5.0% on fev-bench, and 1.7-5.9% on BOOM. However, the deployed policy only gains 0.4-1.3% CRPS compared to selecting the best view per dataset, indicating a major challenge in window-level view selection.
Key points
- Backtest-Gated Augmentation Averaging (BGAA) uses rolling-origin backtests to score structured views of context, discarding underperforming views and averaging predictive distributions.
- On 53 cross-fittable GIFT-Eval configs, the deployed policy gains only 0.4-1.3% CRPS versus 1.4-4.1% when choosing the best view for each dataset configuration.
Context
This research addresses the gap between language model verification techniques and time series forecasting, where causal constraints prevent future information use. The findings suggest that while TTC improves forecasting, significant challenges remain in selecting optimal views for different time windows.
Limitations
The study covers 53 cross-fittable GIFT-Eval configurations but does not report results for all tested TSFM-benchmark combinations. The deployed policy gains are limited to 0.4-1.3% CRPS, suggesting a major remaining challenge in window-level view selection.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 20/11/2026, 05:50:43 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiggivveppcfpjupvgk5mcbclgu7tubd4unxy3xfdmjvl2remgwjjq