cs.AINov 18, 2026

A Verifier from the Past: Test-Time Compute for Time Series Foundation Models

Authors: Arjun Choudhry, Aditya Narayanan, Raghav Mallampalli, Po-han Li, Sai Shankar Narasimhan, Raimi Shah, B. Aditya Prakash, Shubhankar Agarwal

Organizations: Georgia Institute of Technology · Synthefy · The University of Texas at Austin

Abstract

Test-time compute (TTC) improves language models by spending additional inference compute to generate, verify, and select among candidate outputs, but this recipe has not transferred directly to time series forecasting. A language model can verify a proof or vote across samples, but a forecaster must be causal; it cannot use future information to score a candidate. We propose Backtest-Gated Augmentation Averaging (BGAA), a TTC algorithm for time series forecasting that builds a causal verifier from the series's own history, using rolling-origin backtests to score structured views of the context, discard those that underperform the untransformed context, and average the survivors' predictive distributions. BGAA improves every tested TSFM-benchmark combination, reducing MASE by 0.5-1.0% on GIFT-Eval, 1.3-5.0% on fev-bench, and 1.7-5.9% on BOOM, where a 313M model with TTC outperforms the 1B and 2.5B models in its family. We further show that useful views have structured, model-dependent effects consistent with a prior-transport interpretation. On the 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 and 17-19% when choosing the best view separately for each test window. This gap suggests that a major remaining challenge is window-level view selection.

Figures & tables

Appendix figures & tables35 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Apr 13, 2026

TempusBench: An Evaluation Framework for Time-Series Forecasting

Foundation models have transformed natural language processing and computer vision, and a rapidly growing literature on time-series foundation models (TSFMs) seeks to replicate this success in forecasting. While recent open-source models demonstrate the promise of TSFMs, the field lacks a comprehensive and community-accepted model evaluation framework. We see at least four major issues impeding progress on the development of such a framework. First, existing evaluation frameworks comprise benchmark forecasting tasks derived from often outdated datasets (e.g., M3), many of which lack clear metadata and overlap with the corpora used to pre-train TSFMs. Second, these frameworks evaluate models along a narrowly defined set of benchmark forecasting tasks, such as forecast horizon length or domain, but overlook core statistical properties such as non-stationarity and seasonality. Third, domain-specific models (e.g., XGBoost) are often compared unfairly, as existing frameworks do not enforce a systematic and consistent hyperparameter tuning convention for all models. Fourth, visualization tools for interpreting comparative performance are lacking. To address these issues, we introduce TempusBench, an open-source evaluation framework for TSFMs. TempusBench consists of 1) new datasets which are not included in existing TSFM pretraining corpora, 2) a set of novel benchmark tasks that go beyond existing ones, 3) a model evaluation pipeline with a standardized hyperparameter tuning protocol, and 4) a tensorboard-based visualization interface. We provide access to our code on GitHub: https://github.com/Smlcrm/TempusBench and maintain a live leaderboard at https://smlcrm.com/tempusbench.
Oct 21, 2026

Rethinking Agentic Model Orchestration for Time Series Forecasting

Time series foundation models (TSFMs) perform unevenly across forecasting tasks. Recent work uses agentic model orchestration (AMO) to select or combine TSFMs, often outperforming individual forecasters. Yet these gains may reflect stronger or more complementary candidates rather than better orchestration. To disentangle these effects, we introduce FairAMO, a controlled evaluation protocol with frozen candidate forecasts. Experiments on GIFT-Eval show that AMO improves accuracy, but its gains are bounded by the candidate TSFMs. The best orchestrators reduce MASE by less than 6% relative to the best single TSFM across all tasks. These gains depend on complementarity (different candidates excel on different tasks) rather than the number of candidates. Adding members yields diminishing returns, and a redundant member can erase these gains. The value of historical evidence also saturates quickly. A few recent windows capture all gains, and further history adds none. These findings separate system-level forecasting gains from the contribution of orchestration itself.
Oct 8, 2026

Training on the Future: A Delay-Aware Audit of Test-Time Adaptation for Time-Series Forecasting

Test-time adaptation (TTA) methods for time-series forecasting update a deployed model, or a small adapter around it, from incoming ground truth. But the label of an HH-step forecast exists only HH steps later, and real data pipelines add further delay. We build a leakage-free harness in which the label of forecast origin ss is released for updates only at step s+ds+d with d≥Hd \ge H, and enforce this rule inside the released code of four recent TTA methods (TAFAS, COSA, PETSA and DynaTTA), run on their own backbones and checkpoints across five benchmarks (ETTm1, ETTh2, Weather, Electricity and Traffic). As references we add two closed-form correctors: a bank of recursive least squares (RLS) filters combined by a per-coordinate median, with no tunable hyperparameters and 56 microseconds per step on the 7-channel streams, and an ELF-style linear corrector. Under causal delayed labels the picture is asymmetric. On ETTm1 every audited method genuinely adapts, yet the RLS bank still beats three of the four at a fraction of their cost; only DynaTTA beats the bank, only at the minimum causal delay, and at roughly 2,500 times the per-update cost; the ELF-style corrector beats all four. On the other four datasets, the largest statistically significant improvement any published method achieves over its own frozen checkpoint is half a percent, on all four at least one published method is significantly worse than the frozen model at the minimum causal delay, and on drift-heavy ETTh2 longer label delays make every adapter that separates from the frozen model, ours included, significantly harmful. Leaky next-step updates inflate the apparent gains of simple adapters by up to 110%, and the backbone training recipe moves frozen online error by up to a factor of 25, more than any adaptation effect we measure. We release the harness, integration patches and all cached runs.