BGAA Builds a Causal Verifier for Test-Time Compute in Time-Series Forecasting
Summary:BGAA brings test-time compute to time-series foundation models without asking a future-aware verifier to judge forecasts. It creates structured views of the historical context, scores them through rolling-origin backtests, discards views that do not beat the untransformed context, and averages the surviving predictive distributions. Across every tested model–benchmark combination, the paper reports MASE reductions of 0.5%–1.0% on GIFT-Eval, 1.3%–5.0% on fev-bench, and 1.7%–5.9% on BOOM. A 313M-
# BGAA Builds a Causal Verifier for Test-Time Compute in Time-Series Forecasting BGAA adapts a familiar test-time compute pattern to time-series forecasting without allowing a future-aware verifier to score candidate forecasts. Instead, it uses the series’ own history, evaluates structured views through rolling-origin backtests, retains only views that consistently beat the untransformed context, and averages their predictive distributions. The paper reports MASE improvements across every tested benchmark–model combination, but a large gap remains between the deployable policy and per-window oracle selection. ## Core Mechanism: Building a Verifier from History ### 1. Candidate context views Let the original historical context be `x_{ **Source**: [ArXiv AI (cs.AI)](https://arxiv.org/abs/2611.25171) (2026-11-20) — END —
Tags: #Time-Series Foundation Models #Test-Time Compute #Backtest Gating #Predictive Distribution Averaging #Causal Verification
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments