quant research // summarizing the paperLOADING
Alessio Brini · 2026-07-06
The authors test whether nine pretrained "foundation" time-series AI models can forecast realized volatility better than established econometric models like the HAR family, across 50 assets in equities, FX, and futures at three horizons. They find foundation models don't deliver a broad advantage: only one small model (Tiny Time Mixers) narrowly beats a well-specified Log-HAR benchmark at every horizon, and much of its short-horizon edge is just better-scaled forecasts rather than genuinely better prediction of volatility dynamics.
Why it matters: Volatility forecasts feed into risk management, option pricing, and position sizing, so knowing whether new AI models beat classic econometric tools matters for anyone building such forecasts. The practical hint here is that simple, cheap benchmarks like Log-HAR remain competitive, and a plain equal-weight blend of the best AI model and Log-HAR performed robustly across assets without needing to pre-select the best model.
⚠ This is a forecasting-accuracy study on one dataset using zero-shot models, not a live-trading or P&L evaluation, and the measured edge is described as thin.
We ask whether pretrained time series foundation models (TSFMs) improve on established econometric benchmarks for forecasting realized volatility. Using the VOLARE dataset, we conduct the first systematic comparison of nine zero-shot TSFMs against eight econometric specifications, including the Heterogeneous Autoregressive (HAR) family, across 50 assets in equities, foreign exchange, and futures, and three forecast horizons, with formal pairwise and multi-model forecast-comparison tests. Foundation models do not deliver a uniform gain. Pooled losses favor them, but the advantage is concentrated in a few outlier assets; averaging each asset's loss ratio to a well-specified Log-HAR benchmark, so that no single asset dominates, only one small model, Tiny Time Mixers (TTM), beats the benchmark at every horizon, and by a narrow margin. The other foundation models do not improve on Log-HAR, and the econometric benchmarks remain competitive throughout. A Mincer--Zarnowitz recalibration, which removes level and scale bias from every forecast, shows that much of the short-horizon advantage reflects better-scaled forecasts rather than better prediction of volatility dynamics, and only at the monthly horizon does a genuine informational gain remain. Because this edge is thin and even TTM is not best on every asset, a simple equal-weight average of TTM and Log-HAR matches the best single model and enters the Model Confidence Set for 98 to 100\% of assets, more often than either component alone, so a forecaster need not identify the best model for each asset in advance. Our most durable finding is that performance varies so much across foundation-model architectures that choosing the right architecture matters more than the broader choice between foundation and econometric models.
Go deeper: a full research-committee breakdown of this paper, its assumptions and failure modes, and how its method would apply to a specific ticker or your watchlist. See StockTools AI →
AI summary generated from the paper’s public abstract via arXiv; it may miss nuance — read the source before relying on it. Thank you to arXiv for its open-access interoperability; StockTools is not affiliated with arXiv, and all rights remain with the authors. Educational only, not financial advice.