This paper asks whether large pretrained time‑series models can improve forecasts of realized volatility. The short answer is: not reliably. Across 50 assets and three horizons, most pretrained models did not outperform well‑tuned econometric benchmarks. One small foundation model called Tiny Time Mixers (TTM) beat a strong log‑HAR benchmark at every horizon, but only by a narrow margin.

Realized volatility is a common, data‑based measure of how much an asset’s price moves. Accurate forecasts of realized volatility matter for risk management, option pricing, and portfolio decisions. The Heterogeneous Autoregressive (HAR) model and related econometric methods are long‑standing benchmarks because they capture volatility’s slow decay and persistence with a simple structure.

The researchers ran a systematic, zero‑shot comparison. Zero‑shot means the pretrained time‑series foundation models were used as‑is, with no fine‑tuning on the target financial series. They used the VOLARE dataset and forecasted daily realized volatility for 50 assets across equities, foreign exchange, and futures. They evaluated nine foundation models (covering eight different architectures) against eight econometric specifications — including HAR, Log‑HAR, ARFIMA, ARMA, and multiplicative error models — at three horizons: 1, 5, and 22 trading days. Formal forecast tests were applied to control for chance differences and time variation.

On pooled average losses measured with a common volatility loss (QLIKE), some foundation models looked better. But that advantage was driven by a few outlier assets. When the authors averaged performance equally across assets so that no single series could dominate, only TTM beat the Log‑HAR benchmark at all horizons, and its edge was small (roughly 1.3–1.8%). Most other foundation models failed to improve on Log‑HAR. The classic econometric forecasts (Log‑HAR, HAR, ARFIMA, ARMA, MEM) stayed competitive and often ranked within a few percent of each other. More complex HAR variants that add jump or quarticity corrections performed worse in this zero‑shot comparison.