Tail risk forecasting in the U.S. Treasury futures market

U.S. Treasury futures markets are central to global fixed-income price discovery and risk management, making accurate tail-risk forecasting essential for market participants and regulators. We investigate whether high-frequency intraday data can improve Value-at-Risk forecasts in these markets, comparing a broad set of approaches including quantile regression, GARCH, HAR, and CAViaR specifications, as well as forecast combinations. Relative to a Historical Simulation benchmark, models built on realized volatility measures and simple ensemble averaging reduce the average quantile loss by approximately 6% to 12%, with the largest gains at the most extreme quantiles. An equally weighted forecast combination ranks first and the HAR model second across four quantile levels and five evaluation criteria, and both maintain their advantage during tranquil periods and major crisis episodes alike. These findings have important implications for risk management, margin requirements, and regulatory capital calculations in Treasury futures markets. • High-frequency realized volatility substantially improves Value-at-Risk forecasts in U.S. Treasury futures. • An equally weighted forecast combination ranks first and the HAR model second across four quantile levels and five evaluation criteria. • Reductions in quantile loss relative to Historical Simulation range from 6% to 12%, with the largest gains at the most extreme quantiles. • Gains come from simple averaging: an optimization-based combination is markedly less stable across regimes. • Superior performance holds symmetrically across both tails and across crisis regimes, including the GFC, COVID-19, and the 2022–2023 inflation episode.