Climate policy uncertainty, geopolitical risk, and the distribution of crude oil prices: a mixed-frequency interval forecasting approach

We ask whether climate policy uncertainty and geopolitical risk improve forecasts of the entire distribution of crude oil prices — not only its centre, but also the tails in which risk materializes. Using daily Brent data for 1990–2026, we combine the mixed-frequency GARCH-MIDAS model, quantile regression, a recurrent neural network trained with the quantile loss, and conformal calibration of prediction intervals; forecasts are evaluated out of sample on 1,659 days covering the COVID-19 pandemic and Russia's invasion of Ukraine. Although both indexes carry the expected signs in sample, out of sample they improve neither the centre, nor the tails, nor volatility, and the only statistically significant difference runs against them. By contrast, conformally calibrated quantile regression attains empirical coverage of 0.90 at a 0.90 target and passes the 5% Value-at-Risk backtest with a violation share of exactly 0.050, whereas the more flexible uncalibrated network fails the same test. We conclude that the information carried by text-based uncertainty indexes is embedded in the oil price earlier and more precisely, and that the usability of probabilistic forecasts is decided by calibration rather than by enlarging the predictor set. The conclusions survive replacing Brent with WTI, an alternative climate policy uncertainty index, a shorter long-run memory, and the exclusion of 2020.