The substantial fluctuations in oil prices in the wake of the COVID-19 pandemic and the Russian invasion of Ukraine have highlighted the importance of tail events in the global market for crude oil which call for careful risk assessment. In this paper we focus on forecasting tail risks in the oil market by setting up a general empirical framework that allows for flexible predictive distributions of oil prices that can depart from normality. This model, based on Bayesian additive regression trees, remains agnostic on the functional form of the conditional mean relations and assumes that the shocks are driven by a stochastic volatility model. We show that our nonparametric approach improves in terms of tail forecasts upon three competing models: quantile regressions commonly used for studying tail events, the Bayesian VAR with stochastic volatility, and the simple random walk. We illustrate the practical relevance of our new approach by tracking the evolution of predictive densities during three recent economic and geopolitical crisis episodes, by developing consumer and producer distress indices that signal the build-up of upside and downside price risk, and by conducting a risk scenario analysis for 2024.

Early-warning models most commonly optimize signaling thresholds on crisis probabilities. The expost threshold optimization is based upon a loss function accounting for preferences between forecast errors, but comes with two crucial drawbacks: unstable thresholds in recursive estimations and an in-sample overfit at the expense of out-of-sample performance. We propose two alternatives for threshold setting: (i) including preferences in the estimation itself and (ii) setting thresholds ex-ante according to preferences only. Given probabilistic model output, it is intuitive that a decision rule is independent of the data or model specification, as thresholds on probabilities represent a willingness to issue a false alarm vis-à-vis missing a crisis. We provide simulated and real-world evidence that this simplification results in stable thresholds and improves out-of-sample performance. Our solution is not restricted to binary-choice models, but directly transferable to the signaling approach and all probabilistic early-warning models.

Early-warning models most commonly optimize signaling thresholds on crisis probabilities. The ex-post threshold optimization is based upon a loss function accounting for preferences between forecast errors, but comes with two crucial drawbacks: unstable thresholds in recursive estimations and an in-sample overfit at the expense of out-of-sample performance. We propose two alternatives for threshold setting: (i) including preferences in the estimation itself and (ii) setting thresholds ex-ante according to preferences only. We provide simulated and real-world evidence that this simplification results in stable thresholds and improves out-of-sample performance. Our solution is not restricted to binary-choice models, but directly transferable to the signaling approach and all probabilistic early-warning models.