Forecast Combination for Tail Risk with Regulator-Aware Decision Trees

In this paper, we address the challenge of combining forecasts for Value-at-Risk (VaR) and Expected Shortfall (ES) by employing nonparametric weights which are functions of macro-financial state variables. We develop a tailor-made, tree-based machine learning procedure in which optimal combination weights remain constant within economic regimes but vary across them. The continuously varying regimes themselves are identified in a fully data-driven manner. To ensure consistency with supervisory practice, we align the tree construction process with liquidity-horizonspecific risk factors defined by the Basel Committee on Banking Supervision. We further reduce the influence of correlated covariates by proposing a methodology of cluster-wise random variable draws when building each tree. In an application to daily returns of 24 large-cap U.S. stocks from 2000 to 2023, the proposed marketdriven forecast combination delivers significantly lower out-of-sample VaR-ES losses than the standard relative-score benchmark for up to 22 out of 24 assets across multiple liquidity horizons, while also improving coverage properties and economic interpretability.