Journal of Applied Finance & Banking

Regime-Gated Tail Transfer: Reliable Cross-Industry Machine Learning for Value-at-Risk and Expected Shortfall

• [ Download ]
• Times downloaded: 54

• Abstract

   

  Pooling financial series gives machine-learning tail models more extreme observations, but it can also transfer the wrong tail dynamics across assets and regimes. Regime-gated tail transfer (RGTT) combines Student-t GARCH, filtered historical simulation, industry-specific quantile boosting, and pooled cross-industry quantile boosting. FZ0-scored weights depend on nine lagged volatility–breadth states, with shrinkage toward global performance. A coverage controller (RGTT-C) uses past exceedances and preserves ES ≤ VaR. Hyperparameters are selected on 1995–1999 and frozen before a 2000–July 2026 test across ten U.S. industry portfolios (66,840 industry-days). RGTT significantly improves on equal weighting and a state-agnostic gate. RGTT-C attains mean FZ0 −3.4887 and a 2.491% breach rate at the 2.5% target, with significant loss advantages over filtered historical simulation and both quantile-ML experts. Its advantage over Student-t GARCH is not significant at 5% (two-sided p = 0.072). The pooled expert’s mean weight falls from 22.5% to 13.5% between low-volatility/low-breadth and high-volatility/high-breadth states. The framework makes cross-industry transfer state dependent while separating expert weighting from an order-preserving coverage controller.

   

  JEL classification numbers: C22, C45, C53, G17, G32.

  Keywords: Expected shortfall, Value-at-risk, Forecast combination, Quantile machine learning; negative transfer, Regime adaptation, Market risk.