Answers are grounded in this paper and cite their sources.
The risk of downside of the portfolio depends on volatility, drawdown, asset dependence, and regime in a nonlinear way which undermines the stability of risk measurement with fixed distribution. In this paper, we propose a machine-learning based adaptive downside risk measure that utilizes LightGBM quantile regression to predict the conditional tail of the portfolio made up of equally-weighted eight ETFs. The data is composed of daily SPY, IWM, EFA, EEM, TLT, GLD, DBC, and VNQ during the period of 2010 to 2025, and features are constructed by rolling window volatility, semivariance, drawdown, dependence, and market state. Conditional quantiles between 0.25% and 2.50% are forecast and used to construct one-day ahead 97.5% VaR and Expected Shortfall. The adaptive LightGBM in the out-of-sample test produces a pinball loss of 0.0187±0.0011, a VaR violation ratio of 1.04±0.08, and an ES backtest score of 0.021±0.006. Its joint tail loss score significantly outperforms Historical Simulation, Gaussian GARCH, and linear Quantile Regression. When the regime is volatile, the forecast improvement is still significant. Feature ablation suggests that downside-history and asset dependence variables contribute most.
No takes yet. Share an insight, caveat, or question.
Ying Hui (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: