Risk-Sensitive Specialist Routing for Volatility Forecasting

Battle Count: 80

Highly relevant to quantitative trading. Volatility forecasting is a core input for risk management, position sizing, portfolio allocation, and derivative pricing. The regime-dependent model selection insight (GRU dominates in calm markets, GARCH-t/FIGARCH in stressed markets) is directly actionable for trading desks. The 24% reduction in high-volatility forecast loss and 22% reduction in underprediction loss are practically significant for tail-risk management. However, the paper focuses on forecast accuracy rather than trading P&L, and the framework does not directly produce trading signals.

Moderately high complexity. The framework involves four stages with multiple interacting components: online scoring with regime-similarity weighting, threshold-based routing with regime-local/global quantile blending, two specialist branches with different aggregation rules (median vs. 75th percentile with winsorization), a logistic stress-score gate, a second blending gate, and a conditional HAR floor. Hyperparameter tuning across lambda_under, gamma_time, gamma_reg, rho, kappa, c, b, and floor thresholds adds practical complexity. However, the architecture is interpretable and uses pre-specified pools rather than learned gates, reducing some ML engineering burden.

Code and full hyperparameters are publicly available on GitHub. The methodology is clearly specified with explicit formulas, fixed hyperparameters (lambda_under=1, 63-day half-life, gamma_reg=1.5, rho=0.25, kappa=0.35, c=0.55, b=0.12, GRU 20-step/16 hidden units, XGBoost 250 trees/depth 3/lr 0.05), and a detailed walk-forward protocol. However, the exact macro-financial data source and preprocessing pipeline are not fully specified, and the Garman-Klass proxy construction details could be more explicit.