Regime-Gated Residual Mixture-of-Experts for Cross-Sectional Volatility Forecasting

Battle Count: 169

Highly relevant for quantitative trading. The paper directly addresses cross-sectional volatility forecasting, which is critical for options pricing, risk management, portfolio allocation, and trading strategy calibration. The VaR calibration improvements and the finding that gains concentrate during high-volatility periods (4.3x-6.7x amplification) are directly actionable for risk managers and systematic traders. The design principle (regime info for routing, not input) is broadly applicable to any regime-aware trading model.

Moderate complexity. The architecture is built from standard MLP blocks with a softmax gate, making it implementable in PyTorch/TensorFlow. Key components: frozen base network, zero-initialized residual experts, soft gating with regime variables, two-stage training, and three-part loss (MSE + shrinkage + load-balancing). The rolling walk-forward evaluation with 30 seeds and hyperparameter sweeps adds computational overhead. No exotic operations or custom CUDA kernels required.

The paper provides detailed architecture specifications, hyperparameter selection criteria, training procedures (Adam, early stopping, 30 seeds), evaluation protocol (rolling walk-forward with 30 windows), and statistical testing methodology (Diebold-Mariano with Newey-West). Data source (Yahoo Finance) is publicly available. However, no code repository is mentioned, and exact hyperparameter values for all sweeps are not fully enumerated.