Deep g-Pricing for CSI 300 Index Options with Volatility Trajectories and Market Sentiment

Battle Count: 69

Directly applicable to derivatives pricing, hedging, and risk management in quantitative trading. The method provides explicit hedging ratios via automatic differentiation, reduces pricing errors by 32-35% vs BSM, and offers interpretable insights into call/put asymmetry. The sentiment gating configuration provides actionable guidance for model deployment. Particularly relevant for options market-making, volatility trading strategies, and margin risk management. The extreme error reduction (MAPE >100% from 11.51% to 0.77% for OTM) has direct implications for counterparty risk.

High complexity due to: (1) dual-network architecture with shared backbone and dual-head outputs; (2) FBSDE forward recursion with Monte Carlo simulation; (3) automatic differentiation for hedging strategy; (4) two-stage training (pre-training + fine-tuning) with curriculum learning; (5) adaptive activation functions with L2-normalized weights; (6) sentiment gating mechanism with differentiated configurations by moneyness; (7) composite loss function with three components; (8) dependency on external RV forecasting model (TimeXer + CNN + LLM pipeline) and sentiment classification pipeline (BERT + multiple LLMs with voting). Requires careful hyperparameter tuning and GPU resources for training.

Code is available from the corresponding author upon reasonable request. Data availability is restricted due to licensing agreements with data providers. The paper provides detailed algorithm pseudocode (Algorithm 1), hyperparameter configurations (Table 4), and network architecture descriptions, but no public repository is linked. The RV forecasting model references prior team work (Zhang et al., 2025) which is a working paper.