Abstract:

With the rapid expansion of China’s options market， options have become an increasingly important instrument for risk management in commodity and financial markets. Due to the asymmetric payoff structure of options， sellers face substantial downside risk， making effective hedging strategies essential. Traditional option hedging methods are typically based on mathematical pricing models， most notably the Black–Scholes–Merton （BSM） framework， where the optimal hedge ratio is defined as the option Delta. However， these approaches rely on restrictive assumptions such as constant volatility， continuous price dynamics， and frictionless markets. In practice， transaction costs and market frictions significantly reduce the effectiveness of frequent Delta rebalancing， leaving the problem of optimal hedging in incomplete markets unresolved.Recent advances in deep reinforcement learning （DRL） provide a model-free alternative for option hedging by learning optimal hedging strategies directly from data. Nevertheless， DRL methods suffer from high sample requirements， low data efficiency， and weak tail-risk control， which limit their applicability in real financial markets with scarce data.To address these challenges， a Transfer Learning Deep Hedging （TLDH） framework is proposed that integrates transfer learning with DRL. The core idea is to pretrain a hedging agent in a BSM-based simulated environment， thereby embedding classical hedging knowledge into the agent， and then transfer this knowledge to a real-market environment constructed from historical trading data.The TLDH framework adopts the Twin Delayed Deep Deterministic Policy Gradient （TD3） algorithm to handle continuous state and action spaces， and incorporates a Long Short-Term Memory （LSTM） structure to capture temporal dependencies in hedge positions and portfolio returns. Empirical analysis is conducted using high-frequency data from the CSI 300 index options market in China. The proposed method is compared with empirical deep hedging （EDH）， sim-to-real deep hedging （SRDH）， classical Delta hedging， and cost-adjusted Delta hedging strategies.The results show that TLDH consistently achieves higher Sharpe ratios， lower return volatility， and significantly reduced tail risk. These advantages remain robust across different transaction cost levels and investor risk preferences. Overall， it is demonstrated that combining transfer learning with deep reinforcement learning offers an effective and practical solution for option hedging in real-world incomplete markets， and offers valuable insights for both academic research and artificial intelligence applications in financial practice.

Key words: option hedging, deep learning, reinforcement learning, transfer learning