Deep Reinforcement Learning for Option Hedging : Learning Adaptive Hedging Strategies under Transaction Costs

Sammanfattning

This thesis investigates the application of deep reinforcement learning (DRL) for option hedging in markets with transaction costs and discrete trading, and compares its performance with the traditional Black–Scholes–Merton (BSM) delta hedging approach. Although the BSM model provides a theoretically perfect hedge under idealized frictionless conditions, its performance degrades in real-world settings where constant re-balancing incurs costs. To address these limitations, four DRL agents-Deep Q-Network (DQN), Deep Deterministic Policy Gradient (DDPG), Twin Delayed Deep Deterministic Policy Gradient (TD3), and a proposed hierarchical Hybrid DQN-TD3-model were developed and evaluated. The agents were trained and tested across three market scenarios: a frictionless baseline, a realistic market (1% transaction costs), and a stressed market (5% transaction costs). The results demonstrate that all DRL agents converged to the theoretical BSM delta hedging strategy in the absence of frictions. In environments with transaction costs, the DRL agents outperformed the BSM benchmark in terms of mean average costs. These findings suggest that DRL agents can learn adaptive hedging strategies in non perfect markets.