Deep Hedging Using Twin Delayed DDPG

This thesis explores the application of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, a variant of Deep Reinforcement Learning (DRL), to option hedging. Unlike traditional models like Black-Scholes, which assume ideal market conditions, this study considers real-world factors such as discontinuous trading and transaction costs. We evaluate the TD3 agent's performance against the Black-Scholes delta hedging model and the Wilmott delta strategy, focusing on different moneyness levels of options. Our findings reveal that the TD3 agent can effectively balance transaction costs and hedging errors. In a Geometric Brownian motion environment, the agent stays under-hedged to minimize costs, especially when options are deep in-the-money. Coversely, with a stochastic volatility model, the agent over-hedges, highlighting the impact of volatility on the hedging strategy.