Deep Reinforcement Learning for Option Hedging

This thesis investigates deep reinforcement learning as a numerical approach to hedging under frictions, through three incremental experiments. First, we replicate a study formulating hedging as a Markov decision process solved with DDPG under transaction costs, in which two critics estimate the first two moments of the cost still to come. Second, we replace that construction with a single distributional critic, which represents the whole distribution of the cost-to-go through a set of quantiles and from which the actor reads a Conditional Value-at-Risk, so that the objective targets the tail of the hedging cost rather than its dispersion. Third, we sweep the risk aversion parameter of both agents, so that the two objectives are compared as families rather than at one setting each. Throughout, the learned strategies are measured against delta hedging on common held-out paths, under constant volatility and under stochastic volatility, and we examine how the objective shapes hedging behaviour: what it costs in the tail, what it costs in the mean, and how far the resulting position departs from the delta it is benchmarked against.