Deep Hedging with Transaction Costs

This preprint studies option hedging under discrete rebalancing and proportional transaction costs. It compares several rule-based hedging strategies, including Black-Scholes delta hedging, the Leland volatility adjustment, and a delta-band strategy, against a deep learning approach trained to directly minimize the Conditional Value-at-Risk (CVaR) of residual profit-and-loss. The deep hedging model is implemented as a two-layer LSTM trained end-to-end using the Rockafellar-Uryasev formulation of CVaR. The empirical analysis is conducted on Monte Carlo simulations of a European call option under geometric Brownian motion, with additional robustness tests under a jump-diffusion specification. The results show that the learned strategy can reduce tail risk relative to standard rule-based hedging methods, especially when transaction costs are significant. The paper focuses on a single-option setting and is intended as a first preprint version of a broader research project in quantitative finance and machine learning.