Deep Hedging under Transaction Costs: An Auditable NumPy Implementation and Exploratory Study

This record contains the preprint Deep Hedging under Transaction Costs: An Auditable NumPy Implementation and Exploratory Study. The study presents a transparent, from-scratch NumPy implementation of neural hedging for a European call under discrete rebalancing and transaction costs. It makes explicit the terminal-loss convention, the exact finite-sample Conditional Value-at-Risk estimator, the temporal coupling introduced by transaction costs, the analytic network derivatives, and the simulation schemes used for Geometric Brownian Motion and Heston dynamics. The accompanying implementation supports proportional and quadratic transaction costs, state-only and inventory-aware policies, classical hedging benchmarks, automated tests, and centered finite-difference validation of network and pathwise-loss gradients. The numerical experiments are exploratory. In the archived proportional-cost runs, the fitted policy traded less as the cost coefficient increased and produced lower empirical CVaR point estimates than a discretely rebalanced, cost-blind Black--Scholes delta at the positive cost levels examined. These results are descriptive and are not presented as evidence of statistical superiority, global optimality, or live-trading performance. The principal contribution is an auditable and reproducible research framework for studying risk-sensitive neural hedging under transaction costs. The source code, experiment configurations, tests, and reproducibility materials are maintained in the associated GitHub repository.