Classical hedging methods are based on strong generalized assumptions, such as frictionless markets and continuous trading. However, real-world markets are incomplete and are characterized by discrete time trading, transaction costs, and model uncertainty. These frictions render perfect replication or the derivation of the optimal hedging strategy infeasible. Recent advances propose utilizing the flexibility of deep hedging, which involves the use of reinforcement learning to incorporate these constraints and produce an approximate optimal hedging strategy.

The existing deep hedging literature discusses various components, including market dynamics, training methods, objective functions, market frictions, and risk measures. This study identifies key trade-offs in robustness, computational efficiency, and sensitivity to model misspecification.

By forming a rigorous comparison of these approaches within a representative experimental setting, we provide a structured assessment of existing approaches. The results contribute to the literature by clarifying how the design choices of a deep hedging framework affect hedging performance.