Robust deep hedging of equity-linked securities under covariance uncertainty

Risk management under uncertainty is crucial for equity-linked securities (ELS) with complex nonlinear payoffs that challenge traditional hedging methods. We propose a robust dynamic hedging framework for step-down ELS using deep neural networks and stochastic optimization, incorporating covariance ambiguity. By minimizing a weighted combination of CVaR and loss probability, our method reduces tail risk while preserving the low loss probability inherent in step-down ELS. Numerical results show that the proposed robust strategy outperforms a non-robust one across various scenarios, including worst-case covariance and model misspecifications. Although the strategies are trained under Black–Scholes assumptions, they generalize well to more realistic models, demonstrating practical robustness. This framework supports the development of ELS products that offer stronger protection against large principal losses, helping issuers better manage risk and provide safer investment options for retail investors.