Deep hedging techniques for equity-linked life insurance contracts

Details

Supervisors

Barigou, Karim

Faculty

Faculté des sciences

Degree label

Master [120] en sciences actuarielles, à finalité spécialisée

Abstract

This thesis explores the application of deep learning techniques to the hedging of equity-linked life insurance contracts with long-term guarantees, focusing in particular on Guaranteed Minimum Death Benefits (GMDB). Such contracts expose insurers to multiple sources of risk, including equity market volatility, interest rate fluctuations, and mortality uncertainty. We present a unified simulation framework that combines Black–Scholes dynamics for equity prices, Hull–White dynamics for stochastic interest rates, and deterministic (Gompertz–Makeham) and stochastic Ornstein–Uhlenbeck mortality models. Within this setting, we compare traditional Local Risk Minimization (LRM) and a Long Short-Term Memory (LSTM) network. Our numerical experiments show that deep hedging, the LSTM architecture, consistently reduces hedging errors and tail risks (measured by Value-at-Risk and Conditional Value-at-Risk) relative to the classical benchmark. These findings underscore the ability of neural networks to capture path-dependency and complex interactions in incomplete insurance markets. It offers actuaries and regulators some important pointers as to how deep learning is likely to be a key part of the future when it comes to insurance risk management and prudential supervision.