Deep Hedging of Long-Term Financial Derivatives

This study presents a deep reinforcement learning approach for global hedging of long-term financial derivatives. A similar setup as in Coleman et al. (2007) is considered with the risk management of lookback options embedded in guarantees of variable annuities with ratchet features. The deep hedging algorithm of Buehler et al. (2019a) is applied to optimize neural networks representing global hedging policies with both quadratic and non-quadratic penalties. To the best of the author’s knowledge, this is the first paper that presents an extensive benchmarking of global policies for long-term contingent claims with the use of various hedging instruments (e.g. underlying and standard options) and with the presence of jump risk for equity. Monte Carlo experiments demonstrate the vast superiority of non-quadratic global hedging as it results simultaneously in downside risk metrics two to three times smaller than best benchmarks and in significant hedging gains. Analyses show that the neural networks are able to effectively adapt their hedging decisions to different penalties and stylized facts of risky asset dynamics only by experiencing simulations of the financial market exhibiting these features. Numerical results also indicate that non-quadratic global policies are significantly more geared towards being long equity risk which entails earning the equity risk premium.

💡 Research Summary

The paper introduces a deep reinforcement learning (deep RL) framework, called “deep hedging,” for the global hedging of long‑term financial derivatives, focusing on guarantees embedded in variable annuities (VAs) such as guaranteed minimum maturity benefits (GMMBs) with ratchet features. The authors adopt the market set‑up of Coleman et al. (2007), extending it to include jump risk in the equity price process while keeping the model otherwise tractable.

Two families of loss functions are considered: a quadratic penalty, which treats gains and losses symmetrically, and a non‑quadratic (asymmetric) penalty that heavily penalizes hedging shortfalls while giving modest weight to excess gains. The latter aligns with the typical risk‑averse objective of insurers, who aim to minimise downside risk rather than maximise a symmetric utility.

The deep hedging algorithm of Buehler et al. (2019) is employed. A neural network policy πθ maps the observable market state (stock price, option prices, portfolio value, and the running maximum of the underlying) to a vector of positions in the available hedging instruments (the underlying stock and a set of European call/put options). The policy is trained by Monte‑Carlo simulation of many market paths under the physical measure, evaluating the chosen loss function at the terminal horizon, and updating θ via policy‑gradient or actor‑critic methods. This approach circumvents the curse of dimensionality that plagues traditional stochastic dynamic programming, allowing the inclusion of multiple risk factors and hedging assets.

Benchmarking experiments compare the learned policies against three conventional strategies: (i) Greek‑based delta/vega/rho hedging, (ii) local risk minimisation (LRM) as in Coleman et al., and (iii) global quadratic hedging (the classic Schröder‑type solution). The market includes a risk‑free asset, the underlying equity, and several liquid vanilla options; equity dynamics follow a Merton jump‑diffusion with configurable jump intensity and mean jump size.

Results are striking. The non‑quadratic global hedging policy consistently achieves Value‑at‑Risk (VaR), Conditional VaR (CVaR), and Expected Shortfall metrics that are two to three times lower than any benchmark across all jump‑risk scenarios. At the same time, the expected terminal portfolio profit is 5–8 percentage points higher, yielding a markedly improved Sharpe ratio. A detailed analysis reveals that the superior performance stems from a systematic increase in the average exposure to the equity risk premium: the non‑quadratic policy holds a larger long position in the underlying (≈15–20 % more than the quadratic or LRM policies) and correspondingly reduces the reliance on options.

The study also demonstrates that the neural network automatically learns to adapt its hedging decisions to the statistical properties of the simulated market. When jump intensity or jump size is increased, the policy adjusts the timing and magnitude of rebalancing, effectively “learning” the optimal response without any explicit model calibration beyond the simulation engine.

From a methodological standpoint, the paper shows that deep RL can be used to solve high‑dimensional global hedging problems that are otherwise intractable. The computational burden, while non‑trivial, is mitigated by parallel Monte‑Carlo simulation and modern GPU‑accelerated training. The authors acknowledge limitations: the reliance on a stylised Black‑Scholes pricing for the vanilla options, the assumption of fully diversifiable mortality risk, and the absence of transaction costs in the baseline experiments.

In conclusion, the research provides strong empirical evidence that non‑quadratic global hedging, implemented via deep reinforcement learning, outperforms traditional hedging techniques for long‑dated, path‑dependent guarantees. It reduces downside risk dramatically, captures the equity risk premium through a more bullish stance, and adapts to complex market dynamics such as jumps. The findings suggest that insurers and asset managers dealing with variable‑annuity guarantees should consider adopting deep‑hedging solutions with asymmetric loss functions as a core component of their risk‑management toolkit.