Paper: arXiv 2410.22568

Authors: Konrad Mueller, Amira Akkari, Lukas Gonon, Ben Wood

Abstract

Hedging exotic options in presence of market frictions is an important risk management task. Deep hedging can solve such hedging problems by training neural network policies in realistic simulated markets. Training these neural networks may be delicate and suffer from slow convergence, particularly for options with long maturities and complex sensitivities to market parameters. To address this, we propose a second-order optimization scheme for deep hedging. We leverage pathwise differentiability to construct a curvature matrix, which we approximate as block-diagonal and Kronecker-factored to efficiently precondition gradients. We evaluate our method on a challenging and practically important problem: hedging a cliquet option on a stock with stochastic volatility by trading in the spot and vanilla options. We find that our second-order scheme can optimize the policy in 1/4 of the number of steps that standard adaptive moment-based optimization takes.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents advanced mathematical concepts like pathwise differentiability, curvature matrix approximation, and Kronecker-factored preconditioning (KFAC), placing it firmly in high math complexity. It demonstrates empirical rigor by testing the proposed optimization scheme on a realistic hedging problem (cliquet option with stochastic volatility) and reporting specific performance improvements (1/4 reduction in steps) compared to standard methods like Adam.

Research Flowchart

  flowchart TD
  A["Research Goal: Fast Deep Hedging in Realistic Markets"] --> B["Data Inputs: Simulated Market<br>Stochastic Volatility Model"]
  B --> C["Methodology: Second-Order Optimization<br>Pathwise Differentiability & Kronecker Approx"]
  C --> D["Computational Process:<br>Block-Diagonal Curvature Matrix Construction"]
  D --> E["Outcome: 4x Faster Convergence<br>vs. First-Order Optimizers"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a second‑order optimization scheme for deep hedging that uses a block‑diagonal, Kronecker‑factored curvature matrix to precondition gradients.

Source sentence
To address this, we propose a second-order optimization scheme for deep hedging. We leverage pathwise differentiability to construct a curvature matrix, which we approximate as block-diagonal and Kronecker-factored to efficiently precondition gradients.

Assumes 5

  • Proportional transaction costs are assumed and accumulate linearly with trade size. execution
    Source sentence
    For simplicity, we consider proportional transaction costs that accumulate to 𝐶𝑇(𝑢) :=
  • The curvature matrix can be approximated as block‑diagonal and Kronecker‑factored (KFAC). model
    Source sentence
    We leverage pathwise differentiability to construct a curvature matrix, which we approximate as block-diagonal and Kronecker-factored to efficiently precondition gradients.
  • KFAC applicability requires the Hessian to be invertible. statistical
    Source sentence
    We assume this for KFAC applicability; the invariance only requires invertibility.
  • The underlying asset follows a stochastic volatility model. market
    Source sentence
    hedging a cliquet option on a stock with stochastic volatility by trading in the spot and vanilla options.
Show 1 more
  • The cliquet option is locally‑capped and globally‑floored in a standard form. other
    Source sentence
    We consider a standard form of the locally-capped, globally-floored cliquet option, which pays at maturity ...

Data

  • Synthetic cliquet‑option hedging simulation synthetic unknown
    Source sentence
    We evaluate our method on a challenging and practically important problem: hedging a cliquet option on a stock with stochastic volatility by trading in the spot and vanilla options.

Costs

Modeled Costs are part of the reported headline results.

Proportional transaction costs c per instrument are included in the objective as E[C_T(u)] and in the inner Hessian via a scaled L2 cost ˜c = 8c.

Source sentence
For simplicity, we consider proportional transaction costs that accumulate to 𝐶𝑇(𝑢) :=

Finds 3

  • The second‑order DH‑KFAC scheme reaches the target validation loss in about 25% of the steps required by Adam. performance
    Source sentence
    DH‑KFAC progresses significantly faster than Adam and only requires about 25% of Adam’s number of steps to reach the targeted loss level.
  • DH‑KFAC optimizes the policy in one‑quarter the number of steps compared to standard adaptive‑moment methods. performance
    Source sentence
    we find that our second‑order scheme can optimize the policy in 1/4 of the number of steps that standard adaptive moment‑based optimization takes.
  • The learned hedging policy reduces PnL variance, with option hedging cutting the standard deviation by another 50% and increasing skewness. performance
    Source sentence
    Including option hedging reduces the PnL’s standard deviation by another 50% and increases the skewness of the PnL distribution significantly.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 24 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (3 by independent authors): A Topological Approach to Parameterizing Deep Hedging … (uses its method or data)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (42)

  1. Natural Gradient Works Efficiently in Learning (1998)
  2. What Matters In On-Policy Reinforcement Learning ? A Large-Scale Empirical Study (2020) — link
  3. Distributed Second-Order Optimization Using Kronecker-Factored Approximations (2017)
  4. Exact Simulation of Stochastic Volatility and Other Affine Jump Diffusion Processes (2006)
  5. Statistical Hedging (2017)
  6. Deep Hedging (2019)
  7. Deep Hedging of Long-Term Financial Derivatives (2021)
  8. Adaptive Gradient Methods at the Edge of Stability (2024) — link
and 34 more
  1. Natural Neural Networks (2015)
  2. Finding Structure in Time (1990)
  3. Fast Approximate Natural Gradient Descent in a Kronecker Factored Eigenbasis (2018)
  4. Lecture Notes on Neural Net Training Dynamics ( Chapter 3: Metrics ) (2022) — link
  5. Mastering Diverse Domains through World Models (2023) — link
  6. Recent Advances in Reinforcement Learning in Finance (2023)
  7. Delving Deep into Rectifiers : Surpassing Human-Level Performance on ImageNet Classification (2015)
  8. A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options (1993)
  9. Long Short-Term Memory (1997)
  10. Deep Hedging under Rough Volatility (2021)
  11. Tutorial on training recurrent neural networks, covering BPPT, RTRL, EKF and the" echo state network" approach (2002)
  12. A Natural Policy Gradient (2001)
  13. Universal Statistics of Fisher Information in Deep Neural Networks: Mean Field Approach (2019)
  14. Adam: A Method for Stochastic Optimization (2015)
  15. The Geometry of Neural Nets ’ Parameter Spaces Under Reparametrization (2023)
  16. A Well-Conditioned Estimator for Large-Dimensional Covariance Matrices (2004)
  17. Shrinkage Estimation of Large Covariance Matrices: Keep It Simple, Statistician? (2021)
  18. Robust Deep Hedging (2022)
  19. Learning Recurrent Neural Networks with Hessian-Free Optimization (2011)
  20. Training Deep and Recurrent Networks with Hessian-Free Optimization (2012)
  21. Optimizing Neural Networks with Kronecker-factored Approximate Curvature (2015)
  22. Kronecker-Factored Curvature Approximations for Recurrent Neural Networks (2018)
  23. New Insights and Perspectives on the Natural Gradient Method (2020)
  24. Monte Carlo Gradient Estimation in Machine Learning (2020)
  25. Deep Hedging : Continuous Reinforcement Learning for Hedging of General Portfolios across Multiple Risk Aversions (2022)
  26. Resurrecting Recurrent Neural Networks for Long Sequences (2023)
  27. Neural Networks for Option Pricing and Hedging: A Literature Review (2020)
  28. Fast Curvature Matrix-Vector Products for Second-Order Gradient Descent (2002)
  29. Trust Region Policy Optimization (2015)
  30. Variance- Optimal Hedging in Discrete Time (1995)
  31. Log-Normal Stochastic Volatility Model with Quadratic Drift (2024)
  32. Rich Information Is Affordable : A Systematic Performance Analysis of Second-order Optimization Using K-FAC (2020)
  33. Scalable Trust-Region Method for Deep Reinforcement Learning Using Kronecker-factored Approximation (2017)
  34. Root Mean Square Layer Normalization (2019)

0 of 42 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.