Paper: arXiv 2409.13567

Authors: John Armstrong, George Tatlow

Abstract

We train neural networks to learn optimal replication strategies for an option when two replicating instruments are available, namely the underlying and a hedging option. If the price of the hedging option matches that of the Black–Scholes model then we find the network will successfully learn the Black-Scholes gamma hedging strategy, even if the dynamics of the underlying do not match the Black–Scholes model, so long as we choose a loss function that rewards coping with model uncertainty. Our results suggest that the reason gamma hedging is used in practice is to account for model uncertainty rather than to reduce the impact of transaction costs.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 4.5/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper employs advanced mathematics including rough-path theory, Itô calculus, and neural network optimization, yet its empirical approach relies on synthetic data and simulation without real-world backtesting or published code, placing it in the theoretical ‘Lab Rats’ quadrant.

Research Flowchart

  flowchart TD
  A["Research Goal: Why is Gamma Hedging<br>Used in Practice?"] --> B{"Methodology"}
  
  subgraph B ["Key Methodology"]
      B1["Train Neural Networks<br>to Learn Replication Strategies"]
      B2["Use Two Instruments:<br>Underlying & Hedging Option"]
      B3["Define Loss Function<br>Rewarding Model Uncertainty Coping"]
  end
  
  B --> C{"Data & Inputs"}
  
  subgraph C ["Data & Inputs"]
      C1["Black-Scholes<br>Priced Hedging Option"]
      C2["Non-BS Dynamics<br>Underlying Asset"]
  end
  
  C --> D{"Computational Process"}
  
  subgraph D ["Computational Process"]
      D1["Neural Network<br>Optimizes Hedging Strategy"]
      D2["Minimize Loss Function<br>with Uncertainty Reward"]
  end
  
  D --> E{"Key Findings / Outcomes"}
  
  subgraph E ["Key Findings"]
      E1["Network Learns<br>BS Gamma Hedging Strategy"]
      E2["Effective Even When<br>Underlying Dynamics Deviate from BS"]
      E3["Gamma Hedging's Purpose:<br>Manage Model Uncertainty"]
  end
  
  E --> F["Conclusion: Gamma hedging is used<br>to account for model uncertainty"]

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 using neural networks to learn optimal replication strategies that implement gamma hedging with two instruments, the underlying and a hedging option.

Source sentence
We train neural networks to learn optimal replication strategies for an option when two replicating instruments are available, namely the underlying and a hedging option.

Assumes 6

  • Option prices are given by the Black‑Scholes formula with a fixed volatility σ. model
    Source sentence
    We will assume that at all times t ∈[0, T ′] the price of the option with strike Kα is given by the Black‑Scholes formula Vαt = BS(St, Kα, r, T −t, σ).
  • The underlying follows dSt = St(µη dt + ση dWt) with random parameters µη, ση and random initial condition Sη. market
    Source sentence
    where µη and ση are random parameter values and Sη is a random initial condition.
  • Proportional transaction costs pα are incurred when purchasing each derivative. execution
    Source sentence
    We will assume that there are proportional transaction costs pα given when purchasing the derivative α.
  • All derivative contracts are European options with smooth payoff functions at maturity. model
    Source sentence
    we want the gamma of all these derivative contracts to be finite on the whole of [0, T ′]. To achieve this we will assume that the derivative contracts are all European options with smooth payoff functions at time T ′.
Show 2 more
  • Trading occurs at evenly spaced discrete time points {0, δt, 2δt, …, (n‑1)δt} with nδt = T′. execution
    Source sentence
    The trader will only trade at evenly spaced discrete time points {0, δt, 2δt, . . . , (n −1)δt} where nδt = T ′.
  • The payoff functions at time T′ are given by the Black‑Scholes formula for call options with maturity T > T′. model
    Source sentence
    convenience we will further assume that the payoff functions at time T ′ are given by the Black‑Scholes formula for the price of call options with maturity at time T, where T > T ′.

Data

  • Synthetic market data based on S&P 500 parameters synthetic unknown
    Source sentence
    Our volatility values and our proportional transaction cost values in section 4.3 are based on market data for the S&P 500.

Costs

Modeled Costs are part of the reported headline results.

Proportional costs p1 and p2 are set to 0.0 % (no cost), 0.005 %/0.25 % (normal) and 0.5 %/0.5 % (high) per trade.

Source sentence
We evaluate the performance of this network under three different sets of proportional transaction costs: no transaction costs (p1 = p2 = 0.0%); normal transaction costs (p1 = 0.005%, p2 = 0.25%) based on market data for the S&P 500; and high transaction costs (p1 = p2 = 0.5%).

Finds 5

  • When the hedging option price matches Black‑Scholes, the network learns the Black‑Scholes gamma‑hedging strategy even if the underlying dynamics differ, provided the loss rewards coping with model uncertainty. theoretical
    Source sentence
    If the price of the hedging option matches that of the Black–Scholes model then we find the network will successfully learn the Black‑Scholes gamma hedging strategy, even if the dynamics of the underlying do not match the Black‑Scholes model, so long as we choose a loss function that rewards coping with model uncertain
  • Deep‑hedging strategies using both the underlying and an option outperform those using only the underlying, both with known dynamics and under model uncertainty. performance
    Source sentence
    an option markedly outperform strategies which only use the underlying.
  • As the hedging interval shrinks, the robust optimal strategy converges to the gamma‑hedging strategy. theoretical
    Source sentence
    as we increase the values of n and thus δt →0, the optimal robust hedging strategy tends to the gamma‑hedging strategy.
  • When transaction costs are introduced, blindly following gamma hedging becomes suboptimal; deep‑learning discovers a nuanced adaptation that substantially outperforms other strategies. performance
    Source sentence
    Whilst gamma hedging can offer an improved performance to delta hedging when focusing on robust hedging, once transaction costs are introduce, blindly gamma hedging is suboptimal. Deep‑learning can identify a more nuanced adaptation of gamma‑hedging which substantially outperforms the other strategies.
  • Gamma hedging is more valuable for coping with model uncertainty than for reducing transaction costs. theoretical
    Source sentence
    Together, these results suggest that in practice gamma hedging is more valuable for its ability to cope with model uncertainty than for reducing transaction costs.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 31 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive (2 by independent authors)

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 (19)

  1. Deep Hedging (2018) — link
  2. Lecture notes for 7CCMFM18 , M achine L earning (2022)
  3. Deep Hedging: Hedging Derivatives Under Generic Market Frictions Using Reinforcement Learning (2019) — link
  4. Gamma Hedging and Rough Paths — → 2309.05054
  5. A Nonparametric Approach to Pricing and Hedging Derivative Securities Via Learning Networks (1994) — link
  6. Lecture Notes Learning to Trade III: Deep Hedging with Impact, Deep Bellman Hedging, Open Research Questions (2022) — link
  7. Deep Hedging: Learning to Simulate Equity Option Markets (2019) — link
  8. Deep Bellman Hedging (2022) — link
and 11 more
  1. Quantum D eep H edging (2023) — link
  2. Adam: A Method for Stochastic Optimization (2017) — link
  3. Robust Hedging in Incomplete Markets (2018) — link
  4. Deep Hedging under Rough Volatility (2021) — link
  5. Deep Hedging: Continuous Reinforcement Learning for Hedging of General Portfolios across Multiple Risk Aversions (2022) — link
  6. Deep Hedging of Long-Term Financial Derivatives (2020) — link
  7. Deep Hedging of Derivatives Using Reinforcement Learning (2020) — link
  8. Deeper Hedging: A New Agent-based Model for Effective Deep Hedging — → 2310.18755
  9. Stochastic differential equations (1992)
  10. The Tracking Error Rate of the Delta-Gamma Hedging Strategy (2012)
  11. On the convergence of higher order hedging schemes (2009)

2 of 19 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.