Paper: arXiv 2508.02759

Authors: Eduardo Abi Jaber, Louis-Amand Gérard

Abstract

We investigate the use of path signatures in a machine learning context for hedging exotic derivatives under non-Markovian stochastic volatility models. In a deep learning setting, we use signatures as features in feedforward neural networks and show that they outperform LSTMs in most cases, with orders of magnitude less training compute. In a shallow learning setting, we compare two regression approaches: the first directly learns the hedging strategy from the expected signature of the price process; the second models the dynamics of volatility using a signature volatility model, calibrated on the expected signature of the volatility. Solving the hedging problem in the calibrated signature volatility model yields more accurate and stable results across different payoffs and volatility dynamics.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper is mathematically dense, involving advanced concepts from tensor algebra and stochastic calculus to define path signatures, and includes theoretical proofs like universal approximation theorems. However, it primarily presents a methodology and comparative analysis without providing code, detailed backtesting metrics, or specific dataset results, focusing more on theoretical and algorithmic contributions.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Hedging exotic derivatives<br>under non-Markovian<br>stochastic volatility"] --> B["Input Data:<br>Simulated price paths<br>with complex volatility dynamics"]
  B --> C["Deep Learning:<br>Signatures as features<br>in Feedforward Neural Networks"]
  B --> D["Shallow Learning:<br>Signature Volatility Model<br>+ Regression Strategies"]
  C --> E["Compare with<br>LSTM baseline"]
  D --> F["Calibration:<br>Match expected signatures<br>of volatility process"]
  C & E & F --> G["Key Findings:<br>Signatures outperform LSTMs<br>with less compute.<br>Signature models yield<br>more stable hedging strategies."]

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 path signatures as features in feed‑forward neural networks for hedging exotic derivatives, showing they outperform LSTMs with far less training compute.

Source sentence
In a deep learning setting, we use signatures as features in feedforward neural networks and show that they outperform LSTMs in most cases, with orders of magnitude less training compute.

Assumes 6

  • The spot price follows a geometric Brownian motion for the polynomial payoff experiment. model
    Source sentence
    where S is a geometric Brownian motion in order to have an explicit solution to compare the methods to.
  • The European call payoff assumes the underlying follows the Heston model with CIR variance dynamics. model
    Source sentence
    where S follows the Heston model with its variance Vt = Σ2 t having CIR dynamics (5.2).
  • The Asian call payoff assumes a geometric price process with delayed‑equation volatility. model
    Source sentence
    where S will be geometric with delayed-equation (DE) volatility (5.3).
  • The methods assume access to the expected signature of the price or volatility, either exactly or via Monte‑Carlo simulation. data
    Source sentence
    We assume throughout this section that we are given such data (either exactly or through Monte Carlo simulations).
Show 2 more
  • Linear signature regressions are agnostic to the underlying dynamics and can be applied in general contexts. other
    Source sentence
    By contrast, the linear signature regressions are agnostic to the underlying dynamics and can be applied in more general contexts.
  • Training and evaluation are performed over multiple random initialisations (50 for SNN, 10 for RNN). execution
    Source sentence
    All numerical experiments have been averaged over 50 iterations for the SNN and 10 for the RNN, lower for obvious time reasons.

Data

Synthetic data only.

Costs

Not modeled A trading or execution result is reported without costs.

Finds 5

  • Signature‑based feed‑forward networks achieve better hedging performance than LSTMs while requiring orders of magnitude less training compute. performance
    Source sentence
    In a deep learning setting, we use signatures as features in feedforward neural networks and show that they outperform LSTMs in most cases, with orders of magnitude less training compute.
  • Calibrating a signature volatility model and solving the hedging problem within it yields more accurate and stable results across payoffs and volatility dynamics. performance
    Source sentence
    Solving the hedging problem in the calibrated signature volatility model yields more accurate and stable results across different payoffs and volatility dynamics.
  • Both signature and log‑signature neural networks plateau after truncation order 2, with the signature NN achieving lower mean‑squared P&L than the log‑signature NN under the Heston model. performance
    Source sentence
    Figure 12 shows that the signature and log‑signature have a significantly different impact on training with the Log‑Signature NN plateauing relatively high and early in the truncation, whereas the Signature NN goes lower and does not seems to plateau under the Heston model.
  • Most of the hedging performance can be attributed to the first two or three orders of the signature. theoretical
    Source sentence
    conclusions are much less clear but we can still see that most of the performance can be attributed to the first 2 or 3 orders of the signature.
  • Higher truncation orders beyond 3 provide little to no additional benefit and may even delay convergence. negative
    Source sentence
    However, apart from a small improvement between truncation 2 and 3 for the look‑back call option, it seems that higher orders either don’t help or actually delay the convergence.

Checked by others

Contradicted by

  • Option pricing under non-Markovian stochastic volatility models: A deep signature approach (2025)
    On the question: How does increasing truncation order affect model convergence or approximation error?
    Compare the findings

    This paper: Higher truncation orders beyond 3 provide little to no additional benefit and may even delay convergence.

    That paper: Approximation errors for option prices decrease monotonically with higher truncation levels for both SDE and PDE schemes.

    A finds orders >3 give little benefit and may delay convergence, while B finds errors decrease monotonically with higher truncation levels.

  • Sig-Splines: universal approximation and convex calibration of time series generative models (2023)
    On the question: How does truncation order affect model performance?
    Compare the findings

    This paper: Higher truncation orders beyond 3 provide little to no additional benefit and may even delay convergence.

    That paper: Higher orders of truncation improve autocorrelation metrics, indicating better capture of serial independence and volatility clustering.

    A finds higher orders improve autocorrelation metrics, while B finds orders >3 give little benefit and may delay convergence.

Cited by 2 papers in the archive (2 by independent authors): Deep kernel 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 (58)

  1. A primer on the signature method in machine learning (2016) — link
  2. Deep signature transforms (2019)
  3. Optimal stopping with signatures (2023)
  4. Joint calibration to SPX and VIX options with signature-based models — → 2301.13235
  5. Non-parametric Pricing and Hedging of Exotic Derivatives (2020) — link
  6. Lie Elements and an Algebra Associated With Shuffles (1958)
  7. The algebra of iterated stochastic integrals (1994) — link
  8. Perfect hedging in rough Heston models (2017) — link
and 50 more
  1. A Guided Tour Through Quadratic Hedging Approaches (2001) — link
  2. Signature Volatility Models: Pricing and Hedging with Fourier (2025) — link
  3. Deep Hedging (2018) — link
  4. SigFormer: Signature Transformers for Deep Hedging (2023) — link
  5. Discretely sampled signals and the rough Hoff process (2016) — link
  6. Pricing American options under rough volatility using deep-signatures and signature-kernels — → 2501.06758
  7. Learning stochastic differential equations using RNN with log signature features (2019) — link
  8. Calculation of Iterated-Integral Signatures and Log Signatures (2017) — link
  9. Long Short-Term Memory (1997) — link
  10. Attention is All you Need (2017) — link
  11. Adam: A Method for Stochastic Optimization (2017) — link
  12. A Theory of the Term Structure of Interest Rates (1985)
  13. A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options (1993) — link
  14. A new look at the Heston characteristic function (2009) — link
  15. Simulating integrated Volterra square-root processes and Volterra Heston models via Inverse Gaussian — → 2504.19885
  16. Path-dependent processes from signatures (2024) — link
  17. Hedging of Non-redundant Contingent Claims (1985) — link
  18. B rownian Motion and Stochastic Calculus (1988)
  19. Pricing under rough volatility (2016) — link
  20. Integration of Paths, Geometric Invariants and a Generalized Baker- Hausdorff Formula (1957)
  21. Volatility models in practice: Rough, Path-dependent or Markovian? — → 2401.03345
  22. Modeling and analysis of stochastic differential equations driven by point processes (1978) — link
  23. Modeling and approximation of stochastic differential equations driven by semimartingales † (1981) — link
  24. Signature-Based Models: Theory and Calibration (2023) — link
  25. Pricing American options under rough volatility using deep-signatures and signature-kernels — → 2501.06758
  26. Differential Equations Driven by Rough Paths: 'E cole d’ 'E t 'e de Probabilit 'e s de Saint-Flour XXXIV - 2004 (2007) — link
  27. S ignatory: differentiable computations of the signature and logsignature transforms, on both CPU and GPU (2021)
  28. Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation (2016) — link
  29. Dota 2 with Large Scale Deep Reinforcement Learning (2019) — link
  30. Grandmaster level in StarCraft II using multi-agent reinforcement learning (2019) — link
  31. Stochastic Control with Signatures (2024) — link
  32. Learning from the past, predicting the statistics for the future, learning an evolving system (2016) — link
  33. Universal approximation theorems for continuous functions of c`adl`ag paths and L'evy-type signature models (2023) — link
  34. Parameter estimation and bias correction for diffusion processes (2009) — link
  35. Decoupled Weight Decay Regularization (2019) — link
  36. Martingale property and moment explosions in signature volatility models — → 2503.17103
  37. Problems in stochastic analysis. Connections between rough paths and non-commutative harmonic analysis (2003)
  38. Extracting information from the signature of a financial data stream (2014) — link
  39. Application of the Signature Method to Pattern Recognition in the CEQUEL Clinical Trial (2016) — link
  40. A signature-based machine learning model for distinguishing bipolar disorder and borderline personality disorder (2018) — link
  41. Learning Spatial-Semantic Context with Fully Convolutional Recurrent Network for Online Handwritten Chinese Text Recognition (2017) — link
  42. Developing the Path Signature Methodology and its Application to Landmark-based Human Action Recognition (2019) — link
  43. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation (2014) — link
  44. Kernel Density Estimators (1992) — link
  45. Efficiently Modeling Long Sequences with Structured State Spaces (2022) — link
  46. Sig-SDEs model for quantitative finance (2020)
  47. Neural networks for option pricing and hedging: a literature review (2020) — link
  48. Deep Hedging under Rough Volatility (2021) — link
  49. Signature Trading: A Path-Dependent Extension of the Mean-Variance Framework with Exogenous Signals — → 2308.15135
  50. Signature SDEs from an affine and polynomial perspective — → 2302.01362

8 of 58 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.