Paper: arXiv 2309.07843

Authors: Abir Sridi, Paul Bilokon

Abstract

Stochastic volatility models, where the volatility is a stochastic process, can capture most of the essential stylized facts of implied volatility surfaces and give more realistic dynamics of the volatility smile/skew. However, they come with the significant issue that they take too long to calibrate. Alternative calibration methods based on Deep Learning (DL) techniques have been recently used to build fast and accurate solutions to the calibration problem. Huge and Savine developed a Differential Machine Learning (DML) approach, where Machine Learning models are trained on samples of not only features and labels but also differentials of labels to features. The present work aims to apply the DML technique to price vanilla European options (i.e. the calibration instruments), more specifically, puts when the underlying asset follows a Heston model and then calibrate the model on the trained network. DML allows for fast training and accurate pricing. The trained neural network dramatically reduces Heston calibration’s computation time. In this work, we also introduce different regularisation techniques, and we apply them notably in the case of the DML. We compare their performance in reducing overfitting and improving the generalisation error. The DML performance is also compared to the classical DL (without differentiation) one in the case of Feed-Forward Neural Networks. We show that the DML outperforms the DL. The complete code for our experiments is provided in the GitHub repository: https://github.com/asridi/DML-Calibration-Heston-Model

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced stochastic calculus and differential equations for the Heston model, alongside complex deep learning and regularization techniques, indicating high mathematical density. It includes a public GitHub repository with implementation code, rigorous comparisons between Differential Machine Learning and classical methods, and focuses on calibration speed and accuracy for practical application, demonstrating strong empirical readiness.

Research Flowchart

  flowchart TD
  A["Research Goal"] --> B["Data Generation"]
  A["Research Goal"] --> C["Methodology"]
  B --> C
  C --> D["Computational Process"]
  D --> E["Findings"]

  subgraph A ["Research Goal"]
      A1["<b>Apply DML to Heston Calibration</b><br>Improve speed/accuracy of SV model calibration"]
  end

  subgraph B ["Data & Inputs"]
      B1["<b>Simulated Data</b><br>Features: Moneyness, Maturity<br>Labels: Option Price<br>Differentials: Vega, Delta"]
  end

  subgraph C ["Key Methodology"]
      C1["<b>Neural Network Training</b><br>1. Standard DL (MSE Loss)<br>2. DML (inc. diff. terms)"]
      C2["<b>Regularization</b><br>Applied to reduce overfitting"]
  end

  subgraph D ["Computational Process"]
      D1["<b>Optimization Loop</b><br>Calibrate Heston params (k, θ, ρ, σ, v0)<br>to match network predictions"]
  end

  subgraph E ["Key Findings"]
      E1["<b>DML Outperforms DL</b><br>Higher accuracy, lower error"]
      E2["<b>Significant Speedup</b><br>Drastic reduction in calibration time vs. traditional methods"]
  end

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 applies Differential Machine Learning (DML) to price vanilla European put options under the Heston stochastic volatility model and uses the trained network for fast model calibration.

Source sentence
The present work aims to apply the DML technique to price vanilla European options (i.e. the calibration instruments), more specifically, puts when the underlying asset follows a Heston model and then calibrate the model on the trained network.

Assumes 4

  • The underlying asset follows the Heston stochastic volatility dynamics under the risk‑neutral measure. model
    Source sentence
    The Heston model is a stochastic volatility model where the dynamic of the underlying S = {St, t ≥0} and its volatility √vt can be described by the following Stochastic Differential Equation (SDE): (dSt = St(rdt + √vtdW1t ) dvt = κ(θ −vt)dt + σ√vtdW2t )
  • The Feller condition is satisfied for the variance process. model
    Source sentence
    In the sequel, we will work with the dataset generated when the feller condition is satisfied.
  • Synthetic data are generated i.i.d. using Random Latin Hypercube sampling. data
    Source sentence
    assumption about the features and labels (xi, yi), i = 1, . . . , m, is often made. Thus the data needs to be generated by a pseudo‑random method such as simple Monte Carlo (MC)
  • Risk‑free rate r is constant during pricing and calibration. market
    Source sentence
    we generate synthetic dataset for the input parameters for the model (θ, κ, σ, ρ and v0), the option and the market ones, which include m, τ and r within appropriate ranges

Data

  • Synthetic Heston calibration dataset synthetic unknown
    Source sentence
    we generate synthetic dataset for the input parameters for the model (θ, κ, σ, ρ and v0), the option and the market ones, which include m, τ and r within appropriate ranges selected for each parameter.
  • Training set of 16 384 samples synthetic unknown
    Source sentence
    we then decide to retain a dataset of 16 384 samples.

Recognised sources: Yahoo Finance

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • Differential Machine Learning (DML) vastly outperforms classical deep learning (DL) for pricing under the Heston model. performance
    Source sentence
    We show that the DML outperforms the DL.
  • DML can learn effectively from small datasets; with 16 K samples it achieves an MSE of 0.44 bp versus 1.73 bp for a classical network trained on six times more data. predictive-accuracy
    Source sentence
    The MSE on the differential network trained on 16K examples is 0.44 bp, lower than the MSE on the classical network with 6 times the training dataset size, which equals 1.73.
  • The trained neural network dramatically reduces Heston calibration computation time compared to traditional calibration. pricing-calibration
    Source sentence
    the trained neural network dramatically reduces Heston calibration’s computation time.
  • Calibration with three parameters is as accurate as with five but considerably faster, especially when using the DNN‑based price. pricing-calibration
    Source sentence
    We found that the calibration accuracy with three parameters is similar to that with five. The calibration of three parameters is faster regardless of the price computation type (real pricer or DNN‑based).

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 32 related papers). Coverage is partial, so this is neither replication nor refutation.

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

Code

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