Neural Networks Achieve 1000x Speedup for Financial Derivatives Pricing Models: New Calibration Scheme for Heston and SABR Stochastic Volatility Models

An open-source project by Abdellah Kahlaoui, a Master's student in Applied Mathematics from Morocco, replaces the traditional Levenberg-Marquardt optimizer with a feedforward neural network, reducing the calibration time of Heston and SABR stochastic volatility models from 2-5 seconds to under 1 millisecond, providing a practical toolchain for scenarios like high-frequency trading in quantitative finance.

In the field of financial derivatives pricing, calibrating stochastic volatility models is a core task for quantitative analysts. The Heston and SABR models are industry-standard tools. Traditional methods rely on numerical optimizers like Levenberg-Marquardt to solve nonlinear least squares problems, taking 2-5 seconds per calibration. This becomes a fatal bottleneck in high-frequency trading and real-time risk management scenarios.

The project's core innovation is defining the calibration problem as a supervised learning task: generate a large amount of synthetic implied volatility surface data via Monte Carlo simulation, then train a feedforward neural network to directly learn the mapping from volatility surfaces to model parameters. During inference, a single forward pass can output parameters in under 1 millisecond, achieving an approximately 1000x speedup. Moreover, inference can be parallelized, making it suitable for GPU deployment.

The project adopts a modular design: Data Generation Layer: Includes Heston/SABR Monte Carlo pricing engines and parameter sampling modules, generating training data via large-scale random sampling; Neural Network Layer: Builds a feedforward network based on PyTorch to process grid inputs of volatility surfaces, with loss functions targeting parameter estimation errors; Benchmark Comparison Layer: Retains the classic Levenberg-Marquardt optimizer for performance evaluation; Visualization Layer: Integrates Plotly to generate 3D volatility surface visualizations to show fitting degrees.

The project uses RMSE and MAE for evaluation, verifying that the neural network achieves a three-order-of-magnitude speedup while maintaining reasonable accuracy. The trade-off between accuracy and speed has practical value: in fast approximation scenarios, it can replace traditional optimizers; in high-precision scenarios, it can first use the neural network to generate initial guesses before refinement.

Suitable scenarios:

• High-frequency trading market making (continuously updating volatility surfaces);
• Real-time risk reporting (frequently calculating Greeks and risk indicators);
• Historical backtesting analysis (shortening the calibration cycle for large-scale data);
• Teaching and research (helping understand numerical methods for stochastic volatility models). Deployment depends on Python libraries: NumPy, PyTorch, SciPy, Plotly, ensuring portability and community support.

The author is a Master's student in Applied Mathematics at FST Settat, Morocco. The project is open-sourced under the MIT license, allowing commercial use. The open-source value lies not only in the code itself but also in demonstrating the idea of transforming traditional numerical computation problems into machine learning tasks, which can be extended to parameter estimation problems of other financial models.