Neural Networks Speed Up SPX/VIX Calibration – Risk.net

Jointly calibrating models for SPX options and the VIX has become faster thanks to the application of neural networks, according to research published on February 4, 2026. The advancement addresses a longstanding computational challenge in pricing volatility derivatives, which traditionally relies on nested Monte Carlo simulation – a process particularly burdensome for complex models like the four-factor Markov path-dependent volatility model developed by Guyon.

The research, conducted by Fabio Baschetti, Giacomo Bormetti, and Pietro Rossi, builds upon the work of Julien Guyon and proposes a solution to accelerate model calibration. The team’s approach leverages feed-forward neural networks to learn patterns from the volatility surface, utilizing a grid-based method similar to those explored by Hernandez (2017), Horvath et al. (2021), and Rømer (2022). This involves training the neural network to recognize a collection of pixels representing the volatility surface across a two-dimensional grid defined by time to expiry (T) and strike price (k).

A key component of the model is a dictionary-based structure, where a “MODEL” includes a family of neural networks (“MODEL[“nets”]”) and a function (“V_AND_MUY”) that calculates volatility and drift as outputs from these networks. The implementation requires Python 3.6 or later, along with the torch and torchsde libraries (versions 1.13.0 and 0.2.5 respectively). Data for calibration is sourced from the Wharton Research Data Service Optionmetrics database, specifically SPX and VIX bid-ask data in CSV format.

The research highlights the potential for real-time calibration, a significant improvement over traditional methods. The speed gains are particularly relevant for practitioners dealing with the complexities of pricing and hedging volatility products. The use of neural networks allows for a more efficient exploration of the parameter space, leading to faster convergence and reduced computational costs.

The work was presented at the FinTech 2026 conference in Pavia, Italy, where it was reportedly awarded “Best Paper,” according to a LinkedIn post by Mauro Cesa. The development represents a step forward in the application of machine learning to quantitative finance, offering a practical solution to a computationally intensive problem.

The model utilizes a specific structure for defining the neural network, as demonstrated in a sample implementation. This includes sequential layers with linear transformations and hyperbolic tangent (tanh) activation functions. The function V_AND_MUY, crucial for calculating volatility and drift, incorporates constraints to ensure the resulting correlation coefficient (Rho) remains within a reasonable range (-0.99999 to +0.99999). This constraint is enforced using minimum and maximum functions applied to the output of the neural network.

Further research, as detailed in a paper available on arXiv (2212.08297), focuses on developing a unified method for fast and accurate pricing of both SPX and VIX derivatives. This method relies on functional quantization and neural networks to ensure efficient calibration and a fair comparison between different models.