Approximation Rates for Deep Calibration of (Rough) Stochastic Volatility Models

Francesca Biagini, Lukas Gonon, Niklas Walter

Introduction

In modern quantitative finance, the calibration of sophisticated stochastic volatility models is a cornerstone of risk management and derivative pricing. Models such as the Heston model or the rough Bergomi model offer a realistic depiction of market dynamics but come with a significant computational burden. Traditionally, these models require intensive Monte Carlo simulations to price options, making real-time calibration to market data difficult.

To address this, "Deep Calibration"—the use of deep neural networks (DNNs) to learn the mapping between model parameters and option prices—has emerged as a powerful heuristic. However, until recently, the theoretical justification for this approach was incomplete. Specifically, it was unclear whether DNNs could approximate these pricing functions without succumbing to the "curse of dimensionality" (CoD), where the required network size grows exponentially with the number of assets. This paper provides the first rigorous mathematical proof that DNNs can indeed approximate these prices with polynomial complexity, establishing a robust theoretical foundation for deep calibration techniques.

The Curse of Dimensionality in Financial Modeling

The primary challenge in high-dimensional financial engineering is the exponential growth of computational resources required as the number of underlying assets ($d$) increases. Most classical approximation methods fail in high dimensions because their complexity scales poorly. For a DNN to be a viable alternative, one must prove that the network size (number of neurons and layers) required to achieve a target accuracy ($\epsilon$) grows only polynomially with respect to $d$ and $1/\epsilon$. This paper sets out to prove that for a wide class of financial models, DNNs can bypass this curse.

Methodology: Constructive Approximation

The authors employ a constructive proof strategy to derive quantitative error bounds. Instead of merely showing that a network could exist, they establish a functional relationship between the target accuracy, the dimension of the asset vector, and the network architecture.

1. Emulating the Euler-Maruyama Scheme

The proof hinges on the fact that the Euler-Maruyama scheme—a standard method for discretizing continuous-time price dynamics—consists of a sequence of simple algebraic operations and function compositions. Since DNNs are themselves compositions of simple functions (like the ReLU activation), the authors show that a network can be constructed to "mimic" each step of the discretization process, including drift, volatility calculations, and the final Monte Carlo averaging.

2. Overcoming Non-Lipschitz Constraints

A significant hurdle in this proof is that many popular models, such as the Heston model, involve square-root processes that are not Lipschitz continuous near zero. Standard approximation theory often requires smoothness that these models lack. The authors overcome this by introducing "stopping times" and "truncated models." By restricting the process to a regime where the functions behave predictably and proving that the error introduced by truncation is negligible, they extend the approximation guarantees to these non-smooth cases.

3. Addressing Rough Volatility

Rough volatility models, like the rough Bergomi model, present an additional challenge: they are non-Markovian, meaning the future state depends on the entire history of the process. To handle this "memory," the authors utilize a hybrid scheme combining Riemann sums with Wiener integrals to approximate the underlying Volterra processes. This allows the non-Markovian dynamics to be compressed into a DNN architecture while maintaining polynomial complexity.

Key Results

The central contribution of the paper is the proof that the network size is $O(d^p \epsilon^{-q})$. Because the exponent $p$ is fixed and does not grow with the dimension, the model is formally "CoD-free."

Specific applications include:

  • Theorem 2.1 (The Core Result): For a general class of stochastic volatility models, there exists a DNN such that the error in approximating the pricing function is less than $\epsilon$, with the network size bounded polynomially.
  • Multi-Asset Heston Models: The authors successfully extended these bounds to high-dimensional Heston models with complex cross-correlations, a staple of modern multi-asset derivatives pricing.
  • Rough Bergomi Bounds: They proved that for the rough Bergomi model (where volatility is "rougher" than a standard Brownian motion), the approximation rate remains polynomial, provided the payoff functions are bounded.

Discussion and Limitations

While this work provides a vital "existence proof," it is important to distinguish between the existence of an accurate network and the ability to train one. The paper proves that a small, accurate network exists within the hypothesis space, but it does not guarantee that standard optimization algorithms, such as Stochastic Gradient Descent (SGD), will actually find it. The "optimization gap" remains an open area of research.

Furthermore, the results for rough volatility models are currently limited to bounded payoff functions. Extending these proofs to unbounded payoffs, such as standard European Call options, is difficult because the moments of the rough Bergomi model are not yet fully understood in the literature. Finally, the current framework assumes a fixed maturity $T$; future work may explore networks that generalize across different time horizons and strike prices simultaneously.

Conclusion

By proving that deep neural networks can bypass the curse of dimensionality in the context of (rough) stochastic volatility models, this research provides the mathematical justification for the continued adoption of deep learning in quantitative finance. It confirms that the success of deep calibration seen in practice is supported by the fundamental approximation capabilities of neural network architectures, even when dealing with complex, non-Markovian market dynamics.

Summary updated

BibTeX
@article{biagini2024approximation,
  author = {Francesca Biagini and Lukas Gonon and Niklas Walter},
  title = {{Approximation Rates for Deep Calibration of (Rough) Stochastic Volatility Models}},
  journal = {SIAM J. Financial Math.},
  year = {2024},
  url = {https://openreview.net/forum?id=65FOdm0xKn},
}
Figure