Optimal Transport for Model Calibration

Ivan Guo, Grégoire Loeper, Jan Obłój, Shiyi WangSSRN 2021 · DOI

TL;DR

This paper presents a unified framework for calibrating financial models to market option prices using optimal transport theory. It demonstrates how this approach can accurately calibrate various models, including local volatility, local-stochastic volatility, and joint SPX/VIX models, without requiring a continuum of option prices.

What Problem Does This Paper Address?

Calibrating financial models to market prices is a fundamental task in quantitative finance. Traditional methods often rely on fitting models to a vast number of observed option prices, which can be computationally intensive and may not guarantee a perfect fit across all instruments. Specifically, calibrating models to complex derivatives like path-dependent options or jointly calibrating multiple correlated assets (like the S&P 500 index and its volatility index, VIX) presents significant challenges. The paper addresses the need for a robust and efficient framework that can handle these complexities, offering a way to derive model parameters that precisely match observed market prices for a given set of options.

Methods

The core of the paper's contribution lies in its application of optimal transport (OT) theory to model calibration. The authors leverage the dual formulation of optimal transport, which transforms a difficult optimization problem into solving a system of partial differential equations (PDEs), specifically Hamilton-Jacobi-Bellman (HJB) equations.

Here's a breakdown of the key methodological aspects:

  • Probabilistic Formulation: The problem is framed as finding a probability measure (representing a market model) that satisfies certain constraints (e.g., matching expected option payoffs) while minimizing a cost function related to the model's dynamics.
  • Dual Formulation: This is the crucial step. Instead of directly optimizing over probability measures, the authors solve for a dual variable (represented by a function $\phi$) that satisfies an HJB equation. This dual formulation is often more tractable.
  • PDE Solving: The HJB equations are solved using numerical methods, typically finite difference schemes. The paper mentions using implicit finite difference methods and policy iteration for nonlinearities.
  • Gradient Descent: The dual variable $\lambda$ is updated iteratively using a gradient descent algorithm. The gradient is derived from the HJB equation and represents the difference between model-implied option prices and market prices, providing a clear direction for improvement.
  • Model-Specific Cost Functions: For each type of model (local volatility, local-stochastic volatility, etc.), a specific convex cost function $F$ is defined. This function encodes the desired properties of the model and is central to the OT framework.
  • Path-Dependent PDEs (PPDEs): For path-dependent options, the framework naturally leads to infinite-dimensional PDEs. The paper notes that by identifying relevant path-dependent state variables, these can be reduced to finite-dimensional PDEs, making them numerically solvable.

Datasets

The paper primarily uses synthetic data generated from known models to demonstrate the calibration capabilities of their framework. In several sections, they also refer to market data for specific applications:

  • Local Volatility Calibration: Uses synthetic data generated from a local volatility model with a specified volatility surface.
  • Local-Stochastic Volatility Calibration: Employs FX option market data from [22], which includes 10 maturities and 5 European options per maturity.
  • VIX/SPX Joint Calibration: Uses market data as of September 1st, 2020, consisting of monthly SPX options (17 and 45 days maturities) and VIX futures/options (15 days maturity).
  • Path-Dependent Volatility Calibration: Uses synthetic data for barrier options.

The paper does not introduce new benchmark datasets but rather applies its method to existing market data or synthetically generated data representative of market conditions.

Key Insights and Ideas

  • Unified Framework: Optimal transport provides a remarkably unified mathematical framework to tackle calibration problems for a diverse range of financial models, from simple local volatility to complex joint SPX/VIX and path-dependent options.
  • Exact Calibration without Continuum: A significant insight is that the OT approach allows for exact calibration to a finite set of option prices, mirroring the spirit of Dupire's formula but without requiring a continuum of prices. This is a major practical advantage.
  • Dual Formulation Simplifies Complexity: Transforming the calibration problem into solving an HJB equation via the dual formulation is a powerful simplification. It turns a complex, high-dimensional optimization problem into a more manageable PDE problem.
  • Gradient as Price Difference: The interpretation of the gradient of the dual objective function as the difference between model-implied and market option prices provides an intuitive and actionable feedback mechanism for the iterative calibration process.
  • Handling Path-Dependence: The paper highlights that the OT framework can be extended to path-dependent options by reducing infinite-dimensional PPDEs to finite-dimensional ones through the identification of relevant state variables.

Key Results

The paper demonstrates that the proposed optimal transport framework can:

  • Accurately calibrate local volatility, local-stochastic volatility, and joint SPX/VIX models to market data.
  • Achieve exact calibration to the prices of a given set of European options, meaning the model's implied prices for these specific options match market prices precisely.
  • Produce smooth volatility surfaces and skews for complex models, often requiring a reference iteration method to refine the results, especially for local volatility models.
  • Successfully calibrate to FX option market data using the OT-LSV model, showing smooth volatility skews even without extensive iteration.
  • Jointly calibrate SPX and VIX models to market data, satisfying constraints for SPX options, VIX options, VIX futures, and a singular contract to ensure specific model properties (like zero terminal variance).
  • Handle path-dependent options by reducing complex PPDEs to solvable finite-dimensional PDEs.

Limitations and Future Work

  • Numerical Stability and Convergence: While the paper demonstrates successful calibration, the numerical methods (finite difference, gradient descent) can be sensitive to parameters and may face convergence challenges in highly complex scenarios. The "reference iteration method" suggests that initial iterations might produce spiky or unstable surfaces, requiring smoothing.
  • Computational Cost: Solving HJB equations and performing gradient descent can still be computationally intensive, especially for high-dimensional models or a large number of calibration instruments.
  • Assumptions: The framework relies on the existence of a suitable convex cost function $F$ and the ability to reduce PPDEs to finite-dimensional PDEs. The choice of $F$ and the identification of state variables are crucial.
  • Future Work: The paper implicitly suggests future work in several areas:
  • Further exploration of numerical methods for solving the HJB equations more efficiently.
  • Extending the framework to more exotic derivatives and more complex market structures.
  • Investigating the theoretical properties of the calibrated models, such as their behavior in terms of arbitrage and stability.
  • Applying the framework to real-time calibration and risk management.

Summary updated

BibTeX
@misc{guo2021optimal,
  author = {Ivan Guo and Grégoire Loeper and Jan Obłój and Shiyi Wang},
  title = {{Optimal Transport for Model Calibration}},
  year = {2021},
  doi = {10.2139/ssrn.3876854},
  howpublished = {SSRN},
}

Loading related content...

Figure