Existence, uniqueness and positivity of solutions to the Guyon-Lekeufack path-dependent volatility model with general kernels

Basic Information

• Paper ID: 2408.02477
• Title: Existence, uniqueness and positivity of solutions to the Guyon-Lekeufack path-dependent volatility model with general kernels
• Authors: Hervé Andrès (Milliman R&D, Paris & CERMICS, École des Ponts), Benjamin Jourdain (CERMICS, École des Ponts, INRIA)
• Classification: q-fin.CP (Computational Finance)
• Publication Date: October 15, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2408.02477

Abstract

This paper investigates the mathematical theoretical foundations of the path-dependent volatility model proposed by Guyon and Lekeufack (2023). The model describes stock index prices and their spot volatility, with trend and activity characteristics expressible as stochastic Volterra equations (SVE) with non-convolution, unbounded kernels and non-Lipschitz coefficients. The authors first establish existence and uniqueness of SVE solutions under integrability and regularity assumptions on the kernel functions, as well as a second kernel condition ensuring a positive lower bound for the activity characteristic. Subsequently, under the additional assumption that the kernel weighting past returns is exponential and satisfies a condition involving an inequality of logarithmic derivatives of both kernels with respect to the second variable, they prove positivity of the volatility process as a nonlinear function of the SVE solution. Numerical experiments demonstrate that the choice of exponential kernel weighting past returns has minimal impact on model calibration quality, and the calibrated kernel functions satisfy the logarithmic derivative inequality condition.

Research Background and Motivation

Problem Background

1. Challenges in volatility modeling: Traditional volatility models often fail to adequately capture path-dependent characteristics of volatility, particularly leverage effects (volatility increases when prices fall) and volatility clustering (high volatility periods tend to follow high volatility periods).
2. Empirical success of the Guyon-Lekeufack model: The path-dependent volatility model proposed by Guyon and Lekeufack (2023) demonstrates excellent empirical performance, explaining over 80% of volatility changes in major stock indices, demonstrating that volatility is primarily path-dependent.
3. Missing mathematical theory: Despite strong empirical performance, the mathematical theoretical foundations of this model, particularly regarding existence, uniqueness of solutions and positivity guarantees for volatility, remain unestablished.

Research Motivation

1. Theoretical refinement: Provide rigorous mathematical theoretical foundations for the Guyon-Lekeufack model, ensuring mathematical soundness.
2. Generalization extension: Extend existing results for specific kernels (such as exponential kernels) to more general kernel classes.
3. Practical considerations: Ensure theoretical results provide guidance for practical applications while maintaining mathematical rigor.

Core Contributions

1. Existence and uniqueness theorem: Establishes existence and uniqueness of solutions to the stochastic Volterra equation corresponding to the path-dependent volatility model under general kernel conditions, significantly extending existing results.
2. Positivity guarantee: Establishes sufficient conditions for positivity of the volatility process, providing practical positivity verification criteria.
3. Hölder continuity: Proves local Hölder continuity of solutions, providing theoretical support for numerical methods.
4. Empirical verification: Validates the satisfiability of theoretical conditions in practice through calibration experiments on multiple stock index datasets.
5. Methodological contribution: Develops localization techniques for handling stochastic Volterra equations with non-Lipschitz coefficients.

Detailed Methodology

Problem Formulation

Consider the following path-dependent volatility model: $\frac{dS_t}{S_t} = \sigma_t dW_t$

where the volatility is given by: $\sigma_t = \beta_0 + \beta_1 R_{1,t} + \beta_2 \sqrt{R_{2,t}}$

The trend characteristic and activity characteristic are respectively defined as: $R_{1,t} = \int_{-\Delta}^0 K_1(s,t)\sigma_s dB_{-s} + \int_0^t K_1(s,t)\sigma_s dW_s$$R_{2,t} = \int_{-\Delta}^t K_2(s,t)\sigma_s^2 ds$

where $\beta_0, \beta_2 \geq 0$, $\beta_1 \leq 0$, and $K_1, K_2$ are weighting kernel functions.

Model Architecture

The model can be rewritten as the following stochastic Volterra equation: $R_t = g(t) + \int_0^t K_1(s,t)\gamma(R_s)dW_s + \int_0^t K_2(s,t)b(R_s)ds$

where:

• $R_t = (R_{1,t}, R_{2,t})^T$
• $\gamma(r_1, r_2) = (\beta_0 + \beta_1 r_1 + \beta_2\sqrt{r_2}, 0)^T$
• $b(r_1, r_2) = (0, (\beta_0 + \beta_1 r_1 + \beta_2\sqrt{r_2})^2)^T$
• $g(t)$ contains initial condition terms

Technical Innovations

1. Localization technique: Due to the non-Lipschitz nature of coefficients, the authors construct local solutions using truncation function sequences $\{b_n, \gamma_n\}$, then prove that local solutions are actually global.
2. Extended comparison principle: Extends the comparison theorem for stochastic differential equations to Itô processes, used to prove lower bounds on volatility positivity.
3. General kernel handling: Addresses non-convolution, potentially unbounded kernels through appropriate integrability and regularity conditions.

Experimental Setup

Datasets

• Stock index data: S&P 500, STOXX Europe 50, FTSE 100 indices
• Volatility data: VIX, VSTOXX, IVI implied volatility indices; realized volatility computed from 5-minute returns
• Time span: Training set from January 1, 2000 to December 31, 2018; test set from January 1, 2019 to May 15, 2022

Kernel Function Choices

Compared four kernel function combinations:

1. Two time-shifted power-law (TSPL) kernels
2. Two exponential kernels
3. Convex combination of two exponential kernels
4. $K_1$ exponential kernel, $K_2$ TSPL kernel

Evaluation Metrics

Model fitting quality assessed using the $R^2$ coefficient of determination.

Experimental Results

Main Results

Table 1: $R^2$ scores for different kernel function choices

Key Findings

1. Similar performance: The four kernel function choices exhibit very similar performance, indicating that the choice of exponential kernel has limited impact on fitting quality.
2. Positivity condition satisfaction: For the exponential + TSPL combination, all calibration results satisfy the positivity condition $2\lambda\delta \geq \zeta$.
3. Theory-practice integration: The exponential + TSPL combination achieves comparable fitting performance to other choices while guaranteeing theoretical positivity.

Related Work

Stochastic Volterra Equation Theory

• Classical results: Pioneering work by Berger and Mizel (1980), establishing existence and uniqueness for Lipschitz coefficient SVE
• Singular kernels: Cochran et al. (1995) and Coutin-Decreusefond (2001) addressing unbounded kernel cases
• Non-Lipschitz coefficients: Wang (2008) extension under Yamada-Watanabe conditions

Path-Dependent Volatility Models

• Guyon-Lekeufack model: First proposal and verification of path-dependency importance
• Markovian approximation: Nutz and Riveros Valdevenito (2024) existence and positivity results for exponential kernel case
• Rough volatility: Related SVE theory developments, such as work by Abi Jaber et al.

Theoretical Results

Theorem 1 (Existence and Uniqueness)

Under assumptions (I.1)-(I.6), equation (6) admits a unique continuous solution with local $\gamma^*$-Hölder continuity, where $\gamma^* \in (0, \min(\gamma, \frac{1}{2\alpha_1^*}, \frac{1}{\alpha_2^*}))$.

Key assumptions:

• (I.1)-(I.2): Integrability conditions on kernel functions
• (I.3): Boundedness of initial conditions
• (I.4): Higher-order integrability of kernel functions
• (I.5): Hölder continuity of kernel functions
• (I.6): Positive lower bound condition for activity characteristic

Theorem 2 (Positivity)

Under assumptions (I.1)-(I.5) and (II.1)-(II.4), the volatility process $\sigma_t$ has a positive lower bound: $\sigma_t \geq \sigma_0 \exp\left(\beta_1\int_0^t K_1(s,s)dW_s + h(t) - h(0) - \frac{1}{2}\beta_1^2\int_0^t K_1(s,s)^2 ds\right)$

Additional assumptions:

• (II.2): Exponential structure $K_1(s,t) = f(s)e^{h(t)}$
• (II.3): Logarithmic derivative relationship between kernels: $\partial_t K_2(s,t) - 2h'(t)K_2(s,t) \geq 0$

Conclusions and Discussion

Main Conclusions

1. Theoretical refinement: First establishment of a complete mathematical theoretical framework for the Guyon-Lekeufack model with general kernels.
2. Practical guidance: Provides theoretical criteria for kernel selection, particularly conditions guaranteeing volatility positivity.
3. Numerical verification: Theoretical conditions are satisfiable in practical calibration, demonstrating theoretical utility.

Limitations

1. Kernel restrictions: Positivity results require $K_1$ to be exponential, limiting kernel selection flexibility.
2. Truncation issues: Theory does not apply to kernels with finite truncation lags, commonly used in practice.
3. Computational complexity: Numerical implementation of general kernels is more complex than the Markovian case.

Future Directions

1. Relaxing positivity conditions: Seek more general conditions guaranteeing volatility positivity.
2. Numerical methods: Develop efficient numerical solution algorithms.
3. Multi-factor extension: Generalize to higher-dimensional factor models.

In-Depth Evaluation

Strengths

1. Theoretical rigor: Rigorous mathematical reasoning, clear and reasonable assumptions, advanced proof techniques.
2. Practical value: Theoretical results directly guide practical applications, providing actionable kernel selection criteria.
3. Methodological innovation: Localization techniques and extended comparison principles provide new tools for addressing similar problems.
4. Empirical support: Comprehensive numerical experiments validate theoretical effectiveness and practicality.

Weaknesses

1. Assumption limitations: Some assumptions (such as exponential kernel functions) may be overly restrictive in practice.
2. Technical complexity: Complex proof techniques may limit generalization and application of results.
3. Truncation handling: Inadequate treatment of kernel function truncation commonly encountered in practice.

Impact

1. Academic contribution: Makes important contributions to stochastic Volterra equation theory and mathematical finance.
2. Practical guidance: Provides theoretical foundations for practical applications of path-dependent volatility models.
3. Methodological value: Developed techniques have reference value for related problems.

Applicable Scenarios

This research is particularly suitable for:

• Volatility modeling for stocks and indices
• Volatility risk management in derivative pricing
• Financial risk models requiring path-dependency consideration
• Theoretical research on stochastic Volterra equations

References

Main references include:

• Guyon, J. and Lekeufack, J. (2023). Volatility is (mostly) path-dependent. Quantitative Finance.
• Nutz, M. and Riveros Valdevenito, A. (2024). On the Guyon-Lekeufack volatility model. Finance and Stochastics.
• Zhang, X. (2010). Stochastic Volterra equations in Banach spaces. Journal of Functional Analysis.