Local-Stochastic Volatility Models

• Local-stochastic volatility models are hybrid frameworks that integrate deterministic local volatility functions with stochastic variance to capture market dynamics.
• They employ rigorous calibration techniques using PDEs, particle methods, and kernel regression to accurately match market-observed option prices.
• These models are pivotal for pricing exotic derivatives and managing financial risk by effectively reproducing complex volatility surfaces.

Local-stochastic volatility (LSV) models form a broad and technically rich class of stochastic processes for financial assets in which the instantaneous volatility exhibits both deterministic (local) and random (stochastic) components. These models combine the calibration properties of local volatility models, which are constructed to fit the full implied volatility surface, with the pathwise and term-structure flexibility of stochastic volatility models, such as Heston or SABR. The resulting hybrid models are crucial in the pricing and risk management of path-dependent and exotic derivatives, as well as for real-time risk systems and scenario generation.

1. Canonical Model Structure and Mathematical Formulation

A generic local-stochastic volatility model for the asset price $S_t$ under the risk-neutral measure $\mathbb{Q}$ is governed by stochastic differential equations of the form: $$\begin{aligned} dS_t &= S_t\, \ell(t, S_t)\, \sqrt{V_t}\, dW_t + S_t(r-q)\,dt \ dV_t &= \mu(V_t)\,dt + \sigma(V_t)\,dZ_t, \ \end{aligned}$$ with $\corr(dW, dZ) = \rho$. The function $\ell(t, S)$, commonly called the leverage or local volatility function, is chosen to ensure correct marginal laws at each $(t,S)$. The process $V_t$ may follow mean-reverting (Heston), log-normal (SABR), or rough paths (rough Bergomi), among others (Guerrero et al., 2022, Dall'Acqua et al., 2022, Pirjol et al., 2024, Bank et al., 2023). The leverage function is often constructed to satisfy

$$\ell^2(t,S) = \frac{\sigma_{\text{Dup}}^2(t,S)}{\mathbb{E}[V_t \mid S_t=S]},$$

where $\sigma_{\text{Dup}}$ is the Dupire local variance inferred from market option prices (Saporito et al., 2017, Bayer et al., 2022, Djete, 2022, Mustapha, 2024).

In the McKean–Vlasov formulation, the volatility coefficient may depend directly on the law of $(S_t, V_t)$, i.e.

$$dS_t = S_t\, \ell(t, S_t, \mu_t)\, \sqrt{V_t}\, dW_t,$$

where $\ell$ is a functional requiring computation or approximation of conditional expectations via particles or kernel methods (Bayer et al., 2022, Djete, 2022, Mustapha, 2024, Friz et al., 12 Jun 2025).

2. Calibration and Inverse Problem Framework

Calibration of LSV models centers around solving the so-called inverse problem: for a given implied volatility surface (or set of vanilla option prices), determine the leverage function $\ell$ such that the model reproduces these prices marginally. This is typically realized by Gyöngy's mimicking theorem and Dupire's formula, leading to the nonlinear, and possibly singular, fixed-point condition

$$\ell^2(t,S) = \frac{\sigma_{\text{Dup}}^2(t,S)}{\mathbb{E}[V_t \mid S_t = S]}.$$

The conditional expectation is generally intractable, requiring solution of the joint forward Kolmogorov (Fokker–Planck) PDE for $(S_t, V_t)$ and evaluation of the quotient of marginal and joint densities. The PDE systems are high-dimensional and nonlocal; regularization in $S$ (“Tikhonov” or other smoothness priors) is employed for stability, especially in low-probability regions (Saporito et al., 2017). Numerical approaches range from ADI or finite-element methods for the PDE, to particle and regression-based conditional moment estimators (Bayer et al., 2022, Ogetbil et al., 2020, Mustapha, 2024).

In multi-factor or hybrid settings (e.g., 4-factor FX LSV models with CIR++ stochastic rates), the calibration extends to leverage functions that must account for conditional moments not just of $V_t$, but also interest rate factors, under a risk-neutral or forward measure (Cozma et al., 2017, Ogetbil et al., 2020, Cozma et al., 2015).

3. Particle and Numerical Schemes: Forward and Backward Methods

To circumvent dimensionality constraints of PDE approaches, particle methods are ubiquitous in LSV calibration (Cozma et al., 2017, Friz et al., 12 Jun 2025, Mustapha, 2024). An interacting particle system is simulated forward in time, dynamically updating the leverage function “on-the-fly” as a running estimator of the conditional expectation $\mathbb{E}[V_t \mid S_t = S]$, either via kernel regression or Gaussian smoothing. The general particle update for the $i$th path at time $t_j$ is: $$S^i_{j+1} = S^i_j + \ell(t_j, S^i_j) \sqrt{V^i_j} S^i_j \Delta W^i_j,$$ with empirical estimation: $$\widehat{\mathbb{E}}[V_t | S_t = s] = \frac{\sum_{i} V^i_t \, K_h(S^i_t - s)}{\sum_i K_h(S^i_t - s)},$$ for a suitable kernel $K_h$ of bandwidth $h$. For accuracy, a variety of variance-reduction control variates are employed, such as leveraging PDE-based solutions in simpler 2D LSV submodels for conditional moments (Cozma et al., 2017).

Strong and weak convergence of these schemes is well-understood in regular settings with quantitative weak error bounds derived for Euler and half-step methods (Friz et al., 12 Jun 2025, Cozma et al., 2015). For example, particle methods achieve convergence rates of $O(N^{-1/2})$ for the propagation of chaos, and $O(h)$ in time discretization, under suitable regularity and ellipticity assumptions (Friz et al., 12 Jun 2025, Bayer et al., 2022, Mustapha, 2024).

4. Model Variants and Functional Frameworks

4.1 Heston/SLV and Hybrids

In the Heston SLV model,

$$\begin{aligned} dS_t &= rS_t dt + \sqrt{V_t}S_t\ell(t,S_t)[\rho dW_t + \sqrt{1-\rho^2} d\widetilde{W}_t], \ dV_t &= \kappa(\theta - V_t)dt + \gamma\sqrt{V_t}dW_t, \end{aligned}$$

the simulation of the CIR variance process requires careful numerical schemes, with backward Euler (“almost exact”) and truncated Euler providing robust convergence and positivity preservation for $V_t$ (Cai et al., 29 Sep 2025). Backward Euler is more accurate but computationally expensive; explicit (truncated) Euler is preferable for real-time calibration (Cai et al., 29 Sep 2025, Cozma et al., 2015).

4.2 Rough Volatility and Fractional Models

For rough-Heston local volatility (HMLV) models driven by fractional (Volterra) kernels,

$$V_t = V_0 + \int_0^t K(t-s)[\kappa(\theta - V_s) ds + \nu\sqrt{V_s} dW_s],$$

the leverage function in the Markovian lift is again computed via the conditional mean, preserving vanilla calibration by design (Dall'Acqua et al., 2022). Exact small-time asymptotics, such as the $H+\frac{3}{2}$ skew rule relating local and implied volatility skews, are explicitly preserved by interpolation/extrapolation schemes in the leverage function grid (Dall'Acqua et al., 2022).

4.3 Scale-Invariant and Pathwise Local Volatility

Alternative LSV approaches such as the dimensionless (“scale-invariant”) formulation model “surprise” indices $I_t = S_t/M_t$ with relative kernels to achieve regime stability and robust scenario generation. The leverage function is constructed as a function of the scale-invariant index and calibrated directly to regime-dependent volatilities such as VIX, leading to parsimonious and stable parameterizations (Lipton et al., 2023).

5. Analytical and Approximate Solutions

For certain classes of LSV and SLV models, explicit or semi-explicit pricing approaches are feasible. The Wei-Norman factorization combined with Lie algebra techniques reduces time-dependent, non-autonomous PDEs to products of (shifted, mixed) heat kernels, yielding analytically tractable fundamental solutions for European payoffs (Guerrero et al., 2022). In multifactor expansions, closed-form asymptotic pricing and implied volatility expressions can be obtained via polynomial expansion and Dyson series for European options, with rigorous error bounds under uniform ellipticity (Lorig et al., 2013).

Short-maturity asymptotics for both spot and variance options are accessible in local-stochastic volatility models, with rate functions for option prices derived from two-dimensional variational or Hamilton-Jacobi problems. These provide explicit leading-order expansions for implied volatility smiles and realized-variance options, applicable for both Heston- and SABR-type processes (Pirjol et al., 2024, Pirjol et al., 2024).

6. Well-posedness, Regularization, and Theoretical Analysis

The McKean-Vlasov structure underlying LSV calibration introduces significant analytical complexity because the SDE coefficients depend on conditional laws or functionals thereof. Strong existence and uniqueness for the two-factor LSV SDE, specifically when the volatility factor is finite-valued and with regular leverage functions, are established under careful ellipticity and regularity conditions (Mustapha, 2024). Regularizations using reproducing kernel Hilbert spaces (RKHS) are introduced for singular models to guarantee well-posedness and propagation of chaos for the particle system, ensuring the validity of the numerical calibration algorithms (Bayer et al., 2022).

In more general settings where coefficients are only L^1-continuous in law, existence can still be established using compactness and fixed-point techniques on the Fokker–Planck system, though uniqueness may not hold without further assumptions (Djete, 2022). These results highlight the subtlety of model selection and regularization for ensuring both theoretical soundness and effective practical calibration.

7. Applications and Impact in Derivative Pricing and Risk

LSV models are essential for exotic derivative pricing, accurate replication of observed volatility surfaces, path-dependent product valuation (e.g., barrier, Asian, autocallable products), and P&L attribution. They enable robust scenario generation and risk margin calculation (e.g., VaR, expected shortfall), with scale-invariant and relative-parameter models enhancing stability across regimes (Lipton et al., 2023). Hybrid models that incorporate stochastic rates (LV2SR, SLV2SR, LSV–CIR++) are crucial for currency and rate-sensitive products, requiring careful calibration to maintain no-arbitrage conditions in both spot and interest rate markets (Cozma et al., 2017, Ogetbil et al., 2020, Cozma et al., 2015).

The combination of theoretical calibration guarantees, efficient simulation methods, and variance reduction techniques establishes LSV models as the preferred industry standard for high-dimensional risk management and exotic pricing engines.

---

References:

• (Saporito et al., 2017) "The Calibration of Stochastic-Local Volatility Models - An Inverse Problem Perspective"
• (Bayer et al., 2022) "A Reproducing Kernel Hilbert Space approach to singular local stochastic volatility McKean-Vlasov models"
• (Djete, 2022) "Non--regular McKean--Vlasov equations and calibration problem in local stochastic volatility models"
• (Mustapha, 2024) "Strong existence and uniqueness of a calibrated local stochastic volatility model"
• (Friz et al., 12 Jun 2025) "On the Weak Error for Local Stochastic Volatility Models"
• (Cai et al., 29 Sep 2025) "Simulation of the Heston stochastic local volatility model: implicit and explicit approaches"
• (Ogetbil et al., 2020) "Calibrating Local Volatility Models with Stochastic Drift and Diffusion"
• (Cozma et al., 2015) "Convergence of an Euler scheme for a hybrid stochastic-local volatility model with stochastic rates in foreign exchange markets"
• (Dall'Acqua et al., 2022) "Rough-Heston Local-Volatility Model"
• (Guerrero et al., 2022) "Stochastic Local Volatility models and the Wei-Norman factorization method"
• (Halperin et al., 2013) "USLV: Unspanned Stochastic Local Volatility Model"
• (Lorig et al., 2013) "Explicit implied volatilities for multifactor local-stochastic volatility models"
• (Lipton et al., 2023) "SPX, VIX and scale-invariant LSV"
• (Pirjol et al., 2024) "Short-maturity options on realized variance in local-stochastic volatility models"
• (Pirjol et al., 2024) "Short-maturity asymptotics for VIX and European options in local-stochastic volatility models"
• (Bank et al., 2023) "Rough PDEs for local stochastic volatility models"
• (Guo et al., 2019) "Calibration of Local-Stochastic Volatility Models by Optimal Transport"
• (Cozma et al., 2017) "Calibration of a Hybrid Local-Stochastic Volatility Stochastic Rates Model with a Control Variate Particle Method"
• (Bally et al., 2010) "Bounds on Stock Price probability distributions in Local-Stochastic Volatility models"