05 · Hybrid Volatility · Exotics Pricing

The Stochastic
Local Volatility Model.

A study of the SLV framework, which blends a stochastic volatility process with a local volatility leverage function so that the model reprices the vanilla smile exactly while retaining realistic forward-smile dynamics for exotic and path-dependent payoffs.

Context Personal study — Hybrid volatility modelling

Subject Exotic and barrier option pricing

Model Stochastic Local Volatility (SLV)

Method Leverage function via particle methods

01 / Introduction

The best of both worlds.

Local volatility models, such as Dupire's, match the entire vanilla smile exactly by construction, but generate a forward smile that flattens unrealistically fast with time. Stochastic volatility models, such as Heston, produce persistent and realistic forward smiles, but generally cannot fit every strike and maturity of a given surface exactly without an unwieldy number of parameters.

The stochastic local volatility (SLV) framework combines both ingredients: a genuinely stochastic variance process supplies realistic dynamics, while a deterministic leverage function, calibrated on top of it, forces the model to reproduce the market smile exactly. This hybrid construction has become an industry standard for pricing exotic and path-dependent payoffs that are sensitive to forward-smile behaviour.

02 / Dynamics

The SLV stochastic differential equations.

Under the risk-neutral measure \(\mathbb{Q}\), the underlying price \(S_t\) and its variance \(v_t\) follow:

Underlying dynamics

\[ dS_t = (r-q)S_t\,dt + L(S_t,t)\sqrt{v_t}\,S_t\,dW_t^{S} \]

Variance dynamics

\[ dv_t = \kappa(\theta-v_t)\,dt + \xi\sqrt{v_t}\,dW_t^{v} \]

Correlation

\[ d\langle W^S, W^v \rangle_t = \rho\,dt \]

The variance process is typically taken to be a standard stochastic volatility model such as Heston, with parameters \((v_0,\kappa,\theta,\xi,\rho)\) calibrated to capture the realistic term structure of the forward smile. The function \(L(S,t)\), known as the leverage function, is the new ingredient of the SLV model: it scales the instantaneous volatility deterministically as a function of spot and time.

03 / Leverage Function

Forcing an exact fit to the smile.

The role of the leverage function is to correct the smile generated purely by the stochastic variance process so that the model matches the market-implied local volatility exactly at every point. Gyongy's mimicking theorem shows that any one-factor diffusion sharing the same marginal distributions as \(S_t\) must have a diffusion coefficient equal to Dupire's local volatility. Applying this to the SLV process yields the defining relation of the leverage function:

Leverage function calibration condition

\[ L(K,T)^2 = \frac{\sigma_{\mathrm{loc}}^2(K,T)} {\mathbb{E}^{\mathbb{Q}} \left[v_T \,\middle|\, S_T=K\right]} \]

Here \(\sigma_{\mathrm{loc}}^2(K,T)\) is the Dupire local variance implied by the vanilla market smile, and the denominator is the conditional expectation of the stochastic variance given that the underlying is at level \(K\) at time \(T\). By construction, the SLV model's own implied local volatility, averaged over the paths of \(v_t\), coincides exactly with the market's local volatility surface, so the model reprices every European vanilla exactly.

04 / Calibration

A circular problem solved by particle methods.

The calibration condition is circular: computing \(L(K,T)\) requires the conditional expectation \(\mathbb{E}^{\mathbb{Q}}[v_T\mid S_T=K]\), which itself depends on the law of \(S_T\) under the SLV dynamics, which in turn depends on \(L\). No closed-form solution exists in general, and the leverage function is instead built numerically, forward in time.

The most widely used approach, following Guyon and Henry-Labordère (2012), is a particle method. A large number of paths \((S_t^{(i)},v_t^{(i)})\) are simulated jointly on a discrete time grid. At each time step \(T_j\), the conditional expectation is estimated from the simulated particles using a kernel or bucket regression:

Particle estimator of the conditional expectation

\[ \mathbb{E}^{\mathbb{Q}} \left[v_{T_j}\,\middle|\,S_{T_j}=K\right] \approx \frac{ \sum_{i=1}^{M} v_{T_j}^{(i)}\, \phi_h\!\left(S_{T_j}^{(i)}-K\right) }{ \sum_{i=1}^{M} \phi_h\!\left(S_{T_j}^{(i)}-K\right) } \]

where \(\phi_h\) is a smoothing kernel of bandwidth \(h\). The leverage function at \(T_j\) is then computed from this estimate and the independently pre-computed Dupire local volatility surface, and used to simulate the particles one step further to \(T_{j+1}\). The procedure is repeated forward across the time grid, so the leverage surface is built one maturity slice at a time, each slice depending on the particle distribution obtained at the previous one.

05 / Application

Why exotics need SLV.

A vanilla European option only depends on the terminal distribution of the underlying, so a pure local volatility model, which matches that distribution by construction, prices it correctly. Barrier options, cliquets, and other path-dependent or forward-starting payoffs, however, are sensitive to the joint dynamics of spot and volatility along the path, and in particular to the shape of the forward smile — a feature local volatility gets structurally wrong.

Because SLV inherits its forward-smile dynamics from the underlying stochastic volatility process while still matching the vanilla surface exactly, it is generally regarded as the more reliable choice for these products, and remains a widely used industry benchmark for equity and FX exotics desks.

06 / Limitations

Numerical pitfalls and model limitations.

Particle noise near the wings

Where few simulated particles land near a given strike — typically deep in the wings or at short maturities — the kernel regression estimating the conditional expectation becomes noisy, and the resulting leverage function can be unstable or require heavy regularization.

Regression noise

Computational cost

Building the leverage surface requires simulating a large number of joint spot-variance paths forward in time, re-estimating a nonparametric regression at every time step, which is considerably more expensive than calibrating Dupire or Heston in isolation.

Simulation cost

Sensitivity to the underlying stochastic vol model

The leverage function corrects the marginal distributions but not the joint dynamics; the choice of the base stochastic volatility process (e.g. Heston parameters) still materially affects exotic prices even after calibration, so its own parameters must be chosen carefully rather than fit arbitrarily.

Model risk

Bias from discretization

Because the leverage function is built step by step on a discrete time grid, the fit to the vanilla surface is only exact in the continuous-time limit; coarse time grids introduce a discretization bias that must be controlled by grid refinement.

Time discretization

07 / Summary

Exact smile fit with realistic dynamics.

01 A stochastic variance process, typically Heston, supplying realistic forward-smile dynamics.

02 A deterministic leverage function, grounded in Gyongy's mimicking theorem, forcing an exact fit to the market's Dupire local volatility surface.

03 A circular calibration condition solved numerically, most commonly through a forward particle method.

04 A practical benchmark for pricing barrier, cliquet, and other path-dependent payoffs sensitive to forward-smile behaviour.

05 Structural limitations: particle-regression noise, computational cost, and residual sensitivity to the choice of the underlying stochastic volatility model.

SLV does not replace Heston or Dupire; it composes them. The stochastic factor supplies dynamics, the leverage function supplies exactness, and the resulting hybrid remains one of the most widely used frameworks for consistent vanilla-and-exotic pricing on a single desk.

← Back to Projects

The Stochastic Local Volatility Model.