Stochastic Volatility: Heston Model Explained

Published: April 12, 2026

Article by Ryan O'Connell, CFA, FRM

Options traders know that implied volatility varies across strikes — deep out-of-the-money puts typically trade at higher implied volatility than at-the-money options. This pattern, known as the volatility smile or skew, cannot be explained by the Black-Scholes model, which assumes a single constant volatility. The Heston model (1993) addresses this limitation by treating volatility as a random process with its own dynamics, correlated with the underlying asset price. This article explains how stochastic volatility models work, what each Heston parameter controls, and why negative correlation generates the characteristic equity skew.

What Is a Stochastic Volatility Model?

A stochastic volatility model is a class of options pricing framework where volatility is itself a random variable driven by its own stochastic process, rather than a fixed parameter. Instead of specifying a single number for volatility as Black-Scholes does, these models give volatility its own equation — it can rise, fall, cluster during crises, and revert toward a long-run level over time.

The Heston model (1993) is the workhorse of stochastic volatility modeling because it offers semi-closed-form pricing via characteristic functions. While the pricing formula still requires numerical integration, this is far more tractable than full Monte Carlo simulation. Other stochastic volatility models include SABR (popular in interest rate markets), the Bates model (Heston plus jumps), and more recent rough volatility models.

The key intuition is that in real markets, fear (volatility) itself rises and falls — it clusters, spikes during market stress, and slowly reverts back to normal levels. Stochastic volatility models capture this behavior directly, while constant-volatility models like Black-Scholes cannot.

Why Constant Volatility Fails

The Black-Scholes model assumes volatility is a single, constant number. Under this assumption, all options on the same underlying with the same expiration should have the same implied volatility — a perfectly flat volatility surface.

Empirical reality tells a different story. When you back out implied volatility from market prices using the Black-Scholes formula, you get different numbers for different strikes — proving the constant-volatility assumption is wrong. Specifically:

• Fat tails: Black-Scholes uses a lognormal distribution, but actual returns have fatter tails — extreme moves happen more frequently than the model predicts.
• Leverage effect: Equity volatility tends to rise when prices fall. Black-Scholes has no mechanism for this correlation.
• Volatility clustering: High volatility begets high volatility. Periods of calm and periods of turbulence cluster together. Black-Scholes ignores this persistence entirely.

Heston Model Formula: Mean-Reverting Variance

Steven Heston published “A Closed-Form Solution for Options with Stochastic Volatility” in 1993. The key innovation was finding an analytically tractable solution using characteristic functions, despite volatility being random. The model uses two coupled stochastic differential equations (SDEs).

Where (r – q) is the risk-free rate minus dividend yield — the risk-neutral drift used for pricing. The diffusion term uses √v instead of constant σ, where v(t) is the instantaneous variance at time t.

This is a Cox-Ingersoll-Ross (CIR) type process. When variance v is above its long-run level θ, the drift is negative (pulling v down). When v is below θ, the drift is positive (pulling v up). The √v scaling in the diffusion term ensures variance remains non-negative.

The two Brownian motions are correlated:

The Feller condition provides a sufficient condition for variance to stay strictly positive:

Think of variance as a rubber band stretched toward its long-run level θ. The parameter κ controls how hard the band pulls back, and ξ controls how much it quivers along the way. The stock price responds to wherever the rubber band currently sits.

Heston Model Parameters

The Heston model has five parameters. Understanding what each one controls is essential for interpreting calibration results and model behavior.

κ (kappa) — Mean reversion speed: High κ means variance snaps back to θ quickly; the process is tightly anchored. Low κ means variance can drift far from its long-run level for extended periods.

θ (theta) — Long-run variance: This is the level variance gravitates toward over time. If θ = 0.04, then long-run volatility is √0.04 = 20%.

ξ (xi) — Volatility of volatility: Controls how much variance itself fluctuates. Higher ξ produces fatter tails and more pronounced wings in the volatility surface. Note that ξ governs curvature; it does not by itself create skew.

ρ (rho) — Correlation: For equities, ρ is empirically negative (typically -0.5 to -0.8). When ρ < 0, stock price falls tend to coincide with variance rises — the leverage effect. This is the single most important parameter for generating realistic equity skew. Positive ρ can occur in some commodity markets where price rallies are associated with rising volatility.

v₀ — Initial variance: The starting value for the variance process, typically calibrated jointly with other parameters to fit the current volatility surface.

Important: Calibrated Heston parameters are risk-neutral pricing parameters, not historical estimates. They are derived from option prices under the risk-neutral measure Q, not from historical return data under the real-world measure P. The risk-neutral parameters embed a volatility risk premium that may cause them to differ significantly from what you would estimate from realized returns.

How the Heston Model Explains the Volatility Smile

The volatility smile and skew article documents what traders observe empirically — the shape of the implied volatility surface across strikes and maturities. This section explains why the Heston model produces these patterns mechanistically.

The mechanism depends primarily on the correlation parameter ρ (though skew magnitude also depends on ξ and maturity):

• ρ = 0: Heston produces a symmetric smile — both wings (deep OTM puts and calls) trade at higher implied volatility than ATM. This is pure kurtosis from ξ > 0.
• ρ < 0: The smile tilts left into a skew — OTM puts become more expensive (higher IV) than OTM calls. This is the characteristic equity skew.

The economic intuition: investors price crash risk. When markets fall sharply, volatility spikes simultaneously (because ρ < 0 links downward stock moves to upward variance moves). Put options that pay off in crashes therefore cost more — they pay off precisely when volatility is elevated, making them more valuable as insurance.

The term structure of the smile is governed by κ and θ jointly. High κ means short-dated smiles are steep but long-dated smiles flatten toward the long-run level θ. Low κ means the smile shape persists more uniformly across maturities.

Stochastic Volatility Model vs Local Volatility vs GARCH

Stochastic volatility is one of several approaches to modeling non-constant volatility. Each framework has distinct strengths and use cases.

Stochastic vs Local Volatility: Dupire’s local volatility model fits any observed vanilla surface exactly by construction — which sounds ideal. The problem is that its forward smile dynamics are unrealistic; in local vol, the smile tends to flatten as time passes in ways that don’t match observed behavior. Heston’s smile can shift and persist more realistically. In practice, many trading desks use Stochastic Local Volatility (SLV) models that blend both approaches.

Stochastic Volatility vs GARCH: GARCH models live in discrete time and are estimated from historical price data. They excel at econometric forecasting of realized volatility and Value at Risk calculations. Stochastic volatility models live in continuous time and are calibrated to option prices under the risk-neutral measure. They answer different questions: GARCH asks “what will volatility be next week?” while Heston asks “what is the risk-neutral distribution for pricing options today?”

Practical Challenges with Stochastic Volatility Models

While Heston improves significantly on Black-Scholes, it comes with practical difficulties:

Calibration complexity: Five parameters must be jointly optimized to fit the market volatility surface. The optimization problem is non-convex — multiple parameter sets can produce similar vanilla fits but differ significantly in their exotic pricing and hedging behavior.

Computational cost: Unlike Black-Scholes, Heston does not have a simple closed-form price. The semi-analytical solution requires numerical integration of a characteristic function (Fourier inversion). This is slower than Black-Scholes but far faster than full Monte Carlo simulation.

Greeks and hedging: In Heston, you need to hedge not just delta (∂V/∂S) but also vega (∂V/∂v). Because variance v is not directly observable or tradeable, vega hedging requires using other traded options as instruments. The hedge ratios are more complex than Black-Scholes.

Parameter instability: Heston parameters calibrated to today’s surface may differ significantly from parameters calibrated to yesterday’s surface. The model is not necessarily stable across time — this is the “recalibration problem.”

Limitations of the Heston Model

Even with stochastic volatility, the Heston model has fundamental limitations:

Single-factor variance: Heston uses one stochastic variance factor. Real volatility surfaces exhibit multi-factor behavior — short-dated and long-dated smiles can move independently. Extensions like time-dependent Heston, double Heston, or the Bates model address some of these limitations.

No jumps: Heston assumes continuous price paths. Equity returns exhibit sudden jumps from overnight gaps, earnings announcements, or market events. The Bates model (1996) extends Heston with a Poisson jump component to capture this behavior.

Rough volatility: Recent empirical research by Gatheral and Rosenbaum finds that realized volatility has fractal-like “rough” behavior not captured by the smooth CIR process. Rough volatility models like rBergomi better fit certain features of the ATM skew term structure.

Feller condition trade-off: Calibrations that impose the Feller condition for theoretical cleanliness often produce worse fits to the observed surface. Practitioners frequently allow Feller violations, accepting the limitation for better empirical accuracy.

Common Mistakes

Understanding these common errors helps avoid pitfalls when working with stochastic volatility models:

1. Confusing ρ and ξ effects on the smile
Mistake: Assuming that vol-of-vol (ξ) is responsible for the skew.
Correction: ξ controls the curvature (wings) of the smile, not the tilt. The tilt — the asymmetry between put and call implied vols — comes from ρ < 0. A model with ξ > 0 but ρ = 0 produces only a symmetric smile.

2. Confusing risk-neutral and real-world parameters
Mistake: Using historically estimated variance parameters directly for option pricing, or interpreting calibrated parameters as forecasts of future realized volatility.
Correction: Calibrated Heston parameters are risk-neutral (Q-measure) quantities derived from option prices. They embed a volatility risk premium and may differ substantially from P-measure parameters estimated from historical returns. Use the right parameters for the right purpose.

3. Calibrating to a single expiry only
Mistake: Calibrating Heston parameters to fit the smile for one expiry and using the result for all maturities.
Correction: Heston has a specific term structure of volatility implied by κ and θ. Calibrating to a single slice can produce parameters that wildly misprice options at other maturities. Always calibrate to a multi-expiry surface.

4. Ignoring the Feller condition in simulation
Mistake: Allowing the Feller condition to be violated in Monte Carlo simulation without implementing a fix for near-zero or negative variance.
Correction: Even if calibration violates Feller for better fit, the simulation discretization must handle near-zero variance paths explicitly. Common fixes include the full truncation scheme or the QE (Quadratic Exponential) scheme from Andersen (2007).