Heston Model Explained: Stochastic Volatility for Option Pricing 2026

What Is the Heston Model?

The Heston model is a stochastic volatility model published by Steven Heston in 1993 that extends the Black-Scholes model by treating volatility as a random process rather than a fixed constant. It captures real-world pricing patterns - like the volatility smile - that Black-Scholes fundamentally cannot produce.

In Black-Scholes, you plug in a single volatility number and it applies to every strike price and expiry. The Heston model rejects this simplification. Instead, it models the variance of the stock price as its own stochastic differential equation, driven by its own source of randomness. The stock price process and its variance process are correlated, meaning a drop in the stock price can coincide with a rise in volatility - exactly what we observe in equity markets.

The model's popularity in 2026 stems from a powerful combination: it's realistic enough to capture the key features of market-implied volatility surfaces, yet it has a semi-analytical solution that makes pricing and calibration computationally efficient. Most stochastic volatility models require full Monte Carlo simulation for option pricing. The Heston model doesn't. Its characteristic function approach lets you price European options using numerical integration, which is orders of magnitude faster.

Since its publication, the Heston model has become one of the most widely cited and implemented models in quantitative finance. It serves as a baseline stochastic volatility framework on trading desks, in academic research, and in risk management systems across the industry. If you're working with option pricing models beyond Black-Scholes, the Heston model is almost certainly the first one you'll encounter.

---

Why Do We Need the Heston Model?

We need the Heston model because Black-Scholes assumes constant volatility, which produces a flat implied volatility surface. Real markets show pronounced smiles and skews across strikes and maturities, and any model used for serious option pricing must account for this.

If you've ever looked at implied volatilities across different strike prices for the same expiry, you've seen the volatility smile. Out-of-the-money puts on equity indices consistently trade at higher implied volatility than at-the-money options. Black-Scholes has no mechanism to produce this pattern. Under its assumptions, one volatility number applies everywhere - the implied volatility surface would be completely flat.

This isn't just an academic problem. If your pricing model can't reproduce the volatility smile, it will systematically misprice options away from the money. It'll also give you incorrect Greeks - your delta hedge will be wrong, your vega exposure will be miscalculated, and your risk management will rest on a flawed picture of reality.

The Heston model addresses this by introducing two key features:

Stochastic variance. The variance follows a mean-reverting process with its own randomness. This means the distribution of future stock prices has fatter tails than a lognormal distribution - extreme moves are more likely, which pushes up the implied volatility of far-from-the-money options.

Correlation between price and variance. When the stock falls, variance tends to increase - a pattern sometimes called the asymmetric volatility effect. This negative correlation produces the skew we see in equity markets - implied volatility is higher at lower strikes than at higher strikes.

Together, these two features let the Heston model generate implied volatility surfaces that closely match what we observe in real markets. That's why it has been a workhorse of quantitative finance since its publication and remains widely used in 2026.

---

The Heston Model Equations

The Heston model is defined by two coupled stochastic differential equations - one for the stock price and one for the variance. These two processes evolve together, linked by a correlation parameter that controls how price movements and volatility changes relate to each other.

The stock price process:

dS = mu * S * dt + sqrt(v) * S * dW_S

The variance process:

dv = kappa * (theta - v) * dt + xi * sqrt(v) * dW_v

The two Brownian motions W_S and W_v are correlated:

Corr(dW_S, dW_v) = rho * dt

Here's what each component means:

• S - the stock price at time t
• v - the instantaneous variance at time t (variance, not volatility - variance is volatility squared)
• mu - the drift rate of the stock (expected return)
• kappa - the speed of mean reversion for the variance process
• theta - the long-run equilibrium variance level
• xi - the volatility of variance, often called "vol of vol"
• rho - the correlation between the stock's Brownian motion and the variance's Brownian motion
• v0 - the initial variance at time zero

The stock price equation looks similar to geometric Brownian motion, except that the volatility term sqrt(v) is now random rather than constant. The variance equation is a Cox-Ingersoll-Ross (CIR) process - it's mean-reverting (the kappa * (theta - v) term pulls variance back toward theta) and the sqrt(v) factor ensures variance stays non-negative, provided the Feller condition holds: 2 * kappa * theta > xi^2.

For option pricing under the risk-neutral measure, the drift mu is replaced by the risk-free rate r:

dS = r * S * dt + sqrt(v) * S * dW_S

This is the version used for pricing and calibration. The real-world drift mu doesn't affect option prices - only the risk-free rate, the volatility dynamics, and the correlation matter.

---

Understanding the Parameters

Each of the Heston model's five parameters controls a specific aspect of the variance dynamics and the resulting implied volatility surface. Calibrating the model means finding the parameter combination that best matches observed market prices, and understanding what each parameter does is essential for interpreting the results.

Initial variance (v0). This sets the starting point for the variance process. If v0 = 0.04, that corresponds to an initial volatility of 20% (since volatility = sqrt(variance)). It directly controls the at-the-money implied volatility level for short-dated options.

Long-run variance (theta). This is the level that variance gravitates toward over time. If v0 is below theta, the model predicts variance will tend to increase, and vice versa. The difference between v0 and theta shapes the term structure of at-the-money implied volatility - whether it slopes upward, downward, or stays flat.

Mean reversion speed (kappa). This controls how quickly variance reverts to theta after a shock. A high kappa means shocks to variance dissipate quickly - the volatility surface looks more "normal" at longer maturities. A low kappa means volatility shocks persist, keeping the smile pronounced across the term structure. This parameter has a strong influence on the slope of the ATM volatility term structure.

Vol of vol (xi). This is the volatility of the variance process itself. Higher xi means more randomness in variance, which fattens the tails of the stock price distribution. On the implied volatility surface, this shows up as greater convexity - both wings of the smile are pushed higher relative to the centre. If you set xi to zero, there's no randomness in variance, and the model collapses back to Black-Scholes.

Correlation (rho). This is the most important parameter for equity applications. For stocks and equity indices, rho is typically negative, commonly in the range of -0.7 to -0.9. Negative rho means that when the stock price falls, variance tends to rise - producing the characteristic skew where lower strikes have higher implied volatility. Setting rho to zero produces a symmetric smile. Positive rho (rare in equities but sometimes seen in commodities) produces a reverse skew where higher strikes have higher implied volatility.

The interplay between these parameters is what gives the Heston model its flexibility. Rho controls the tilt of the smile, xi controls its curvature, and kappa and theta control how the smile evolves across maturities. Fitting these five numbers to a market surface of hundreds of option prices is the calibration problem.

---

The Heston Model Characteristic Function

The key advantage of the Heston model over many other stochastic volatility models is that it has a known characteristic function in closed form. This allows European options to be priced efficiently using Fourier inversion, without needing Monte Carlo simulation.

The characteristic function gives you the expected value of exp(i * u * ln(S_T)) under the risk-neutral measure, where u is a complex argument and i is the imaginary unit. Once you have this function, the price of a European call or put can be computed by evaluating a one-dimensional integral - specifically through the method developed by Carr and Madan (1999) or the original formulation by Heston.

In practice, this means pricing a single European option under the Heston model takes milliseconds rather than the seconds required by Monte Carlo. When you're calibrating the model and need to reprice hundreds or thousands of options at each optimisation step, this speed advantage is critical.

The characteristic function approach works because the stochastic differential equations governing the Heston model have an affine structure - the drift and variance of the log-price are affine (linear) functions of the state variables. This mathematical structure makes the characteristic function tractable. Not all stochastic volatility models share this property, which is one reason the Heston model specifically became so popular.

For practical implementation, the most common approach is to use the formulation from Albrecher et al. (2007) or the "little Heston trap" correction, which avoids numerical issues with branch cuts in the complex logarithm that can affect naive implementations. Getting the integration contour and damping parameter right is essential for stable pricing.

The Fourier pricing approach also extends naturally to pricing other payoffs - digital options, power options, and variance swaps can all be handled through appropriate modifications of the integration formula.

---

Calibrating the Heston Model

Calibrating the Heston model means finding the five parameters (v0, theta, kappa, xi, rho) that produce model prices matching observed market option prices as closely as possible. It's an optimisation problem where you minimise the distance between theory and reality.

The standard workflow looks like this:

1. Collect market data. Gather a set of European option prices (or implied volatilities) across multiple strikes and maturities. A typical calibration set might include 50 to 200 options spanning the liquid part of the volatility surface.

2. Define the objective function. The most common choice is to minimise the sum of squared differences between model and market implied volatilities:

Objective = Sum over all options of (IV_model - IV_market)^2

Some practitioners weight the errors - giving more weight to at-the-money options where liquidity is highest, or weighting by vega to normalise for each option's sensitivity to volatility.

3. Choose an optimiser. The objective function is non-convex with multiple local minima, so gradient-based methods alone can get stuck. Common approaches include:

• Differential evolution or particle swarm optimisation for global search
• Levenberg-Marquardt or L-BFGS-B for local refinement
• A two-stage approach: global search to find a good starting region, then local optimisation to fine-tune

4. Apply constraints. Parameters must satisfy economic constraints:

• v0, theta, xi > 0
• kappa > 0 (variance reverts toward theta, not away from it)
• -1 < rho < 1
• The Feller condition 2 * kappa * theta > xi^2 is desirable but not always enforced strictly

5. Validate. Check that the calibrated model reproduces the market surface well, not just at the calibration strikes but also at intermediate points. Plot the model-implied surface against the market surface and inspect the residuals.

A common pitfall is overfitting to a single maturity. The Heston model has five parameters for the entire surface, so it can't perfectly match every point. The goal is to capture the broad shape of the smile across maturities, not to eliminate every last basis point of error. If you need a perfect fit to today's surface, a local volatility or local-stochastic volatility hybrid is more appropriate. The Heston model's strength is that its dynamics - how the smile evolves over time - are more realistic than local volatility alone.

In 2026, most calibration implementations use a combination of Python (scipy.optimize, or dedicated libraries like QuantLib) and fast characteristic function evaluation. A full calibration typically runs in seconds to minutes depending on the size of the data set and optimiser settings.

---

Implementing the Heston Model in Python

Here's a practical Python implementation of the Heston model using Monte Carlo simulation. This generates correlated paths for both the stock price and variance, then prices a European call option by averaging the discounted payoffs.

import numpy as np


def heston_monte_carlo(
    S0: float,
    v0: float,
    r: float,
    kappa: float,
    theta: float,
    xi: float,
    rho: float,
    T: float,
    K: float,
    n_steps: int = 252,
    n_paths: int = 100_000,
    seed: int = 42,
) -> dict:
    """
    Price a European call option under the Heston stochastic
    volatility model via Monte Carlo with a log-Euler scheme.
    """
    rng = np.random.default_rng(seed)
    dt = T / n_steps

    S = np.full(n_paths, S0, dtype=np.float64)
    v = np.full(n_paths, v0, dtype=np.float64)

    for _ in range(n_steps):
        Z1 = rng.standard_normal(n_paths)
        Z2 = rng.standard_normal(n_paths)

        # Correlated Brownian increments via Cholesky decomposition
        W_S = Z1
        W_v = rho * Z1 + np.sqrt(1 - rho**2) * Z2

        # Prevent negative variance (absorption at zero)
        v_pos = np.maximum(v, 0.0)

        # Log-Euler step for the stock price
        S = S * np.exp(
            (r - 0.5 * v_pos) * dt + np.sqrt(v_pos * dt) * W_S
        )

        # Euler step for the variance process
        v = v + kappa * (theta - v_pos) * dt + xi * np.sqrt(v_pos * dt) * W_v

    # Discounted payoff for a European call
    payoffs = np.maximum(S - K, 0.0)
    call_price = np.exp(-r * T) * np.mean(payoffs)
    std_error = np.exp(-r * T) * np.std(payoffs) / np.sqrt(n_paths)

    return {
        "call_price": round(call_price, 4),
        "std_error": round(std_error, 4),
        "n_paths": n_paths,
        "n_steps": n_steps,
    }


# --- Price a European call under the Heston model ---
result = heston_monte_carlo(
    S0=100,         # spot price
    v0=0.04,        # initial variance (20% volatility)
    r=0.05,         # risk-free rate
    kappa=2.0,      # mean reversion speed
    theta=0.04,     # long-run variance
    xi=0.3,         # vol of vol
    rho=-0.7,       # negative correlation (equity skew)
    T=1.0,          # 1 year to expiry
    K=100,          # at-the-money strike
    n_steps=252,
    n_paths=200_000,
)

print(f"European Call Price: {result['call_price']:.4f}")
print(f"Standard Error:      {result['std_error']:.4f}")
print(f"Paths: {result['n_paths']:,}  |  Steps: {result['n_steps']}")

A few notes on this implementation:

Log-Euler scheme for the stock price. Instead of applying the Euler step directly to S, we exponentiate the log-return. This prevents the stock price from going negative - a problem that can occur with the naive Euler scheme when variance is large.

Absorption at zero for variance. The line v_pos = np.maximum(v, 0.0) prevents negative variance, which the Euler discretisation can produce when the Feller condition isn't satisfied. More sophisticated schemes like the Quadratic Exponential (QE) scheme by Andersen (2008) handle this more accurately and are preferred for production use.

Correlated Brownian motions. The Cholesky decomposition trick converts two independent standard normal samples Z1 and Z2 into correlated increments W_S and W_v with the desired correlation rho.

Extending the code. To generate a full implied volatility smile from the Heston model, run this simulation for a range of strike prices and then invert the resulting call prices through Black-Scholes to extract implied volatilities. This lets you visualise the smile that a given set of Heston parameters produces.

---

Heston Model vs Black-Scholes

The Heston model and Black-Scholes agree on the basic framework - risk-neutral pricing under no-arbitrage conditions - but differ fundamentally in how they treat volatility. This single difference has far-reaching consequences for pricing accuracy, hedging quality, and computational requirements.

When Black-Scholes is sufficient. For quick back-of-the-envelope pricing and for communicating option prices in terms of a standardised metric (implied volatility), Black-Scholes remains the universal language. Every options trader in 2026 still quotes in Black-Scholes implied volatility, even when they use more sophisticated models internally. For near-the-money options where the smile effect is small, Black-Scholes pricing is often close enough.

When you need the Heston model. Any situation where the volatility smile matters - pricing exotic options that depend on the full return distribution (barriers, digitals, variance swaps), managing a portfolio of options across different strikes, or building a consistent risk management framework across the book. If you're trading the skew or relative value across strikes, a model that actually produces a smile is essential.

The practical difference shows up most clearly in hedging. Under Black-Scholes, you compute delta and hedge with shares of the underlying. Under the Heston model, volatility itself is a source of risk. You need to think about vega hedging and how changes in the variance process affect your portfolio. This is more complex but also more realistic - volatility does move, and pretending it doesn't creates hidden exposures.

For a full treatment of Black-Scholes, see our Black-Scholes model guide.

---

Heston Model vs Other Stochastic Volatility Models

The Heston model isn't the only approach to stochastic volatility, and different models suit different use cases. Here's how the main alternatives compare and when you'd choose one over another.

SABR model. The SABR (Stochastic Alpha Beta Rho) model is the standard for interest rate options, where it provides a convenient closed-form approximation for implied volatility as a function of strike. SABR is typically calibrated to a single expiry at a time, while the Heston model calibrates across the full term structure simultaneously. For equities, the Heston model is more common. For swaptions and caps, SABR dominates.

Local volatility (Dupire). Local volatility defines a deterministic volatility function sigma(S, t) that exactly reproduces the observed implied volatility surface. The calibration is perfect by construction, but the forward smile dynamics are unrealistic - the model predicts that the smile flattens too quickly over time. The Heston model sacrifices perfect calibration for more realistic dynamics of how the smile evolves.

Bates model (Heston with jumps). The Bates model adds a jump-diffusion component to the Heston framework. This lets it capture the steep short-dated smile that pure stochastic volatility struggles with. Jumps matter most for short maturities; at longer horizons, the diffusive component dominates. The cost is two or three extra parameters (jump intensity, mean jump size, jump volatility).

Local-stochastic volatility (LSV) hybrids. These combine a local volatility component for exact calibration with a stochastic volatility component for realistic dynamics. LSV models are computationally expensive but represent the current standard on many exotic derivatives desks in 2026. The Heston model often serves as the stochastic volatility backbone in these hybrid approaches.

The choice depends on what you're pricing. For vanilla European options on equities, the Heston model offers the best balance of accuracy, speed, and interpretability. For path-dependent exotics where both calibration accuracy and forward dynamics matter, LSV hybrids are increasingly preferred. For more on how these models fit into the broader pricing picture, see our option pricing models guide.

---

Frequently Asked Questions

What is the Heston model used for?

The Heston model is primarily used for pricing options where the volatility smile matters. This includes European options across multiple strikes and expiries, variance swaps, and structured products with sensitivity to the full return distribution. It's also widely used for risk management, where accurate modelling of the volatility surface is essential for computing Greeks and managing portfolio exposures. Trading desks use calibrated Heston models to identify relative value across the options surface. In 2026, it remains one of the most commonly implemented stochastic volatility models in both academic research and industry practice, serving as a foundation for more complex hybrid approaches.

Is the Heston model better than Black-Scholes?

The Heston model is more realistic than Black-Scholes for capturing the volatility smile and the dynamics of real option markets, but "better" depends on context. Black-Scholes is simpler, faster, and universally understood - it's the standard language for communicating option prices. The Heston model is more accurate for pricing and hedging when the smile shape matters, but it requires five parameters instead of one and is more computationally demanding. For quick pricing and basic hedging of near-the-money options, Black-Scholes is often sufficient. For serious work across the strike and maturity spectrum, you'll want the Heston model or one of its extensions.

How do you calibrate the Heston model to market data?

Calibration involves collecting a set of market option prices across multiple strikes and maturities, then finding the five Heston parameters that minimise the difference between model and market prices. The standard approach uses sum-of-squared-errors in implied volatility space. Because the objective function has multiple local minima, most practitioners use a global optimiser (like differential evolution) to find a good starting region, then refine with a local optimiser (like Levenberg-Marquardt). Constraints ensure parameters stay in economically meaningful ranges. The process typically runs in seconds to minutes depending on data set size and optimiser configuration.

What is the Feller condition in the Heston model?

The Feller condition states that 2 * kappa * theta must be greater than xi^2, where kappa is the mean reversion speed, theta is the long-run variance, and xi is the vol of vol. When this condition holds, the variance process never reaches zero - it stays strictly positive at all times. When the condition is violated, variance can touch zero, which creates issues for some numerical discretisation schemes. In practice, many calibrated parameter sets violate the Feller condition slightly, particularly when xi is high. This isn't necessarily a problem for pricing - it just means you need to use a scheme that handles the zero-variance boundary properly, like the QE scheme or variance absorption techniques.

Can the Heston model capture all volatility smile shapes?

The Heston model is flexible but has limits. It does a good job capturing the broad features of equity volatility surfaces - downward-sloping skew driven by negative rho, curvature driven by xi, and term structure effects driven by the kappa and theta relationship. However, it can struggle with very steep short-dated smiles, which are often driven by jump risk that the basic Heston model doesn't include. The Bates extension (Heston plus jumps) addresses this. The model can also have difficulty fitting both short-dated and long-dated smiles simultaneously with a single parameter set. For cases where an exact fit is essential, local-stochastic volatility models are preferred.

How does the Heston model relate to Black-Scholes?

The Heston model reduces to Black-Scholes when volatility is constant. Specifically, if you set xi = 0 (no randomness in variance), kappa to any positive value, and theta = v0 (variance starts at its long-run level and stays there because there's no randomness to push it away), then v remains constant at v0 for all time. The stock price SDE becomes dS = r * S * dt + sqrt(v0) * S * dW, which is exactly the Black-Scholes geometric Brownian motion with constant volatility sigma = sqrt(v0). Black-Scholes is a special case of the Heston model - and understanding this connection helps clarify what the extra parameters actually buy you.