What Is the Heston Model?

Introduction

The Heston model is an option-pricing model that assumes volatility is not constant, but itself moves randomly over time. That sounds like a small adjustment to Black–Scholes, but it changes the logic of the model in exactly the place markets most clearly disagree with the constant-volatility assumption: the shape of implied volatility across strikes and maturities.

The puzzle the model is trying to solve is easy to state. If Black–Scholes were literally true, then all European options on the same underlying and expiry would imply the same volatility. In real markets, they do not. Out-of-the-money puts often imply higher volatility than at-the-money options, and the pattern changes with maturity. Traders call these patterns the volatility smile or skew. The Heston model exists because a constant-volatility world has too little structure to generate those shapes.

What makes Heston important is not only that it adds stochastic volatility, but that it does so in a way that remains unusually workable. In the original 1993 paper, Steven Heston showed that European option prices could still be computed in closed-form style through the model’s characteristic function, rather than by solving a full high-dimensional partial differential equation numerically every time. That tractability is why the model became a standard reference point in equities, FX, and other derivatives markets.

The central idea is simple enough to hold in your head: the asset price moves randomly, and its variance also moves randomly, pulling future uncertainty up or down as conditions change. Once volatility becomes a state variable rather than a fixed input, option prices stop being forced onto the Black–Scholes surface. Correlation between price shocks and volatility shocks then gives the model a way to tilt the distribution, producing skew rather than just fatter tails.

Why Black–Scholes fails to explain the implied-volatility skew

Black–Scholes assumes that the underlying follows geometric Brownian motion with constant volatility. That assumption is powerful because it compresses the entire future distribution into a single volatility number. But it is also restrictive for the same reason. If volatility is fixed, then the risk-neutral distribution of returns has a shape that is too rigid: once maturity is chosen, changing strike just samples different parts of the same lognormal distribution.

That rigidity leads to a market contradiction. Traders observe that far-downside options often trade rich relative to Black–Scholes at any single volatility input. If you back implied volatilities out of market prices, you do not get a flat line across strikes. You get a surface. Something in the market is saying that downside states are not just lower-price states; they are often higher-uncertainty states as well.

Here is the mechanism Black–Scholes cannot express. In many markets, especially equities, price drops and volatility increases tend to arrive together. That means the distribution is not merely wide; it is asymmetric. A model with fixed volatility can widen or narrow the distribution, but it cannot make volatility itself react to price moves. So it cannot naturally generate the skew built into market option prices.

You can think of Heston as answering a specific question: what if volatility is not an input chosen once at time zero, but a random process that evolves alongside price? Once you ask that question, the implied-volatility surface is no longer an anomaly outside the model. It becomes a consequence of the model’s internal dynamics.

How do price and variance interact in the Heston model?

The Heston model specifies two coupled stochastic processes. The asset price S_t evolves with an instantaneous variance v_t, and that variance itself follows a mean-reverting square-root process. In standard notation, the dynamics are written as dS_t / S_t = (r - q) dt + sqrt(v_t) dW_1 and dv_t = κ(θ - v_t) dt + σ sqrt(v_t) dW_2, with correlation corr(dW_1, dW_2) = ρ.

The notation matters, but only after the intuition is clear. r is the risk-free rate, q is a dividend yield or foreign short rate depending on context, v_t is instantaneous variance, κ controls how fast variance mean-reverts, θ is the long-run variance level, σ is the volatility of variance (often called vol-of-vol) and ρ is the correlation between price and variance shocks. The Brownian motions dW_1 and dW_2 are the random drivers.

Each parameter exists for a reason. Mean reversion through κ(θ - v_t) prevents variance from drifting aimlessly forever. The square-root term sqrt(v_t) makes variance shocks scale with the current level of variance, which helps keep the process economically sensible and related to the Cox–Ingersoll–Ross family. The correlation ρ is the key lever for skew: when ρ is negative, downward moves in the asset tend to coincide with rising variance, which shifts probability weight in a way option markets often demand.

A useful way to read the model is to separate what changes from what pulls it back. The random shocks change both price and variance from instant to instant. The mean-reversion term pulls variance back toward θ. So the model says volatility clusters and wanders, but not without structure. It can spike, decay, and interact with price in a way constant-volatility models cannot express.

Heston parameters explained: v0, θ, κ, σ and ρ

The easiest way to misunderstand Heston is to treat the parameters as an arbitrary list. They are better understood as controls on the risk-neutral distribution’s shape.

The long-run variance θ mostly sets the level around which variance tends to live if you look far enough ahead. If θ is higher, the model wants more uncertainty in the medium and long term. The initial variance v0 controls where the process starts today, so it matters most for shorter maturities. If current market conditions are stressed, v0 can be much more important than θ for the front end of the surface.

The mean-reversion speed κ controls how quickly the distinction between today’s variance and long-run variance gets forgotten. If κ is high, shocks to variance decay quickly and the model pulls back toward θ aggressively. If κ is low, current variance persists, which affects term structure. This is why calibration often struggles to pin down κ cleanly from a sparse cross-section: it acts through time, not through a simple one-step effect.

The vol-of-vol parameter σ affects how variable the variance process itself is. In Heston’s original analysis, increasing σ tends to increase kurtosis (fatter tails) rather than skew if correlation is absent. In ordinary language, it creates more states where volatility wanders far from its mean, making extreme price outcomes more likely. This matters for both wings of the option surface.

The correlation ρ does a different job. It changes skewness. Negative ρ means bad price moves tend to come with higher future variance, which pushes downside options richer relative to upside options. This is one of the most important conceptual points in the entire model: vol-of-vol mainly fattens the distribution; correlation mainly tilts it. Traders often talk loosely about both as “surface shape” parameters, but they do not shape the surface in the same way.

Example: how negative correlation creates option skew under Heston

Suppose an equity index is trading quietly, but investors know that when the market sells off, volatility usually jumps. In the Heston framework, that means current variance v0 may be moderate, long-run variance θ may also be moderate, but the correlation ρ between return shocks and variance shocks is negative.

Now imagine a one-day downside shock. The spot drops, and because ρ is negative, the same random event tends to push variance upward. That matters because option value is not only about where spot ends up; it is about how uncertain the path ahead remains after today’s move. A lower spot combined with higher future variance makes downside protection especially valuable. Out-of-the-money puts are therefore priced not just for ending up in the money, but for landing in scenarios where volatility is elevated as well.

Keep extending that logic across many possible paths. Some paths drift mildly. Some sell off sharply and trigger higher variance. Some recover. The resulting risk-neutral distribution is no longer the smooth, fixed-vol lognormal shape of Black–Scholes. It becomes a mixture generated by many future variance paths, with asymmetry induced by the correlation between returns and variance.

When you then invert market prices back into Black–Scholes implied volatilities, you see a skew. The skew is not an independent object in Heston. It is the visible footprint of the joint dynamics of spot and variance.

Why Heston is computationally practical for European options

At first glance, adding a second stochastic process should make the model much harder to use. And in one sense it does: the pricing equation becomes more complex because option value depends on both S_t and v_t. The reason Heston became so influential is that this extra complexity can still be managed analytically for European-style claims.

The key tool is the characteristic function of the log asset price. A characteristic function is the expectation of exp(i u x) for a random variable x, where u is a real frequency variable and i is the imaginary unit. That may sound abstract, but the intuition is straightforward: it is another way of encoding a probability distribution, often easier to manipulate than the density itself.

Heston derived a closed-form expression for this characteristic function under his stochastic-volatility dynamics. Once you have that function, option prices can be recovered by numerical inversion. This is why people often describe Heston as “semi-closed-form.” The hard probabilistic part is solved analytically; the remaining numerical integration is relatively light compared with solving the full pricing PDE from scratch for every contract.

That computational reduction is a large part of the model’s practical success. It made stochastic volatility usable in settings where speed matters, especially when calibrating to many vanilla options repeatedly. Later pricing methods such as Fourier inversion, FFT-based approaches, and COS-style expansions all build on the same basic fact: if you know the characteristic function, many prices become accessible quickly.

Numerical pitfalls in Heston implementations (the ‘little Heston trap')

There is a trap here. The Heston formula is elegant on paper, but not every algebraically equivalent implementation behaves equally well on a computer. This is not a cosmetic issue. Several later papers showed that some common forms of the characteristic function can run into branch-cut discontinuities in complex logarithms and square roots, leading to visible pricing errors, especially for longer maturities.

This issue became known informally as the little Heston trap. The underlying mathematical point is that complex logarithms and powers are multivalued. Software typically chooses a principal branch. In one formulation of the Heston characteristic function, that choice can make the computed function jump discontinuously as the integration variable moves, even though the theoretical object should remain continuous along the pricing contour.

The practical consequence is severe: a model can look calibrated and reasonable for short maturities, yet quietly misprice longer-dated options because the numerical representation is unstable. Later work showed that an alternative but algebraically equivalent formulation (often called the second formulation or a Duffie/Bakshi-style representation) avoids these discontinuity problems much more robustly. Some implementations also use rotation-count corrections to track the complex argument continuously.

This is a useful lesson about quantitative finance more broadly. A “closed form” does not automatically mean “safe to code naively.” In Heston, implementation choices about the characteristic function, quadrature, and branch handling are part of the model, in the practical sense that they determine whether the numbers you produce actually correspond to the theory.

How traders calibrate Heston to market option prices

In practice, people do not start with Heston parameters because they are economically observable. They infer them from option prices. This is calibration: choose v0, θ, κ, σ, and ρ so that the model reproduces a selected market volatility surface as closely as possible.

That sounds straightforward, but calibration is really an inverse problem. Market prices tell you about the shape of the risk-neutral distribution, not directly about the “true” physical dynamics of variance. Heston’s original framework already noted the role of a volatility risk premium. That matters because the parameters that price options under the risk-neutral measure need not match parameters estimated from historical time series of spot and realized variance.

The geometry of calibration is also subtle. Later work on fast Heston calibration found that the objective function often looks less like a landscape with many clearly separate local minima and more like a narrow valley with a flat bottom. That means several parameter sets can price vanillas similarly well, even if the parameters themselves differ noticeably. A desk may think it has identified “the” Heston parameters, when in fact it has identified one acceptable point in a thin region of near-equivalent fits.

A concrete workflow helps here. A desk might collect liquid vanilla option quotes across strikes and maturities, convert them into prices or implied volatilities, attach a Heston pricing engine, and minimize an error metric such as price error, relative price error, or implied-volatility error. Libraries such as QuantLib expose exactly this pattern through helpers that tie a Heston model to market quotes and a pricing engine. MATLAB examples show the same structure: fit Heston to call prices, inspect the resulting implied-volatility surface, and then reuse the calibrated model to price more complex products by Monte Carlo.

That last step explains why calibration matters. Vanilla options are usually the data source because they are liquid and observable. Exotic options are often the target because they are harder to hedge and mark directly. The desk uses the vanilla surface to infer a stochastic-volatility world, then prices path-dependent or American-style claims within that world.

Heston strengths and limitations: when it works and when it fails

The Heston model gets something important right about markets: volatility is dynamic, clustered, and related to price moves. That is enough to generate skew and smile effects that Black–Scholes cannot produce on its own. It also creates a richer framework for hedging because changes in volatility are endogenous to the model rather than external adjustments pasted onto a constant-volatility engine.

But the model still makes strong assumptions. The variance process is continuous. There are no jumps in the basic specification. That can be a problem when markets reprice discontinuously around earnings, macro releases, or crises. In those settings, a pure diffusion with stochastic variance may still understate short-dated wing behavior unless extended with jumps.

The square-root variance process also introduces boundary behavior that matters numerically and conceptually. A commonly cited positivity condition is the Feller condition, 2 κ θ > σ^2. If it holds, the variance process stays strictly away from zero in a strong sense. If it fails, variance can reach zero, and numerical methods can become delicate. Importantly, the model is still used even when the condition is not satisfied; violating it does not automatically invalidate calibration, but it does mean some solvers or discretizations may behave poorly.

There is also a distinction between fitting vanillas and managing risk in exotics. Research on calibration risk under Heston shows that different plausible calibration choices can lead to materially different exotic prices and different hedging P&L distributions. That is not a paradox. Vanillas constrain the surface, but exotics depend on more of the joint path structure. If several parameter sets fit vanillas almost equally well, they may still disagree on path-dependent risk.

So the right way to think about Heston is not as “the true volatility model,” but as a tractable compromise. It captures more of reality than constant-volatility models, while remaining usable enough for daily pricing and calibration. Its weaknesses are not bugs accidentally left behind. They are the price of keeping the model structured and computable.

Using Heston for exotics: Monte Carlo, PDEs and other pricing methods

Although the original Heston paper focused on European calls and also extended the framework to bond and currency options, the model’s practical life is broader. Once you have calibrated dynamics for spot and variance, you can price products that do not admit a simple closed-form payoff expression by using other numerical methods.

For path-dependent or early-exercise derivatives, practitioners commonly switch from the characteristic-function formula to Monte Carlo, trees, finite-difference PDE schemes, or Fourier-based numerical methods such as COS expansions. The model itself stays the same; only the pricing method changes because the payoff is more complicated. That is why Heston often serves as a base model inside libraries and vendor systems rather than as a single formula used in isolation.

This also explains why implementation ecosystems matter. QuantLib exposes a HestonModel object with standard parameters theta, kappa, sigma, rho, and v0, and calibration helpers that connect market quotes to pricing engines. Commercial and academic workflows follow the same pattern. The Heston model is not just a paper formula; it is a reusable engine component sitting between vanilla market data and downstream pricing or risk calculations.

Heston model in one paragraph: volatility as a state variable

The Heston model is a stochastic-volatility option-pricing model in which the asset price and its variance evolve together, with variance mean-reverting and correlated with price shocks. That single structural change is enough to explain why real markets produce implied-volatility skews and smiles instead of the flat surface implied by Black–Scholes.

Its enduring value comes from a rare combination: richer market behavior and still-manageable mathematics. The characteristic-function solution makes European pricing and calibration fast enough for practical use, while the same calibrated dynamics can feed more complex pricing methods for exotics.

If you remember one idea tomorrow, make it this: Heston works because volatility is treated as part of the state of the market, not as a fixed number chosen outside the model. Once volatility can move (and especially once it can move with price) the option surface stops looking like a flaw in the data and starts looking like a consequence of the mechanism.