The Workhorse: Heston's Closed-Form Stochastic Volatility Model

By 1993 the options market had a problem it could not ignore. Since the 1987 crash, index options had traded with a pronounced skew: out-of-the-money puts commanded far higher implied volatilities than calls. Black-Scholes said this was impossible. Every desk was quietly using a different volatility for every strike, which is a polite way of saying they had abandoned the model while continuing to use its formula as a quoting convention.

Several researchers had already shown how to build models where volatility is random. Hull and White had done it in 1987. The trouble was that those models were slow. To price an option you had to simulate or integrate numerically, and a trading desk repricing a book of ten thousand options every few seconds cannot afford that.

Steven Heston solved the problem by finding a stochastic volatility model that is realistic enough to fit the skew and has a closed-form solution you can evaluate almost instantly. That combination, not any single ingredient, is why the Heston model became the industry workhorse and remains so more than thirty years later.

The problem: realism and speed were in conflict

The demands on a production options model are brutal and contradictory.

It must fit the market, reproducing the observed smile and skew across strikes and maturities, or your prices will be arbitraged.

It must be fast, because a desk calibrates the model to hundreds of market quotes many times a day, and each calibration involves pricing thousands of options.

It must produce stable, sensible Greeks, because the hedges are what actually keep you alive.

Before Heston, you could have any two of these. Black-Scholes was fast and stable but did not fit. Stochastic volatility models fit but were far too slow to calibrate.

The key idea via analogy: a thermostat with a mind of its own

Heston's model has two moving parts.

The stock, which does the usual random walk, except that the size of its steps is governed by the current level of volatility.

The variance, which is itself a random process, and here is Heston's key choice: he made it mean-reverting. Volatility does not wander freely. It is pulled back toward a long-run average, like a thermostat pulling the room temperature back to its setting. When volatility spikes in a crisis, the model expects it to decay back down. When markets are unnaturally calm, the model expects turbulence to return. This matches what volatility actually does, and it is the single most important qualitative feature of real volatility.

The specific process he used, now universally called the Heston variance process, has a further nice property: it cannot go negative. Variance is bounced off zero rather than passing through it, which is essential, because a negative variance is meaningless.

And then the crucial ingredient, the one Hull and White had left out.

Heston let the two random drivers be correlated. The randomness pushing the stock and the randomness pushing volatility are allowed to be linked, and in equity markets the link is strongly negative. When the stock falls, volatility rises. When the stock rallies, volatility drifts down. That single parameter, usually written as rho, is what generates the skew. If volatility rises precisely when the market drops, then bad outcomes are also volatile outcomes, so the left tail of the distribution is stretched out. Downside puts become more valuable. The implied volatility curve tilts. That is exactly what the market shows, and Heston's rho is the dial that produces it.

The parameters, once you strip away the notation, map cleanly onto things a trader already thinks about:

• The long-run variance: where volatility settles in the long run.
• The speed of mean reversion: how fast a volatility spike decays. This controls the term structure, how the smile flattens for longer maturities.
• The volatility of volatility: how violently volatility itself moves. This controls the curvature of the smile, how much it bends.
• The correlation: the tilt, the skew.
• The current variance: where volatility is right now.

Five parameters, each with a distinct and interpretable effect on the shape of the volatility surface. That interpretability is a large part of why practitioners like the model: when the fit is wrong, you can usually tell which knob to turn.

The technical breakthrough: a formula, not a simulation

The mathematical achievement is that Heston found a way to get a semi-closed-form solution. He worked not with the probability distribution directly but with its characteristic function, which is a kind of frequency-domain fingerprint of a distribution. For his particular choice of variance process, that fingerprint has an explicit formula. Recovering the option price from it then requires a single numerical integral, which computers do in microseconds.

That is the whole game. A model rich enough to fit the smile, but which prices in the time it takes to evaluate an integral rather than the time it takes to run a simulation. It is why Heston, and not one of the several other stochastic volatility models proposed around the same time, took over.

Why it mattered

• It became the default. Ask a quant to name a stochastic volatility model and they will say Heston. It is implemented in every serious pricing library, taught in every graduate course, and used as the baseline against which every new model must justify itself.
• It made the characteristic function the standard tool. After Heston, the recipe for a new option pricing model became: define a process, find its characteristic function, invert numerically. Carr and Madan's fast Fourier transform paper (1999) and Duffie, Pan and Singleton's affine jump-diffusion framework (2000) generalise exactly this approach. Heston is the paper that showed it could work.
• It gave the skew a mechanism. The equity skew is not an arbitrary market convention. In Heston's world it is what you must observe if volatility rises when markets fall. That is an economic story, and it is largely correct.
• It underpins variance derivatives. Because the model is written in terms of variance rather than volatility, it connects naturally to variance swaps, the VIX, and the whole modern volatility trading complex.
• It generalises well. Bolt jumps onto it and you get the Bates model. Bolt on stochastic rates and you get models for long-dated hybrids. It is a platform as much as a model.

The honest limitations

• It cannot fit the short-dated smile. This is the most serious, most well-known failure. Real short-maturity options, expiring in days, show an extremely steep, sharply curved smile. To match it, Heston must be calibrated with an absurdly high volatility of volatility, which then ruins the fit everywhere else and implies dynamics nobody believes. The reason is structural: over a short horizon, a continuous volatility process simply cannot move enough to create a steep smile. You need either jumps, or a volatility process that is far rougher than Heston's. The rough volatility literature exists largely because of this failure.
• Calibration is unstable. The five parameters trade off against each other. Different parameter sets can fit today's market almost equally well, and the fitted values can jump around dramatically from one day to the next even when the market has barely moved. That is a serious problem when your hedge ratios depend on those parameters.
• The variance process can misbehave near zero. Unless a particular condition on the parameters holds (the Feller condition), the variance can get arbitrarily close to zero, which causes real numerical headaches in simulation. In practice, calibrated parameters routinely violate this condition, and practitioners have built an entire cottage industry of workarounds.
• The market is still incomplete. Two sources of risk, one stock. You cannot hedge volatility with the stock, so you must hedge vega with other options, and your hedge is only as good as the model. And the price you compute depends on an assumed price of volatility risk, which cannot be arbitraged into existence.
• It has no jumps. Real markets gap. A continuous-volatility, continuous-price model will always underprice genuinely sudden events.
• Its fitted parameters are often economically implausible. The volatility-of-volatility required to fit market smiles is frequently far higher than what is observed in realised volatility data. The model fits the prices by adopting a dynamic that is not the one the world actually follows, which should worry anyone who intends to hedge with it.

The one-line takeaway

Heston made volatility a mean-reverting random process, let it rise when the market falls, and found a characteristic function that turns the whole thing into a fast, computable formula, delivering the first model that could fit the equity skew at production speed, which is why it became the industry standard despite being unable to fit short-dated options at all.