Time Runs at a Random Speed: the Variance Gamma Model

There is a strange and beautiful idea buried in this paper, and it starts with a question about time.

Black-Scholes assumes the stock price is Brownian motion running on ordinary clock time. Every day contributes the same amount of randomness as every other day. But that is obviously not how markets feel. Some days nothing happens: the market drifts sideways, volume is thin, nobody trades. Other days are a torrent: news breaks, volume explodes, and the market experiences more genuine price discovery in an hour than it did in the previous fortnight.

So here is the thought. What if the market does not run on calendar time at all, but on some kind of business time or information time, a clock that ticks fast when news is flowing and slowly when it is not? And what if that clock itself is random?

Madan, Carr and Chang built exactly that model, and it turns out to explain the volatility smile with remarkable economy.

The problem: getting fat tails without bolting on jumps

By the late 1990s everyone knew that returns have fat tails and that Black-Scholes therefore misprices options away from the money. There were two established fixes:

• Add jumps (Merton). The price is a diffusion most of the time, but occasionally it teleports.
• Make volatility random (Hull-White, Heston). The throttle wobbles.

Both work. Both also feel a little like patching. You start with a model you know is wrong and glue on an extra mechanism to compensate. The Variance Gamma model takes a different route: it changes something more fundamental, and gets the fat tails as a consequence rather than as an addition.

The key idea via analogy: a clock that speeds up and slows down

Take an ordinary Brownian motion, the perfectly well-behaved random walk of Black-Scholes. Do not modify it at all.

Now replace the clock.

Instead of evaluating the Brownian motion at time t, evaluate it at a random time G(t), where G is itself a random, always-increasing process. This is called a subordinator, or more evocatively, a time change. The specific one used here is a gamma process, which is why the model is called Variance Gamma: a Brownian motion whose variance is dictated by a gamma-distributed clock.

What does this mean intuitively? Over one calendar day, the market might experience a lot of business time (a busy, news-filled day, so the Brownian motion runs a long way) or very little (a quiet day, so it barely moves). The amount of randomness the market absorbs in a day is itself random.

And now watch what falls out, for free:

• Fat tails. On the rare days when the clock races ahead, the stock experiences an enormous amount of Brownian motion in one calendar day, so it can move a very long way. Those are your tail events. You never had to insert a "jump." You just let time run fast.
• Excess kurtosis. Mixing normal distributions with random variances always produces something more sharply peaked and more heavy-tailed than a normal. The quiet days pile up around zero (the peak) while the busy days populate the tails. Real return distributions have exactly this shape: too many small moves, too many enormous ones, not enough medium ones.
• Skew. Add a drift to the Brownian motion before you time-change it, and the fat tails become asymmetric. A negative drift means the busy periods tend to be down periods, which is exactly the equity market's character: crashes are fast and rallies are slow. One parameter, one skew.

The model has just three parameters beyond the obvious: overall volatility, a kurtosis parameter (how variable the clock is, controlling the fatness of the tails and hence the smile's curvature), and a skew parameter (the drift, controlling the tilt). Each maps directly onto a visible feature of the smile. That interpretability is a genuine selling point.

The deeper structure: it is secretly all jumps

There is a lovely mathematical twist. Although the model is built as "Brownian motion on a random clock," it turns out to be equivalent to a pure jump process: a process with no continuous component at all, made up entirely of jumps, but with infinitely many tiny jumps arriving in every interval, plus rarer large ones.

That is a genuinely different picture of what a stock price is. Black-Scholes says the price is continuous, moving in infinitesimal shivers. Merton says it is continuous with occasional gaps. Variance Gamma says there is no continuous part at all: price movement is nothing but the accumulation of trades, each of which moves the price by a discrete amount, most of them tiny, a few enormous. Which, if you have ever looked at a limit order book, is a rather accurate description of reality.

The model is the best-known member of a family called Levy processes, and this paper is a large part of why that family entered mainstream finance.

Why it mattered

• It fits the short-dated smile, which Heston cannot. This is its killer application. Because the tail-generating mechanism is instantaneous (the clock can race tomorrow) rather than cumulative, Variance Gamma produces a steep, curved smile at short maturities where stochastic volatility models go flat. Combine it with stochastic volatility (which handles the long end) and you get a model that works across the surface.
• It is fast. The characteristic function of the Variance Gamma process has a simple closed form, which means it plugs straight into the Carr-Madan Fourier transform machinery and prices in microseconds. Realism at production speed, which is always the constraint.
• It brought Levy processes into finance. After this paper, the pure-jump, infinitely-active view of price formation became a serious modelling tradition, producing the CGMY model, the normal inverse Gaussian model, and a considerable literature.
• It reframed what volatility is. In this model, "volatility" is not a property of the asset so much as a property of the rate at which information arrives. That is an economically appealing idea: the market is volatile because a lot is happening, not because of some abstract diffusion coefficient.

The honest limitations

• It has no volatility clustering. This is its most serious flaw, and it is a big one. The random clock in Variance Gamma has independent increments: a busy day tells you nothing about whether tomorrow will be busy. But real markets cluster violently, turbulent periods persist for weeks. The model captures fat tails but completely misses the persistence of volatility. It is right about the distribution and wrong about the dynamics.
• The smile it produces flattens too fast. Because the model has independent increments, the central limit theorem takes over as you extend maturity, so the smile flattens with time faster than the market's does. Great at the short end, poor at the long end, which is precisely the opposite of Heston's failure and precisely why the two get combined.
• The market is incomplete. With infinitely many possible jump sizes and one hedging instrument, you cannot replicate. There is no unique price, and the model's answer depends on assumptions about how jump risk is compensated.
• Delta hedging is conceptually awkward. In a pure-jump world, the whole intuition of continuous hedging, adjust a little as the price moves a little, no longer applies cleanly, because the price does not move a little. Hedging in these models is inherently approximate.
• Calibrated parameters are unstable. As with most flexible models, several parameter sets fit today's surface roughly equally well, and the fitted values move around more than the market does.
• The time change is elegant but not observable. Nobody can measure business time, so the model's central object is a latent construct, and its economic interpretation, while appealing, is not directly testable.

The one-line takeaway

Madan, Carr and Chang got the volatility smile not by adding jumps to a random walk but by letting the clock itself run at a random speed, so that busy, news-filled days produce enormous moves and quiet days produce almost none, an elegant reframing that generates fat tails and skew from a picture of markets as pure accumulated trading activity, and which fits the short-dated smile that stochastic volatility models cannot, while completely failing to capture the fact that volatility clusters.