Pricing a Thousand Options at Once: Carr and Madan's Fourier Transform

By the late 1990s, quantitative finance had a bottleneck, and it was not a shortage of ideas. It was speed.

Researchers had produced a rich collection of realistic models. Heston's stochastic volatility. Merton's jumps. Bates's combination. Variance Gamma. All of them fit the market better than Black-Scholes. All of them shared one practical problem: to use a model, you must first calibrate it, which means adjusting its parameters until its prices match the hundreds of option quotes on your screen. Calibration is an optimisation, and an optimiser will evaluate the model thousands of times before it converges. If a single option price takes a tenth of a second to compute, a calibration takes hours, and you have to do it several times a day.

Peter Carr and Dilip Madan removed the bottleneck by importing a trick from electrical engineering, and the effect on the field was immediate.

The problem: everyone had a characteristic function and nobody could use it fast enough

Here is a fact that had become clear by the late 1990s. For a very wide class of realistic models, you cannot write down the probability distribution of the stock price in closed form. But you can write down its characteristic function.

A characteristic function is a distribution's fingerprint. It contains exactly the same information as the distribution, encoded in a different language, the language of frequencies rather than of outcomes. It is the Fourier transform of the density. If you know the characteristic function, you know everything about the distribution, in principle.

Why does this matter? Because for models built out of processes with independent increments, and that includes Heston, Merton, Bates and Variance Gamma, the characteristic function turns out to have a simple explicit formula, even when the density does not. The information is there. It is just in the wrong format.

So the standard recipe was: write down the characteristic function, then run a numerical integral to invert it back into a price. It works. It is also slow, and it computes one option, at one strike, per integration. To price a hundred strikes you do a hundred integrations.

The key idea via analogy: the equaliser on a stereo

Think about how music is stored. You could store the pressure of the air at every instant, which is the waveform, the "outcome" view. Or you could store how much of each frequency is present, which is the spectrum, the "characteristic function" view. Both descriptions are complete. Converting between them is what a Fourier transform does.

The reason your phone can do this in real time is the Fast Fourier Transform (FFT), an algorithm published by Cooley and Tukey in 1965 that reduced the cost of that conversion from unmanageable to trivial. It is one of the most consequential algorithms ever written, and it powers digital audio, image compression, medical imaging and mobile phones.

Carr and Madan's contribution was to reformulate option pricing so that the FFT could be applied directly. And the crucial payoff is not just that it is fast, it is what it is fast at.

The FFT does not compute one number. It computes an entire grid of outputs in one pass. So Carr and Madan set the problem up so that the "frequency" axis is the characteristic function and the "output" axis is the log-strike. One FFT then returns the prices of options at hundreds or thousands of strikes simultaneously, for the cost of a single computation.

That is exactly, precisely what a calibration routine needs. A calibrator does not want one option price. It wants the whole strike-by-strike smile for a given maturity, so it can compare it to the market. Carr and Madan gave it that, in one shot.

The technical obstacle, and the fix

There is a wrinkle, and dealing with it is a real part of the paper's contribution.

The Fourier transform of a call option's price does not, strictly, exist in the ordinary sense. The problem is at the low-strike end: as the strike goes to zero, the call price does not decay to zero, it converges to the stock price. That non-decaying tail makes the integral ill-behaved and the transform undefined.

Carr and Madan's fix is a piece of applied cleverness. They multiply the call price by a damping factor, an exponential that decays with the strike, which forces the whole thing to be well-behaved and square-integrable. Transform the damped price, do the FFT, then undamp at the end. The damping parameter is a free choice, and choosing it well matters: too small and the integral misbehaves, too large and you get numerical noise. Selecting it sensibly is one of the practical arts of implementing the method.

They also provided a variant specifically designed for options very close to expiry, where the price is nearly a kink and the standard approach becomes numerically unstable.

Why it mattered

• It made advanced models usable in production. This is the whole point. Before Carr-Madan, calibrating a Heston model to a live surface was painful enough that many desks did not bother. After, it took seconds. The gap between "there is a better model in the literature" and "we are trading on it" collapsed.
• It created a standard recipe for building models. After 1999, the design pattern for a new option pricing model became explicit: find a process whose characteristic function you can write down. If you can, pricing is free, courtesy of the FFT. If you cannot, your model is a curiosity. This is a strange but real constraint, and it shaped which models the industry adopted. Duffie, Pan and Singleton's affine framework, published the following year, is essentially a systematic catalogue of processes that satisfy it.
• It spawned a family of methods. The COS method (a cosine-series expansion), the Lewis approach, fractional-FFT variants, and various refinements all descend from this paper. Fourier pricing is now a standard chapter in every computational finance textbook.
• It exemplifies borrowing from other fields. The FFT was fifteen years old and utterly standard in engineering when Carr and Madan brought it to options. A great deal of quantitative finance progresses this way, and this is one of the cleanest examples.

The honest limitations

• It only works for European options. The method prices payoffs that depend solely on the terminal price. Anything path-dependent, a barrier, an Asian, an American, is out of reach, because the characteristic function of the terminal price says nothing about the path taken. Those still need trees, PDEs or Monte Carlo.
• You must have a closed-form characteristic function. That is a real restriction on what models you can use, and it has arguably distorted the field: models get chosen partly for their analytical convenience rather than their realism. Rough volatility models, which fit the data better than almost anything, are computationally painful precisely because their characteristic functions are not simple.
• The damping parameter is a fudge, and it matters. There is no single correct choice, the best value depends on the model and the maturity, and a bad choice degrades accuracy in ways that are not always obvious. Implementations differ, and so do their results.
• The FFT returns a grid, not arbitrary points. You get prices at a set of log-strikes determined by the algorithm's spacing, which will not be the strikes actually quoted in the market. You must interpolate, which introduces its own small errors.
• It computes prices, not sensitivities. Greeks require either additional transforms or numerical differentiation, both of which add complexity.
• Speed can flatter a bad model. Being able to calibrate a model in one second makes it very tempting to calibrate it every hour and trade on the results, without ever asking whether the model's dynamics are believable. Fast calibration to today's prices is not the same as a model that will hedge correctly tomorrow.

The one-line takeaway

Carr and Madan noticed that although realistic option models rarely have a usable probability density, they almost always have a simple characteristic function, and by damping the call price and applying the fast Fourier transform they made it possible to price an entire strip of strikes in a single computation, which turned model calibration from an overnight job into a one-second one and is the reason Heston, Bates and Variance Gamma ended up on trading desks rather than in journals.