Asymptotic Analysis: When Volatility Moves Fast

Here is a frustrating reality of stochastic volatility models. You pick a model because it is tractable (Heston, anyone?). You get nice semi-closed-form solutions. But what if the model does not actually describe reality well? You have traded accuracy for mathematical convenience, and in finance, that trade can cost you real money.

Chapter 55 of Wilmott’s book offers a way out. Using asymptotic analysis, you can get approximate solutions that work for a wide class of stochastic volatility models. You do not have to commit to one specific model. You just need the model to satisfy some reasonable conditions. The catch? Volatility must mean-revert fast. But as Wilmott shows, empirical evidence says it does.

What Is Asymptotic Analysis?

If you have not encountered asymptotic analysis before, here is the idea in plain terms. Sometimes an equation has a parameter that is very large or very small. Instead of solving the full equation (which might be impossible), you exploit the extreme size of that parameter to find approximate solutions as a series expansion.

Think of it like this. You want to understand a complicated system. One part of the system moves very fast, and another moves slowly. The fast part “averages out” over the timescale of the slow part. You can solve for the slow behavior first (ignoring the fast details) and then add corrections for the fast stuff.

In our context, the stock price moves on one timescale, and volatility fluctuates on a much shorter timescale. The ratio of these timescales gives us a small parameter, and we build our solution as a series in powers of this small parameter.

The Fast Volatility Observation

When people calibrate stochastic volatility models to market data, they consistently find two things. First, the volatility of volatility is large, often bigger than the volatility of the underlying stock itself. Second, volatility mean-reverts quickly.

Wilmott introduces a small parameter epsilon that captures the ratio of the volatility timescale to the stock price timescale. When epsilon is small, volatility fluctuates rapidly around its long-run level while the stock price moves relatively slowly.

This is not just a mathematical trick. It has real empirical support. Wiggins (1987) estimated stochastic volatility models and found that the volatility of volatility was indeed larger than the stock volatility, consistent with a small epsilon. Other researchers have found similar results.

The Leading-Order Solution

With the small parameter in hand, Wilmott looks for the option price as a series:

V = V0 + epsilon * V1 + epsilon^2 * V2 + …

The leading term V0 is the most important. After going through the asymptotic procedure (multiply by the stationary density, integrate over all possible volatilities), V0 turns out to satisfy the plain Black-Scholes equation. But with what volatility?

Not the current volatility. Not the average volatility. The root-mean-square (RMS) volatility with respect to the stationary distribution of the volatility process. This is the effective long-term volatility that the stock “sees” when volatility fluctuates rapidly.

So the zeroth-order approximation is simple: just use Black-Scholes with the RMS volatility. No new math needed. No new numerical methods. Just one number.

This makes physical sense. If volatility jitters around very quickly, the stock price experiences a kind of averaged volatility. It is like shaking a camera during a long exposure: the image blurs, and the blur depends on the average intensity of the shaking, not on the specific path the camera took.

The First Correction: The Skew

The next term in the series, V1, is where things get interesting. This correction is proportional to the correlation between the stock and its volatility. If correlation is zero (stock movements and volatility movements are independent), the first correction vanishes entirely.

For vanilla options, this correction translates into a skew in implied volatility. When you compute the implied volatility corresponding to the asymptotic price V0 + epsilon * V1, you get something that varies linearly with log-moneyness (the log of the forward price divided by the strike).

This is the first-order approximation, and it was previously derived by Fouque, Papanicolaou, and Sircar. Their work stopped here, giving a linear skew. But markets often show implied volatility that is curved, not just tilted. To capture the curvature (the “smile”), you need to go to higher order.

Higher-Order Corrections: The Smile

Wilmott and Rasmussen push the expansion further, to second order and beyond. The second-order term V2 adds curvature to the implied volatility surface. Now the implied volatility becomes a polynomial in log-moneyness, with coefficients that depend on the time to expiry and on model parameters.

At the second order, the correction depends on both the correlation and on the specific stochastic volatility model. At leading order, all models with the same RMS volatility give the same answer. At first order, all models with the same correlation give the same skew. Only at second order and higher do model-specific details start to matter.

This is a powerful insight. It tells you exactly how much your choice of stochastic volatility model matters. For the overall level: not at all (just use RMS volatility). For the skew: only the correlation matters. For the curvature and higher-order smile features: you need to know the details of the model.

The Implied Volatility Formula

For vanilla options, the asymptotic analysis produces a formula for implied volatility as a fifth-order polynomial in forward moneyness. The coefficients of this polynomial depend on:

• The RMS volatility (from the stationary distribution)
• The correlation between stock and volatility
• Certain integrated moments of the volatility process (the F1, F2, gamma functions Wilmott defines)
• The time to expiry

This formula is practical. You can calibrate it to market implied volatilities by fitting a few parameters. Because the formula is explicit (no numerical PDE solving needed), calibration is fast. And because it works for any model satisfying the basic conditions, you are not locked into a specific stochastic volatility specification.

Which Models Qualify?

The conditions for the analysis to work are mild. The volatility process must be autonomous (its dynamics depend only on volatility, not on the stock price). It must have a stationary distribution. And certain integrals involving the stationary distribution must be finite (basically, the RMS volatility must exist).

Most popular models satisfy these conditions. Scott’s model (log-volatility follows an Ornstein-Uhlenbeck process) works. The Heston model (variance follows a square-root process) works. The 3/2 model works. Mean-reverting geometric Brownian motion works.

One notable exception: the lognormal model without mean reversion fails. Without mean reversion, there is no stationary distribution, and the whole machinery breaks down. This is another reason to include mean reversion in your volatility model: it is both empirically justified and mathematically necessary for this kind of analysis.

Limitations

The main limitation is right there in the setup: the time to expiry must be much longer than the characteristic timescale of volatility fluctuations. For very short-dated options, the asymptotic expansion breaks down because the volatility does not have time to average out. The option expires before the fast process has done its job.

For such cases, other approximations exist (Hull and White, Lewis, and others have worked on near-expiry asymptotics). But for options with reasonable time to expiry, say a few months or more, the fast mean reversion approximation works well.

Also, the expansion is an approximation, not an exact solution. For very large values of epsilon (when volatility is not actually mean-reverting fast), the higher-order terms may not converge quickly, and the approximation degrades.

Why This Matters

The practical message of this chapter is liberating. You do not have to pick the “right” stochastic volatility model. At the level of accuracy most people need, the specific model matters much less than having a reasonable estimate of the RMS volatility and the correlation.

This is Wilmott’s recurring theme in these chapters on volatility. Models are tools, not truths. The less your pricing depends on the exact specification of the model, the more robust it is. Asymptotic analysis quantifies exactly how much model detail matters at each level of precision, and the answer is: less than you might think.

For a quant trying to calibrate a model to market smiles, this chapter provides a fast and model-agnostic alternative to brute-force calibration of specific models. Fit the polynomial formula to observed implied volatilities, extract the parameters, and you have a description that is accurate enough for most purposes while being honest about the limits of your knowledge.

Previous post: Stochastic Volatility Meets Mean-Variance Analysis

Next post: The Cliquet Option: A Volatility Case Study