Is the Normal Distribution Good Enough for Finance?
Chapter 16 is a short but important one. It asks a question that every quant should think about deeply: is the normal distribution actually a good model for financial returns? The answer is “mostly yes, but catastrophically no.” And that “catastrophically no” part has wiped out entire firms.
Why We Love the Normal Distribution
The Central Limit Theorem (CLT) is the reason the normal distribution shows up everywhere. The idea is simple: if you add up lots of random numbers drawn from the same distribution, the sum ends up being normally distributed. It does not matter what the original distribution looks like. Coin flips, dice rolls, weird asymmetric distributions. Add enough of them together and you get a bell curve.
The fine print has three conditions:
- All the random numbers must come from the same distribution (you cannot mix different sources).
- Each draw must be independent from the others.
- The mean and standard deviation of the original distribution must both be finite.
Since daily stock returns are arguably the sum of thousands of individual trade-level returns during the day, the CLT suggests they should be approximately normal. And the normal distribution only needs two parameters (mean and standard deviation), making the math relatively clean.
Normal vs. Lognormal: A Quick Clarification
Students often confuse these two. Here is the deal:
The Black-Scholes model assumes that stock returns are normally distributed. But the stock price itself is lognormally distributed. How does that work?
Start with stock price S0. After each period, multiply by (1 + R) where R is the random return. After N periods:
S_N = S0 times (1 + R1) times (1 + R2) times … times (1 + R_N)
Take the logarithm of both sides. The product becomes a sum of log(1 + Ri) terms. By the CLT, this sum is approximately normal. So log(S_N) is normal, which means S_N itself is lognormal.
That is the whole derivation. Returns are normal, prices are lognormal. The CLT does the heavy lifting.
The Problem: Fat Tails
Now for the bad news. Compare the actual distribution of S&P 500 daily returns with a normal distribution that has the same mean and standard deviation. You will see two things:
- The actual distribution has a higher peak in the middle.
- The actual distribution has fatter tails (more extreme events than the normal predicts).
The high peak is annoying but not dangerous. The fat tails are a different story entirely.
The 1987 Crash Test
Wilmott uses the crash of October 19, 1987 to make the point unforgettable. On that day, the S&P 500 fell 20.5% in a single day.
Empirical probability: Using 24 years of daily data (about 6,048 trading days), this happened once. So the empirical probability is roughly 1 in 6,000, or 0.000165. A rare event, but it happened within a human lifetime.
Theoretical probability (using normal distribution): The average daily standard deviation of SPX returns was about 1.06%. A 20% move is about 19 standard deviations away from the mean. Under the normal distribution, the probability of this is approximately 1.8 times 10 to the power of minus 79.
Let me write that out. The normal distribution says a 20% single-day crash should happen once every 2 times 10 to the 76th power years. The universe is roughly 1.4 times 10 to the 10th power years old. So the normal distribution says this event should basically never happen in the entire history of the universe, multiplied by itself several times over.
In reality, it happened on a regular Tuesday in October.
Empirical answer: once every 25 years. Theoretical answer: once every 2 times 10 to the 76th power years. That is the gap between the normal distribution and reality when it comes to tail events.
Why Not Just Use a Different Distribution?
This seems like an obvious fix. The normal distribution underestimates crashes, so use a distribution with fatter tails. Problem solved, right?
Not so fast. The distributions that best fit real return data have tails so fat that their standard deviation is infinite. Yes, infinite. The standard deviation, the number that the entire finance industry uses as its primary measure of risk, simply does not exist for these distributions.
Wilmott lays out the consequences in a stark comparison:
| Normal Distribution | Fat-Tailed Distribution |
|---|---|
| Math is easy | Math is hard |
| Underestimates crashes | Good estimate of crashes |
| Standard deviation = a finite number (volatility) | Standard deviation = infinity |
| Delta hedging works | Delta hedging does not work |
| Risk preferences do not matter | Must model risk preferences |
| Local models, differential equations | Global models, integral equations |
If the standard deviation is infinite, delta hedging becomes impossible. The concept of volatility loses its meaning. You cannot use the Black-Scholes equation anymore. Instead, you need partial integro-differential equations that relate option values at ALL stock prices simultaneously, not just locally.
This is not a minor inconvenience. It means throwing away most of the theory that modern finance is built on.
Serial Autocorrelation
There is another reason the normal distribution might fail. The CLT requires each return to be independent. But there is evidence of serial autocorrelation in stock returns. This means that an up move might be more likely to be followed by another up move (or a down move by another down), especially within a single day.
If returns are correlated, the CLT conditions are violated, and the sum of returns may not converge to a normal distribution. Wilmott notes that very little has been written about pricing derivatives when serial autocorrelation exists, but he promises to tackle it later in the book (Chapter 65).
Wilmott’s Personal Preference
Here is where Wilmott shows his practical side. Despite all the evidence against the normal distribution, he says he prefers to use it most of the time, and treat tail events separately.
His reasoning: the normal distribution gives you tractable math and a framework that works well for typical market conditions. For the extreme events, do not rely on the model at all. Instead, take practical precautions:
- Buy tail protection (like out-of-the-money puts)
- Diversify your portfolio so you are not wiped out by one stock crashing
- Always keep in mind that a dramatic crash can happen, regardless of what your model says
This is the quant equivalent of wearing a seatbelt. You do not refuse to drive because crashes exist. You drive carefully and wear protection.
The Takeaway for Real Life
If you work in finance or invest your own money, this chapter has one message: your risk models almost certainly underestimate the probability of extreme events. The 2008 financial crisis, various flash crashes, and pandemic market drops were all “impossible” according to normal distribution models.
The solution is not to abandon quantitative models. It is to know their limits. Use the normal distribution for everyday pricing and hedging. But never forget that the tails are fatter than your model suggests, and plan accordingly.
Key Takeaways
- The Central Limit Theorem explains why the normal distribution appears everywhere, but it has conditions that financial markets may violate.
- Stock returns are approximately normal, stock prices are approximately lognormal. The CLT connects the two.
- Fat tails are real. The normal distribution underestimates extreme events by absurd margins (factor of 10 to the 73rd for the 1987 crash).
- Distributions that fit the tails better often have infinite standard deviations, breaking most of classical finance theory.
- Wilmott’s pragmatic advice: use normal distributions for everyday work, but always prepare for extreme events that the model says cannot happen.
Previous post: The Binomial Model Part 2: Trees, Greeks, and the Continuous Limit
Next post: What Blackjack and Gambling Teach Us About Investing