Against the Gods Ch. 3: From Dice to the World — Sampling, the Bell Curve, and Regression

Investment Background

Chapter 2 solved a dice problem. Dice have a special property: their probabilities are knowable a priori.

A six-sided die gives each face a one-in-six chance. You need not roll it once to know that — you only count the faces.

The real world does not work that way.

What is the probability distribution of next year's stock returns? There are no faces to count. You can only infer from history, and history is a sample, not the whole.

This chapter covers that leap — from counting to inferring. It required three ideas, and all three are in your portfolio.

The Wall Street Translation

Idea One: The Law of Large Numbers (Jacob Bernoulli, 1713)

Bernoulli asked a question nobody had formally posed: if I do not know an event's true probability, can I estimate it by observing?

His answer — the law of large numbers — is yes, and the more observations, the closer your estimate to the true value.

This sounds obvious. It is not.

It is effectively a mathematical proof that induction works, which is a philosophically very hard problem.

But Bernoulli also discovered a fact that troubled him, and it matters enormously to investors:

Convergence is very slow.

Specifically: to improve the precision of an estimate tenfold, you need one hundred times the observations.

That square-root relationship is among the most useful mathematical facts in this book for a retirement investor.

Its direct consequences:

What you want to estimate	Data required
Whether a coin is fair (within 5%)	roughly 400 tosses
The stock market's long-run average return (within one percentage point)	centuries

Which is why Chapter 6 of Stocks for the Long Run says that two centuries containing ten non-overlapping twenty-year periods is very weak evidence.

That is not pessimism. It is a direct corollary of Bernoulli's law.

Idea Two: The Normal Distribution (de Moivre, Gauss)

The second idea: when many independent random factors combine, the result takes a particular shape — the bell curve.

Its practical value is that an entire distribution can be described by two numbers: the mean and the standard deviation.

That is a remarkable compression. You need not enumerate every possible outcome — only two numbers.

Most of modern finance is built on that compression: volatility is a standard deviation. The Sharpe ratio is a mean divided by a standard deviation. Every Monte Carlo simulation on this site samples from some distribution by default.

And a qualification must follow immediately, because it is the entire subject of another book in this library.

The Black Swan and Antifragile argue that financial market returns do not follow a normal distribution. Extreme events occur far more often than the bell curve predicts. Taleb calls this Extremistan.

This book does not repeat that argument. What it adds is the historical side:

The normal distribution was not an error imposed on finance. It was a tool of extraordinary success in physical measurement, reasonably borrowed into a domain where it does not fully apply.

Understanding this matters, because it determines the correct response. The response is not "abandon all statistics" but "know where this tool is reliable and where it is not."

Idea Three: Regression to the Mean (Francis Galton, 1880s)

The third idea has an interesting origin: Galton was studying heredity, not finance.

Measuring the heights of parents and children, he found: tall parents' children tend to be shorter than their parents, and short parents' children taller.

He called it "regression toward mediocrity."

In investing this phenomenon is everywhere, and it has a very concrete practical consequence:

The best-performing funds, sectors, or asset classes tend to perform worse in subsequent periods — not because some force punishes success, but because extreme performance usually contains luck, and luck does not repeat.

This directly explains two things:

1. Why chasing recent top-performing funds is a losing strategy — the mechanical explanation behind the SPIVA persistence data.
2. Why rebalancing works — it mechanically sells what has risen and buys what has fallen.

An Important Warning

Regression is the most easily misused of the three ideas.

Misuse one: assuming regression is inevitable and time-bound.

"This stock has fallen a lot, so it is due for a bounce." That is the gambler's fallacy. Regression describes a statistical tendency of a distribution, not a guarantee about any single path.

Japan's market peaked in 1989 and has not recovered real purchasing power in over thirty years — the case we used in Chapter 6 of Stocks for the Long Run to question inevitable long-run reversion.

Misuse two: assuming a fixed mean exists to revert to.

This is precisely the reflexivity argument in Chapter 1 of The Alchemy of Finance: in some conditions, the mean itself is moving.

So the correct statement is: regression to the mean is a statistical tendency, reliable in large samples, guaranteeing nothing about a single path, and conditional on that mean being stable.

Executable Trading Rules

1. For any conclusion drawn from historical data, ask the sample size first. The chapter's most practical line. "This strategy worked for the past five years" — how many independent market cycles does five years contain? Usually zero or one.

2. Understand the square-root relationship and use it to discount short-run data. Three years of data is not one-third as reliable as ten years but roughly one-fifth. Short-run performance contains almost no information.

3. Treat "best performing" as a regression warning rather than a buy signal. Concretely: when you want to buy something because it has performed brilliantly, ask how much of that performance was luck. Galton's answer is usually "more than you think."

4. Use statistical tools while assuming the tails are fatter than the model shows. The correct posture toward the normal distribution — not rejection, but adding a margin of safety. In practice: no leverage, and more cash than the model suggests.

5. Do not treat regression as a basis for timing. See both misuses above. It explains why rebalancing works; it cannot tell you when the bottom has arrived.

Relevance to a Retirement Portfolio

These three ideas together produce every number in your retirement plan, and together bound how much those numbers can be trusted.

When our Advanced Withdrawal Simulator reports "85% success," it is doing three things:

1. Inferring a distribution from historical data (Bernoulli) — limited by sample size.
2. Assuming a distributional shape (Gauss) — the tails may be fatter than assumed.
3. Implicitly assuming long-run mean reversion (Galton) — usually true, not guaranteed.

This is not to say the tools are useless. They are extremely useful, and vastly better than planning by feel.

It is to say their output is an estimate with an error band, not a prophecy.

And that recognition has one very concrete operational consequence:

Do not make complex adjustments to move a success rate from 85% to 87%.

Those two percentage points are far smaller than the model's own error band.

Put your effort into the things that do not depend on model precision: raising your savings rate, lowering fees, holding a cash buffer, avoiding leverage, diversifying globally.

Those measures work under any distributional assumption — which is exactly where their value lies.

Chapter 4 covers the first large-scale practical application of these ideas: insurance, humanity's first systematic conversion of risk into a tradeable commodity.