The trap, one coin at a time
The fan above is Figure 1. The next two use the same coin and the same colours: ochre for the money, slate for the people. One shows where the expected value went. The other shows how to stop feeding it.
Where the money is
Every player who throws the same number of heads ends with the same wealth, so the whole game can be counted exactly. The slate hump is where the players end up. The ochre hump is where the money ends up.
Take hold of either hump and drag it. The two humps walk away from each other. The expected value sits between them, in a place where there are almost no players and almost none of the money.
Figure 2. Final wealth after the chosen number of rounds, on a logarithmic scale. Slate bars count players and ochre bars count dollars. This is an exact binomial count, not a simulation, and each hump is scaled to its own peak. The white marker is you, from Figure 1.
The best expected value on offer, and the worst outcome for the person holding it.
The way out
Same coin, same flips. The only thing you change is how much of your wealth rides on each one. Staking everything is the fan at the top of the page.
Stake a quarter and the typical player starts to grow. That fraction is Kelly's. Past a half, the typical player shrinks again, although every single bet still has positive expected value.
260 players, 300 rounds, drawn to the same scale as the fan above
Growth of the typical player, per round, by fraction staked. Drag the dot.
The stake follows your scroll, from all in down to Kelly. Touch the slider to take over.
Figure 3. The same 260 sequences of coin flips at every stake. Only the fraction of wealth wagered each round changes. The curve gives the typical player's growth per round at each stake. The white line is your own run of flips from Figure 1, replayed at this stake.
Figures 1 to 3 are one essay from the book. There is one like it each month. Get the next one by email
Essays
Three other essays, one figure each. They use different games, and the same habit of mind: ask what a single player lives through before trusting the average.
St. Petersburg Paradox
Flip until the first tails. If it comes on flip n, you win $2n. The expected payout is infinite, which is the ergodic trap with the ceiling taken off.
The line is your average winnings per game. It jumps on every long run of heads and sags in between. It has no limit to settle on.
Figure 4. Average payout per game against games played, with games on a logarithmic axis. The dashed line is log₂ of the number of games, which is roughly how fast the average drifts upward (Feller).
Monty Hall Problem
Three doors, one prize, two goats. Pick a door. The host, who knows where the prize is, opens another to show a goat. Stick or switch?
Play it both ways and the graph keeps score. The two lines settle on one third and two thirds, which is not what it feels like at the door.
Pick a door.
No rounds played yet.
Figure 5. Share of rounds won by each strategy as rounds accumulate. The dashed lines mark one third and two thirds.
The Cult of 0.05
Nothing in this figure is real. Every test compares two groups drawn from the same population, so every finding is noise. Choose how many things the study measures, then run it.
With one test, a false alarm turns up one time in twenty. With twenty tests, most studies of nothing find something, and that something is what gets written up.
Figure 6. Each dot is the p-value of one test in a single study where nothing is going on. When there is no effect a p-value is equally likely to land anywhere between 0 and 1, so each is one uniform draw. The axis is on a square-root scale to open up the left end. The bar counts repeated studies that found at least one result below 0.05, and the white tick is what arithmetic predicts.
With twenty tests, most studies of nothing find something.
from The Cult of 0.05How sure is sure?
The figures above are about being wrong on average. This one is about knowing how wrong you might be. It takes two minutes and nobody sees your answers but you.
Ninety Percent Sure
For each question, give a low number and a high number so that you are 90% sure the true answer lies between them. You are not asked to know the answer. You are asked to know how much you do not know.
If your ranges are honest, about nine of the ten will catch the truth. Make them as wide as you need. Do not look anything up.
Figure 7. After scoring, each bar is one of your ranges, drawn to the same width so they can be compared, and each dot is the true answer: slate when your range caught it, ochre when it did not. A dot far from the bar means the truth was several range-widths away. The exercise is an old one in decision research, and in published versions most people catch far fewer than nine. A range that is too narrow is the same mistake as a risk model that is too confident.
The expected value is real. It is just not yours.
from The Ergodic TrapWhere to next
Essays
St. Petersburg, Monty Hall, the cult of 0.05, and a two-minute test of how sure you are. One figure each.
Play them →The book
Elegant Errors, a book in progress, and the writing drawn from it.
See the book →Five papers
Shannon, Becker, Lucas, Rabiner and Black. The idea, the mistake it kills, and a figure to move.
Read them →Notebook
Hidden Markov models and wavelets, worked out as long visual guides.
Open the notebook →A range that is too narrow is the same mistake as a risk model that is too confident.
from Ninety Percent SureThe book
Elegant Errors is a book of short essays on probability and statistics, each built around a beautiful idea and the expensive mistake people make with it. From Cardano's dice to the copula formula, from the bell curve to the cult of 0.05. A book in progress.
A few of the essays.
- The Ergodic Trap in Wilmott, graphs on the home page
- St. Petersburg graph
- Monty Hall game
- Fat Tails and Infinite Variance
- The Cult of 0.05 figure
- The Great Formula That Ate Wall Street
- The Model That Destroyed Itself
- and more
Writing
Essays on the places where the arithmetic of chance and the experience of living through it come apart. Most are drawn from the book below.
-
Wilmott, November 2026
The Ergodic Trap: Why Positive Expected Value Can Still Destroy You
A backtest that averages across worlds, a portfolio that lives in one. Jensen's inequality, Kelly, Thorp, and Peters. Figures 1 to 3 on the home page are its argument.
Read it in the November 2026 issue Open the interactive version at Wilmott
-
Coming 2027
The Model That Destroyed Itself
What happened to the Phillips curve happens to every trading model that gets used. On alpha decay, reflexivity, and the difference between a pattern and a law.
Notebook
Mathematical ideas I have thought about for far longer than was sensible, drawn out as long visual guides.
Hidden Markov Models
States you cannot see, read from the things you can. Viterbi, Baum-Welch, and the model's relatives in one long guide.
Open the guide →Wavelets
How to take a signal apart by scale and by time at once, and why a sine wave cannot tell you when something happened.
Open the guide →Method in the Madness
Five papers that still blow my mind: Signal, Noise, and three others.
Each one has an idea, a mistake it kills, and a figure you can move.
A Mathematical Theory of Communication
Claude E. Shannon, Bell System Technical Journal
- The idea
- Information is surprise, and it can be counted in bits. Every noisy channel has a capacity, and below that rate a message can get through with almost no errors.
- The mistake it kills
- That a noisy line means a garbled message. Noise limits how fast you can send. It does not limit how accurately.
On this site: Kelly's betting rule in Figure 3 came out of this paper eight years later.
Figure 8. Bits of information in one flip, by how biased the coin is. Drag the dot.
Irrational Behavior and Economic Theory
Gary S. Becker, Journal of Political Economy
- The idea
- Demand curves slope down even when every buyer chooses at random. A higher price shrinks what anyone can afford, and that alone does the work.
- The mistake it kills
- Reading the order in a market as proof that the people in it are rational.
Figure 9. Left: each slate dot is a buyer who spends a fixed income at a random point on the budget line, and the ochre dot is their average. Right: that average at every price.
Econometric Policy Evaluation: A Critique
Robert E. Lucas, Jr., Carnegie-Rochester Conference Series on Public Policy
- The idea
- A relationship measured under one policy does not survive a change of policy, because people change what they do when the rules change.
- The mistake it kills
- Treating a fitted curve as a lever you can pull.
On this site: The Model That Destroyed Itself, coming in 2027.
Figure 10. Slate dots are the years before anyone used the curve, and the dashed line is the curve fitted to them. Ochre dots are the years after. A toy economy in which only surprise inflation moves unemployment.
A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition
Lawrence R. Rabiner, Proceedings of the IEEE
- The idea
- The state you care about cannot be seen. What you see is the noise it throws off. Three questions about that, and an algorithm for each.
- The mistake it kills
- Taking each observation at face value. One wild day is not a new regime.
On this site: the Hidden Markov Models guide in the notebook.
Figure 11. The wide slate line is the true hidden state, which really does persist 90% of the time. The ochre line is the most likely path given the observations (Viterbi). A white ring marks a wrong reading.
Noise
Fischer Black, Journal of Finance
- The idea
- People who trade on noise are what make a market liquid, and they are also why its prices are never quite right. Black called a market efficient if price is within a factor of two of value.
- The mistake it kills
- That a liquid market is an accurate one. The trading that lets you in and out is the same trading that pushes price away from value.
On this site: The Cult of 0.05, which is also about taking noise for information.
Figure 12. A simulated market on a logarithmic scale. The band runs from half of value to twice value.
About
I have traded quantitative strategies for over twenty-five years. I was educated at the University of Chicago and Columbia University, and held research positions at the University of Chicago and Harvard University.
The name of this site is a joke about the phrase "in the grand scheme of things," which is what people say when they want to average over the worlds they are not in. Most of us get one sequence. The writing here is about what follows from that.
You can reach me at fq@grandschemeofmadness.com.
Subscribe
One essay a month, no more. New pieces, the occasional simulation, and news about the book.
Unsubscribe any time. Your address is never shared.