People treat expected value like a promise. You put the probabilities in a spreadsheet, multiply through, and the number that comes out is supposed to be what you’ll make if you stay in the game long enough.
It isn’t.
That number is what a crowd of you would make, if you could average across them. You don’t get the crowd. You get one path.
That sounds abstract until you make it concrete, and a little ugly. Imagine a game of Russian roulette. Five chambers empty, one live, and a pot large enough that the expected value is positive. A thousand people play once. Most of them walk away richer. The spreadsheet looks excellent.
Now change the question. What happens if you are one person, and you play a thousand times?
The average is no longer a description of your life. It’s a description of a pile of lives you don’t get to live. You cannot collect the winnings of the versions of you that survived. One person pulling the trigger a thousand times is not the same experiment as a thousand people pulling it once, and treating them as the same is how a positive expected value talks you into ruin.
Most of the decisions that look like this don’t involve a gun. They involve a bankroll.
The Napkin Version
Ole Peters laid this out with a coin toss simple enough to do on a napkin. A fair coin. Heads, your wealth grows by 50%. Tails, it shrinks by 40%. The expected value of that bet is +5% per toss. Classic theory says take it, and take it again.
Play it sequentially with one bankroll and the picture flips. Start with $100. Heads takes you to $150. Tails then takes 40% of that, and you are at $90. Reverse the order and you get the same place: tails first leaves you with $60, and a 50% gain on $60 is $90. A head and a tail, in either order, is a 10% loss.
You cannot lose 40%, win 50%, and be back where you started.
The time-average growth rate is a little worse than −5% per round. Keep playing, and a single trajectory is almost certain to go to zero. The average across many players still grows, because it is carried by a handful of extremely lucky paths — the ones that flipped heads enough times in a row to dominate the mean. Those paths are real in the average. They are almost certainly not you.
The two averages disagree because wealth compounds. Gains and losses multiply. The order lives in the account. That is what it means for a process to be non-ergodic: the average over many parallel worlds is not the average over your time.
This is easy to miss, because there are games where the average really is what you get. A casino is one of them. So is an insurer. They run many roughly parallel bets at once, so their year can start to look like an ensemble. They can live on expected value because they are, in a literal sense, the crowd. You, with one net worth and one account, are not the house. Treating your bankroll as if it were a thousand bankrolls is the error.
It would be a parlor trick if it only lived in a lecture. It doesn’t. Almost everything in markets is multiplicative. Your account tomorrow is your account today times a return. A strategy with a beautiful expected value can still be a bad life, if the path includes a loss you cannot come back from. The same thing shows up in a pitch deck: a blended return across funds, vintages, or hypothetical investors, offered as if it were a forecast of the money you will take home. It might be a fair description of the group. It is not a promise to any one member of it.
People learn this the hard way, and they usually give it a softer name than ruin. They size a position as if they were the average of a thousand accounts. Then one bad week takes them out, and the expected value keeps existing in a spreadsheet that no longer has their name on it.
The practical response to this math is older than the word “ergodicity.” Size so that you are optimizing the growth rate of the path you actually walk, not the average of paths you don’t. That’s what the Kelly criterion is doing, whether or not the person using it has ever heard of Ole Peters. I wrote about that earlier, in Why Kelly Criterion Is a Life Philosophy, Not Just a Betting System. The point here is upstream of the formula. If you optimize the wrong average, the formula just helps you be wrong more carefully.
The Ledger, Not the Average
This is why I care more about the shape of a record than about a headline return.
Obsidian Quant has traded a live-market paper account, funded at $100,000 on February 17, 2026, on real prices with broker-simulated fills. No client capital has been deployed. I’m saying that plainly, the same way I have before, in Paper Money Doesn’t Sweat.
It is not a backtest. A backtest is a story you tell after you already know the ending. This is a research ledger: decisions made in sequence, on data that had not happened yet, with the losing days and the software defects left in the file. The performance page is public, and the figures on it are shown as a cumulative path over a stated period, not annualized. A record this short cannot support a yearly number without pretending the next months will behave like an average of the last ones. That’s the same mistake in a different suit.
If the only number you take from that page is the total return, you’ve missed the argument. The useful question is not what a thousand copies of this account would have averaged. It’s what this one account actually did, in order, including the stretches that would have made a person reach for the off switch. That’s the path. That’s the thing you get.
A spreadsheet can tell you the expected value of a bet. It cannot tell you whether you will still be in the game to collect it.
- The coin-toss figures — +5% expected growth per round, a time-average growth rate of a little worse than −5% per round, and $100 becoming $90 after a head and a tail — come from Ole Peters, “The ergodicity problem in economics,” Nature Physics 15, 1216–1221 (2019), and from his explanation in “The infamous coin toss.” The same gamble is developed in Ole Peters and Murray Gell-Mann, “Evaluating gambles using dynamics,” Chaos 26, 023103 (2016). The time-average multiplier per round is √(1.5 × 0.6) ≈ 0.9487, or about a 5.13% decline; the −5% figure in the body follows Peters’ usual rounding.
- Obsidian Quant trades a live-market paper account, funded at $100,000 on February 17, 2026 — real market data, real prices, broker-simulated fills. No client capital has been deployed. Returns on the public page are presented as a cumulative path over the stated period and are not annualized. The performance page is public.
Obsidian Quant publishes a live-market paper account — real prices, simulated fills, no client capital, the losing days left in. Look at the sequence, including the ugly stretches, not at a single number taken out of it. The ledger is public.