Ergodicity: Why the Average Trader’s Return Isn’t Your Return

Here is a bet. I flip a fair coin. Heads, your stake grows by fifty percent. Tails, it falls by forty percent.

Work out the expected value. Half the time you multiply by 1.5, half the time by 0.6. The average multiplier is 1.05. The bet returns five percent per flip, and you may play it as many times as you like.

Take it. Obviously take it. Every framework you have ever been taught says take it, and every one of them is correct.

Now play it a thousand times.

You will be broke. Not unlucky, not down on the year. Broke, with near certainty, and the mathematics knew it before you sat down.

Two averages, and only one of them is yours

The five percent is real. It is simply not a number that describes you.

Imagine ten thousand people play one round of this coin simultaneously. Add up all their stakes at the end and divide by ten thousand. You get 1.05 times what they started with. That number is the ensemble average: it averages across people, at a moment in time.

Now imagine one person plays ten thousand rounds. His stake compounds. A win and a loss, in either order, multiplies his stake by 1.5 × 0.6, which is 0.9. He is down ten percent for every pair of flips. Over many rounds his growth rate per flip converges on the geometric mean, the square root of 0.9, which is minus 5.13 percent.

That number is the time average, and it is the only number that describes you, because you are not ten thousand people. You are one person, moving forwards through time, and your losses reduce the base your next gain is calculated on.

What ten thousand rounds actually looks like

Simulate a hundred thousand players, a hundred rounds each.

The typical player has lost more than ninety-nine percent of his stake. The average player has multiplied his by fifty-three. Both statements are true, and they describe the same simulation.

The average is not a lie. It is being carried, almost entirely, by a handful of players who flipped heads far more often than heads deserves to appear. Ninety-six percent of all the money in the system sits with one percent of the participants. The rest financed it.

A detail worth pausing on: theory says the mean after a hundred rounds should be 131.5×. The simulation of a hundred thousand players produced only 53.2×. Not an error. The mean of this process is carried by outcomes so rare that a hundred thousand samples cannot find enough of them. Even the average of the averages understates the average. That is what a non-ergodic process does to your intuition.

Now say it in trading

You have a genuinely good system. Forty percent win rate, winners run to 2R, losers stop at 1R. Expectancy is +0.20R per trade, and it does not change no matter how much you risk. Risk one percent, risk thirty percent, the expectancy in R is identical.

That last sentence is where traders die.

Because expectancy is an ensemble statement. It tells you what happens to the average of many traders taking this trade once. It says nothing about what happens to you taking it four hundred times with a compounding account.

For that, compute the growth rate: 0.4 · ln(1 + 2f) + 0.6 · ln(1 − f), where f is the fraction of the account you risk.

Read the right-hand column, then read the middle one again. The edge never moves. The outcome inverts.

Above about twenty percent risk per trade, a system with a permanently positive expected value has a permanently negative growth rate. It goes to zero. Not because the edge failed, not because of a bad run, not because of anything you did wrong on any individual trade. It goes to zero because the arithmetic of compounding does not care about your expectancy.

The exact zero-growth point here is 20.39% risk per trade, which is 2.04 times the Kelly fraction of 10%. You may have read that the zero-growth point is exactly twice Kelly. That is true in the continuous limit; for discrete bets it lands slightly above. The distinction does not matter to your account. What matters is that a number exists, that it is far lower than most traders assume, and that crossing it converts a good business into a countdown.

Same edge. Same trades. Three sizes.

Twenty thousand traders, five hundred trades each, identical system.

The mean is spectacular at twenty percent risk. Five million times your stake, on average.

The median trader made forty-two percent over five hundred trades, and a third of them lost nine tenths of everything they had.

If you were choosing a position size by looking at expected outcomes, you would choose twenty percent. The number is enormous. It belongs to somebody else.

Why ruin is different from a very large loss

Everything above depends on one property, and it is worth naming precisely.

Zero is absorbing. There is no path back. A ninety-nine percent drawdown requires a hundredfold gain to recover; a hundred percent drawdown requires a gain that does not exist. This is not a matter of degree. It is a discontinuity, and it is the reason you cannot average your way out of it.

The ensemble average is permitted to include the traders who blew up, because in an ensemble their zeroes are diluted by other people’s fortunes. Your time average is not permitted to do that. Once your sequence hits zero it stops, and every future term is zero, and no subsequent good fortune is defined.

What this actually changes on Monday

Stop evaluating risk with expectancy. Expectancy tells you whether the trade is worth taking. It cannot tell you how much to risk, because it is invariant to how much you risk. Two different questions, and the second one is the one that ends careers.

Size with the geometric growth rate. For a two-outcome trade, the growth rate is p·ln(1 + b·f) + q·ln(1 − f). If that number is negative at your risk fraction, you are decaying, and no improvement to your win rate will save you until you fix f.

Treat the Kelly fraction as a ceiling you never reach. Kelly maximises growth. It also assumes you know your win rate exactly, which you do not, and estimation error pushes you above Kelly far more easily than below it. Half of Kelly buys roughly three quarters of the growth for about half the volatility. That is the subject of its own article.

Distrust every average you are shown. Fund returns, prop firm pass rates, the profits of the trader whose thread you are reading. All of them are ensemble statistics computed over survivors. None of them describe the path any individual walked, and the gap between the two grows with every round of compounding.

The line worth keeping

Taleb has spent thirty years saying one thing, and this is the mathematical version of it.

Do not cross a river that is, on average, four feet deep.

The average is true. It is also not the quantity that determines whether you reach the other side. What determines that is the deepest point on the path you personally take, and the fact that drowning, unlike a bad quarter, is not something you recover from and continue averaging over.

Positive expectancy earns you the right to play. Position sizing decides whether you are still playing when the expectancy finally pays.