Ergodicity and Time

There is a number that looks like a money machine and is actually a slow-motion bankruptcy. The expected value says take the bet, repeatedly. The lived experience says you went broke. Both are correct — they are just answering different questions, and almost everyone confuses them. This is ergodicity: the most important word in probability that your finance class never mentioned. It is the analytical crux of everything Taleb says about risk, ruin, and skin in the game. By the end you’ll see why the average over a crowd and the average over your life can point in opposite directions — and why, when wealth compounds, it is the second one that pays your bills.

Ensemble average vs time average

Analogy. Imagine a thousand people each spinning the same risky roulette wheel, all at noon. The ensemble average is what you get by snapping a photo of the whole crowd at noon and averaging everyone’s chips — an average across people, at one instant. Now ignore the crowd and watch one stubborn player spin that wheel for forty years. The time average is what that one trajectory earns per spin over its lifetime. Ergodicity is the question: do these two averages agree?

Definitions.

• Ensemble average — the expected value $E[\cdot]$ across many parallel copies of the system at a single moment. The mean over the crowd.
• Time average — the long-run average growth rate of a single trajectory followed through time.
• Ergodic — a process where the two coincide. Watching one path forever tells you the same thing as photographing the crowd.
• Non-ergodic — a process where they diverge. The crowd’s fate and your fate are different stories.

Why wealth is non-ergodic. Wealth doesn’t add up — it multiplies. Each period your money is scaled by a growth factor (a return), and there is an absorbing barrier at zero: hit it and the trajectory stops forever. Multiplicative dynamics plus an absorbing wall break ergodicity. As Ole Peters, the physicist who put this on a rigorous footing, puts it: the expected value secretly assumes you can “interact with copies of yourself in parallel universes” and pool the results. You can’t. You live exactly one path, and on that path a single −100% is permanent.

The coin flip that breaks the bank

This is the example. Internalise it and the rest follows. We start with $100 and play this game: a fair coin, each round.

• Heads — multiply your wealth by 1.5 (a +50% gain).
• Tails — multiply your wealth by 0.6 (a −40% loss).

The ensemble average looks delicious. The expected multiplier per round, averaged across a crowd of players, is

$E[\text{factor}] = 0.5(1.5) + 0.5(0.6) = 1.05.$

That’s +5% per round. Photograph a million players after one round and their average wealth is up 5%. It reads like a money printer.

The time average is a slow death. Now follow one player. Over a long run heads and tails arrive in roughly equal numbers, and the multipliers don’t add — they compound. The growth rate of a single path is the geometric mean of the factors, not the arithmetic one:

$g = \sqrt{1.5 \times 0.6} = \sqrt{0.9} \approx 0.9487.$

That’s about −5.1% per round. Concretely: one head then one tail takes $100 → $150 → $90. Two rounds, one of each, and you’re down 10% — about −5% per round, every cycle. The order doesn’t matter ($1.5 \times 0.6 = 0.6 \times 1.5 = 0.9$): any pairing of an up and a down loses you a tenth of your money. Almost every individual player grinds toward zero.

Here is the whole paradox in one table — the same game, two averages, opposite verdicts:

Read that last pair again. After 100 flips the mean ending wealth is about $16,697 — the ensemble average kept its promise. But the median player, the one in the middle, ends with about 52 cents. The mean is hauled skyward by a microscopic handful of astronomically lucky paths that no typical player ever experiences. The crowd’s average is a fiction stitched together from lottery winners; the person you’d actually be went broke.

The misconception this kills: “the EV is +5%, so just take the bet over and over.” Positive ensemble expected value does not imply positive time-average growth. The arithmetic mean answers “what’s the average across the crowd?”; the geometric mean answers “what happens to me?” When wealth compounds, only the second one is decision-relevant.

Multiplicative vs additive: volatility drag and Kelly

Why the two averages split. It comes down to whether your bets add or multiply.

• Additive bets — fixed-dollar wagers from a bankroll you don’t reinvest. Outcomes sum, so the arithmetic mean is decision-relevant and the process is ergodic. The crowd’s average is your average.
• Multiplicative bets — you reinvest, so this period’s wealth scales next period’s. Outcomes multiply, so the geometric mean — equivalently the average of log wealth — is what governs your path. Logs turn products into sums, which is exactly why the log is the right lens for compounding.

Volatility drag. A foundational fact: the geometric mean is always $\le$ the arithmetic mean (the AM–GM inequality), with equality only when there’s no variation. The gap between them is volatility drag — growth that volatility quietly eats. The brutal asymmetry behind it: a −40% loss does not need a +40% gain to recover. It needs

$\frac{1}{1-0.40} - 1 = \frac{1}{0.6} - 1 \approx 0.667 = +66.7\%.$

You lose 40% in one step and have to earn two-thirds back just to break even. Down moves dig a deeper hole than up moves fill, so wiggling around an average drifts you below it. (We met this exact arithmetic in the risk-of-ruin course as the recovery math; here it’s the engine that splits the two averages apart.)

The same idea over a whole range of volatilities: hold the average return fixed and crank up the wobble, and your realised compound growth slides down a parabola — roughly arithmetic mean minus half the variance.

Maximising the time average = Kelly. If the decision-relevant quantity is the geometric mean (expected log wealth), then the right objective is to maximise expected log wealth — and that is precisely the Kelly criterion you met in the kelly-and-cagr course. Kelly isn’t a separate trick bolted on; it is what you get the instant you take ergodicity seriously and optimise the time average instead of the ensemble one.

Worked example — Kelly on our coin. Reframe the bet as risking a fraction $f$ of bankroll at net odds $b$. Our coin pays +50% on a win and costs −40% on a loss, so per unit risked the win pays $b = 0.5/0.5 = 1.25$ and $p = 0.5$. The Kelly fraction is

$f^* = \frac{p(b+1) - 1}{b} = \frac{0.5(2.25) - 1}{1.25} = \frac{0.125}{1.25} = 0.10.$

Bet 10% of bankroll, not 100%. Now re-run the per-round factors at that size:

Same wager, same coin, same edge. The only thing that changed is position size — and it flipped the time-average growth from −5.1% to +0.4% per round, positive. Survival here is a property of sizing, not of the bet. Full exposure to a +5%-EV game is a losing strategy; a tenth of that exposure is a winning one.

The absorbing barrier: ruin makes the path the whole story

The wall at zero. What truly breaks ergodicity is the absorbing barrier: once wealth hits zero, the trajectory stops. There is no rebound, no reversion, no next round. A single ruinous outcome doesn’t just dent the average — it ends the path, which is why the time average is so violently path-dependent. The crowd can absorb a few deaths and report a healthy mean; you cannot absorb your own death and keep playing.

Ruin compounds toward certainty. Suppose each round carries a small probability $p$ of a ruinous outcome. Survive all $n$ rounds only if you dodge ruin every single time, so survival is $(1-p)^n$, which marches to zero as $n$ grows. Even with a glittering positive expected value, $P(\text{ruin}) \to 1$ under enough repetition.

A 1% chance of disaster per round sounds like rounding error. Run it 300 times and you are 95% likely to be ruined. (This is the same machinery as the risk-of-ruin course’s repeated-exposure result — here it’s the reason a non-ergodic process can have a positive expectation and still kill almost everyone who plays it long enough.)

The misconception here is seductive: “a low per-trial ruin probability is safe if I just repeat the game.” The opposite is true — repetition is exactly what compounds a tiny per-trial ruin probability into near-certain ruin. With an absorbing barrier, more plays is more danger, not the safety the law of large numbers seems to promise. (The law of large numbers needs you to keep playing; ruin revokes that permission.)

Why this reframes everything

Ensemble thinking hides the dead. When you observe a population once and average it, the casualties have already been removed — the dead don’t show up in tomorrow’s cross-section. A backtest of “strategies that survived,” a survey of “funds still operating,” an average return computed over “investors still in the market” — each silently deletes the ruined and reports the survivors’ mean as if it were a typical fate. Ergodic reasoning launders ruin out of the data. The time-average view refuses to: it follows one path and counts the absorbing barrier as the permanent, decision-ending event it actually is.

Survival is the real objective. Once you accept that you live one non-ergodic path, the whole Taleb toolkit clicks into place. Skin in the game matters because the decision-maker must share the trajectory, not the ensemble. Avoiding ruin dominates maximising expected value, because a path that touches zero has no future to average over. Antifragility is the strategy of staying in the game long enough for convexity to pay. Peters reframes even the oldest puzzle in the field: risk aversion is “a residual of ergodicity” — not an irrational quirk of human psychology, but the sensible behaviour of someone correctly optimising a time average and steering clear of the absorbing barrier. People aren’t being timid; they’re being non-ergodic, which is to say, correct.

Cost–benefit analysis is dangerous when risk is systemic and repeated. Weighing expected benefits against expected costs is fine for additive, poolable, one-shot risks. It is reckless when the downside is ruinous and the exposure repeats — because the ensemble math says “the average is fine” while the time math says “you are eventually wiped out with probability 1.” As Taleb puts it: “In a strategy that entails ruin, benefits never offset risks of ruin.” No expected value, however gaudy, buys back a path that has ended.

The big picture

Ergodicity is the hinge: ensemble average is the mean across parallel copies at an instant; time average is the growth of one trajectory over time. They coincide only for ergodic processes — and wealth, being multiplicative with an absorbing barrier at zero, is not one. The coin game makes it concrete: +5% ensemble EV sits on a −5.1% time-average growth ($\sqrt{0.9}$), so the crowd’s mean soars while the typical player goes broke. The split is driven by volatility drag (geometric $\le$ arithmetic by AM–GM), the fix is position sizing (maximising log wealth is Kelly), and the hard constraint is the absorbing barrier (a tiny per-round ruin probability compounds to near-certain ruin). The upshot reframes risk itself: survival over time — skin in the game, avoiding ruin, antifragility — is the objective, and in a strategy that entails ruin, benefits never offset it.

Next, we carry this lesson’s lens — you live one non-ergodic path, so protect it — into the practical machinery of staying in the game: barbells, optionality, and convexity, where surviving long enough is the whole strategy.