A bet can look great "on average" across a crowd of people playing it once, and still ruin almost everyone who plays it over and over. Ergodicity is the question of whether those two pictures — the average across many people, and the path one person actually walks through time — agree. For most of life's repeated, compounding decisions (money, health, career bets), they don't.
When a choice is non-ergodic, the math you learned for one-shot decisions — multiply each outcome by its probability and add them up — quietly stops describing what happens to you. It keeps describing what happens to a crowd.
| Type of thinking | Quantitative / probabilistic — distinguishes two different kinds of "average" |
| Best for | Any decision you (or your resources) face more than once in sequence: investing, pricing risk, career moves, health habits, startup bets |
| Danger if misread | High — used to justify both reckless leverage ("the odds favor me") and excessive caution ("never take any risk") |
| Pairs well with | Expected Value, the Kelly Criterion, Antifragility, Margin of Safety |
| Origin | Formalized for economics by physicist Ole Peters, building on the 18th-century St. Petersburg paradox and 20th-century growth-optimal betting theory |
Imagine 1,000 people each flip a coin. Heads, their money grows 50%. Tails, it shrinks 40%. Flip once, and the average result across the whole crowd is a 5% gain (0.5 × 1.5 + 0.5 × 0.6 = 1.05). That crowd-wide average is called the ensemble average — it treats every person's single flip as one sample from a big parallel population.
Now have one person flip the same coin 50 times in a row, each time betting their whole updated balance. That person isn't sampling a population — they're walking a single path through time, and each flip multiplies what the last flip left them. This is their time average.
Because the outcomes multiply instead of add, these two averages are not the same thing, and the gap between them is the entire mental model.
With additive outcomes (gain $10 or lose $8 on a fixed stake), repeating the bet many times converges to the same average whether you look across people or across time. Ergodic systems behave like this — the two averages agree.
With multiplicative outcomes — and almost everything that compounds does multiply: savings, debt, a company's revenue, a reputation, a body's health — one bad multiplier can't be "averaged away" by later good ones the way an addition can. A single ×0 (total ruin) ends the sequence no matter how good the other 49 flips were. The relevant average for a repeated multiplicative process is the geometric mean of the outcomes, not the arithmetic mean — and the geometric mean of our coin (√(1.5 × 0.6) ≈ 0.949) is a 5.1% loss per flip, not a 5% gain.
The seed of this problem is the 1738 St. Petersburg paradox: a coin-flip game whose textbook expected value is infinite, yet no sane person would pay more than a modest amount to play it. Daniel Bernoulli's historical fix was to argue that money has diminishing utility. Physicist Ole Peters offered a different resolution starting around 2011: the paradox isn't about utility at all — it's that the expected-value calculation silently averages across an ensemble of parallel players, a quantity no single, time-bound player ever actually experiences. His 2019 paper "The Ergodicity Problem in Economics" (Nature Physics) formalized this as a general critique of how standard economic theory handles risk.
The same insight underlies the Kelly criterion (John Kelly, 1956), a bet-sizing formula from information theory that maximizes long-run time-average growth rather than one-shot expected value — it is, in effect, the practical answer to "how much should I bet, given that I'll be the one living through every round."
Ergodicity economics is an active, respected research program, not settled consensus. Mainstream economists generally agree the St. Petersburg-style examples are real, but some argue expected-utility theory already handles them adequately (through risk aversion) without needing a new ergodicity framework, and that in practice most real-world bets are diversifiable or poolable, which blunts the critique. Treat "ergodicity" as a sharp, useful lens for spotting ruin risk in repeated multiplicative decisions — not as a fully settled replacement for expected-value thinking in every context.
Set a win size, a loss size, and a win probability for a repeated multiplicative bet. Then compare what a crowd playing once averages to, against what one person playing many rounds in a row actually tends to end up with.
Each click plays one random 50-round sequence for a single wealth path, starting at 1.0. Click it several times — notice how often the line trends toward zero even when the "crowd average" box above stays positive. That gap is ergodicity.
| Domain | The non-ergodic trap | The ergodicity-aware fix |
|---|---|---|
| Personal investing | Using leverage sized to maximize expected return, ignoring that one bad sequence of margin calls can force liquidation at the worst time | Size positions so no realistic sequence of losses forces you out of the game (Kelly-style sizing, not pure expected-value sizing) |
| Startup career bets | A solo founder treating their one startup like a venture fund's diversified portfolio ("most startups fail, but the expected value is huge") | Recognize you only get to live your own single path; stage the bet (keep income, raise smaller tranches) so a single bad draw doesn't end your runway |
| Insurance | Skipping insurance because the expected payout to the insurer is positive (that's why insurers sell it) | Buy insurance against the specific losses that would be ruinous to your one life, even at a negative expected value — the ensemble math isn't the one you live through |
| Health and training | Pushing recovery and intensity like an elite-athlete average, ignoring that one serious injury removes you from training entirely | Keep a margin against the single bad outcome (injury, burnout) that halts your personal compounding, even if it costs average performance |
| Business pricing / risk pooling | A small firm self-insuring a rare catastrophic loss "because the expected cost is low" | Transfer tail risk to a party that genuinely operates at ensemble scale (a reinsurer, a larger pool) instead of carrying ruin risk on one balance sheet |
| Public policy | Justifying a policy by its expected social benefit while ignoring that it concentrates catastrophic downside on a few unlucky individuals who don't get to re-average across the population | Weigh policies partly by how bad the worst individual-level realized path can get, not only by the population-level expected benefit |
Scenario: You're deciding whether to take a loan to double your position in an investment that has a genuinely favorable expected return, because "the math says it pays off on average."
Describe a repeated or compounding decision you're weighing, and get a ready-to-use prompt that walks an AI assistant through an ergodicity check on it.
This packages the Ergodicity model as a portable instruction set you can hand to an AI assistant, so it checks your repeated or compounding decisions for the ensemble-vs-time-average trap automatically.
Describe a specific situation, and get a prompt that applies the Ergodicity skill to it directly.