Ergodicity: Why Average Outcomes Are Not Your Outcomes
Imagine a casino that advertises an average payout of 110 percent. Sounds profitable. Now imagine the game: flip a coin. Heads — your stake grows by 50 percent. Tails — your stake falls by 40 percent. The arithmetic average of +50% and -40% is +5%. But if you play this game repeatedly, you will eventually go broke. The geometric mean — what actually happens to your money over time — is negative. The game is a loser despite a positive arithmetic average.
This is the ergodicity problem. A system is ergodic if the time-average of a single player equals the ensemble-average of many players playing once. Most systems that matter — finance, career risk, health decisions — are not ergodic. What happens to the average person is not what happens to any actual person over time.
The ensemble versus the time series
The distinction sounds academic until you realize it changes every risk calculation you've ever made. When a financial analyst says "the expected return is 7 percent annually," they're computing an ensemble average: average across all possible outcomes, or across many investors. But you are not a diversified portfolio of yourself. You are one sequence of events. If any point in that sequence goes to zero — bankruptcy, serious illness, a single catastrophic loss — the sequence ends regardless of what the ensemble average says.
Nassim Taleb calls this "ruin." Ole Peters, who formalized the ergodicity problem for economics, puts it this way: maximizing expected value is wrong for non-ergodic systems. The correct goal is to maximize the time-average growth rate, which means avoiding outcomes that permanently reduce your capacity to participate.
What this means for decisions
The practical implication is that you should not make irreversible bets — even positive-expected-value bets — if losing them ends the game for you. Kelly's criterion formalizes this: the optimal bet size is the fraction that maximizes the logarithm of wealth, not the raw expected value. For most real-world risks, Kelly implies betting far less than the expected value calculation suggests.
Consider a startup founder choosing between two strategies. Strategy A: 80 percent chance of 2× return, 20 percent chance of total loss. Expected value: positive. Strategy B: certain 30 percent gain. For an ensemble of founders, A wins. For a single founder with no capital buffer and a family to support, B may be rational — because a single instance of the 20 percent tail ends the time series permanently.
India's informal economy and ergodicity
India's hundreds of millions of workers in the informal sector face deeply non-ergodic conditions. A daily wage labourer has no savings buffer; a single month without work — illness, a construction project pause, a monsoon shutdown — can trigger a debt spiral that takes years to exit. The ensemble average of informal sector wages looks reasonable in national statistics. The time-average for an individual facing this sequence of risks does not.
This is why microfinance institutions that ignore ergodicity — pushing loan products that maximize expected portfolio returns but expose borrowers to ruin risk — routinely produce debt crises even when their average borrower does fine. The average borrower and the typical sequence of events for an individual borrower are different objects.
The honest question
Before evaluating any risk, the ergodicity question is: if I lose, can I still play? If the answer is no, the expected value calculation is wrong for your situation regardless of what it says for the ensemble. Survivorship — staying in the game — is a prerequisite for capturing any long-run average. A strategy that risks ruin is a strategy that trades away the time series for a chance at ensemble statistics you will never live long enough to collect.
Quick answers
What is Ergodicity?
An ergodic system is one where the average across many people equals the average over time for one person. Most real-world risks aren't ergodic — and ignoring this leads to ruin, even when expected value looks fine.
