Where the Model Lies

You have spent five lessons learning to distrust the average. That is a powerful reflex — and like every powerful reflex, it can misfire. A model you can only agree with, never criticise, has stopped being a model and become a superstition. So this capstone does the opposite of cheerleading: it names exactly where ergodicity stops applying, where invoking it is empty, and where it gets confused with things it is not. The point isn’t to weaken the tool. It’s to sharpen it — so you draw the blade only when there’s something to cut.

Two mirror-image errors bookend this course. The first, which we spent five lessons on, is treating a non-ergodic process as if it were ergodic — believing the crowd’s +5% is your future. The second, which this lesson is about, is treating everything as non-ergodic — flinching at every average as though it were a trap. Both are wrong. The skill is telling them apart.

Not everything is non-ergodic

Analogy. A smoke detector that shrieks at toast is not a safety device — it’s noise you’ll learn to ignore. If your ergodicity alarm goes off at every average, it can’t tell you anything when a real fire starts. A useful alarm stays silent for burnt toast and screams for the kitchen ablaze.

The precise point. Genuinely additive, bounded, repeatable, non-ruinous bets can be judged on the ensemble mean — and that is exactly the situation expected value was built for. Non-ergodicity needs two ingredients working together: multiplicativity (the outcome scales your whole stake) and a reachable absorbing barrier (a state like ruin you can’t come back from). Strip out either one and the two averages converge again. A bet that merely adds a fixed amount, that can’t compound you toward zero, that you can genuinely repeat many independent times — that bet is effectively ergodic, and the expected value is the honest number.

Worked example. Three cases that look dangerous but aren’t:

• A casino booking millions of tiny independent bets with deep reserves. From the house’s side, each $1 bet adds a sliver to a huge bankroll; no single spin dents it; the realised return per bet converges to the theoretical 2.7% edge — the ensemble mean, made real. The house is effectively ergodic even while every gambler facing it is not.
• A large diversified insurer pooling millions of uncorrelated policies. Any one claim is a rounding error against the pool; the law of large numbers does exactly what the textbook promises, and the expected loss per policy is the number to price on.
• A tiny stake relative to your wealth — one dollar of a millionaire’s fortune. A −40% “multiplicative” loss on that dollar is 40 cents; it’s approximately additive because it can’t move your base. The multiplicativity is technically present but too small to bite.

The correction. Treating every average as a lie is its own error — the exact mirror image of the mistake the course has been hunting. The discipline is symmetric: don’t assume ergodicity by default, but don’t assume non-ergodicity by default either. Run the test each time. Multiplicativity plus a reachable absorbing barrier is what makes the model bite; absent both, the ensemble mean is right.

When to use it

Before you flinch at any average, ask the two-part question: does this scale my whole stake, and is there a state I can’t return from? Only if both are true does the ensemble mean lie. A fixed-dollar side-bet with pocket change fails the test — so trust its expected value and move on.

Ergodicity is about the dynamics, not a mood

Analogy. “Non-ergodic” is a diagnosis, like “this bridge has a fatigue crack.” It is not a temperament, like “I have a bad feeling about bridges.” One points at a measurable feature of the structure; the other is a vibe. Only the first tells an engineer where to weld.

The precise point. Ergodicity is a mathematical property of a process — do the time average and the ensemble average converge, yes or no? It is not a synonym for “be pessimistic” or “expect the worst.” You can be a raging optimist about the future and still correctly observe that your leveraged, all-in position is non-ergodic; the dynamics don’t care about your outlook. And conversely, chanting “non-ergodic!” as a mood — without pointing at actual multiplicative dynamics or a real ruin barrier — is empty. It’s using a precise word as an incantation.

Worked example. Two founders describe the same venture. One says, “I’m betting everything, and if it fails I’m wiped out with no reserve — the dynamics are multiplicative with an absorbing barrier, so I should size for survival, not for the expected valuation.” The other just says, “Startups are risky and scary, so, you know, non-ergodic.” The first has named a mechanism; the second has borrowed a word to dress up a feeling. Same company, but only one of them is actually using the model.

The correction. Keep the concept anchored to its definition. If someone (including you) says “non-ergodic,” the follow-up is always: show me the two averages diverging, or show me the multiplicativity and the barrier. A property you can measure is a model. A mood you can only feel is a horoscope.

When to use it

Use ergodicity as a lens for analysis, not as a permission slip for anxiety. When you catch yourself — or a colleague — invoking it emotionally, translate the claim back into its mechanical form: which quantity multiplies, and what is the absorbing state? If you can’t fill in those blanks, you don’t yet have a non-ergodicity claim; you have a feeling.

It is not the same as risk-aversion

Analogy. Risk-aversion is a taste — like disliking spicy food. Non-ergodicity is a fact about the kitchen — like the fact that this particular dish, cooked this particular way, will actually poison you. A person with no aversion to spice at all should still not eat the poison. The two live in different categories entirely: one is about what you prefer, the other about what the process does.

The precise point. Risk-aversion is a preference — a concave utility curve, a dislike of variance, a willingness to pay to smooth outcomes. Non-ergodicity is a fact about which average is decision-relevant. Crucially, non-ergodicity bites even for a perfectly risk-neutral agent — one who feels nothing about variance and wants only to maximise long-run wealth. Such an agent still refuses the +50%/−40% coin, not because they dislike the swings, but because the time-average growth rate of full-staking is negative. The refusal falls out of the dynamics, with no taste for or against risk anywhere in the argument.

Worked example — the live debate, presented fairly. Here is where honesty matters most, because this is contested science.

• Ole Peters’ provocative claim. Peters and the London Mathematical Laboratory argue that much of what economics models as “risk aversion” — via expected-utility theory, where you posit a concave utility function to explain why people decline attractive-looking gambles — may instead be plain, preference-free time-average optimisation. On this view you don’t need a special utility function to refuse the +50%/−40% coin; you only need the right average. The concave utility, Peters suggests, was a patch invented to reproduce behaviour that the time average predicts directly from the dynamics.
• The mainstream reply, presented just as fairly. Standard expected-utility theory answers that it already declines that bet. Maximise the expected value of a log utility of wealth and you refuse the coin and recover Kelly betting exactly — the same prescription this course reached. So many economists read ergodicity economics as a reframing and a pedagogy — a cleaner, more intuitive derivation of results the field already had — rather than a refutation of expected-utility theory.

The correction. This is a live, contested research program — Peters and the London Mathematical Laboratory on one side, the expected-utility mainstream on the other. The honest teaching move is to hold both: non-ergodicity is genuinely a different category from risk-aversion (fact vs preference), and whether it overturns or merely re-derives expected-utility theory is an open question smart people still disagree about. Teach the debate; don’t pick a winner.

The unfalsifiable-story trap

Analogy. “It was non-ergodic” can decay into the astrologer’s move: whatever happened, the stars foretold it. A prediction that survives every outcome predicts nothing. If “non-ergodic” is the label you paste onto any bad result after the fact, it has become a horoscope with a physics degree.

The precise point. After a loss, it is tempting to intone “well, it was non-ergodic” as though that explained anything. But an explanation that fits every possible outcome is not an explanation. The discipline that keeps the model falsifiable is to name the mechanism in advance: Where is the multiplicativity? Which quantity scales? Where is the absorbing barrier? What is the state you can’t return from? Can you show the two averages diverging? If you can answer those, you have a model that could have been wrong and wasn’t. If you can’t, you have a label, not a model — a comforting story stapled onto hindsight.

Worked example. A trader blows up and shrugs, “non-ergodic markets, what can you do.” Contrast a colleague who, before the trade, wrote: “This position is leveraged 5×, so a 20% adverse move triggers a margin call — an absorbing barrier — and returns compound multiplicatively off the remaining equity; the ensemble mean flatters this, so I’ll size at a quarter of the tempting position.” The second statement is falsifiable: it points at a specific mechanism (5× leverage, the 20% barrier) that either was or wasn’t present. The first is unfalsifiable consolation.

The correction. Demand the mechanism, especially from yourself, and especially before the outcome is known. The test — multiplicativity plus a reachable barrier plus demonstrable divergence — is what separates a working model from a fatalistic slogan. If you only ever reach for “non-ergodic” in the post-mortem, you’re not analysing; you’re grieving.

When the ensemble can become you

Analogy. Alone in a lifeboat, one leak can sink you — a non-ergodic, path-dependent problem. Lash a hundred lifeboats together and share the water, and a single leak is now a hundredth of everyone’s problem. You didn’t repeal physics; you changed the dynamics by pooling the paths, until the fleet’s average became something each boat could actually live.

The precise point. Pooling, insurance, diversification, and risk-sharing can convert a non-ergodic individual problem into an effectively ergodic collective one. When many people can share a path — spreading each person’s downside across the group — the crowd’s average becomes reachable for the individual in a way it wasn’t when they faced the barrier alone. The absorbing state that made the problem non-ergodic gets diluted below the level where it can absorb any single member. This isn’t a loophole; it is arguably the reason cooperation, insurance markets, and diversified funds exist at all — they are machines for making the ensemble mean claimable.

Worked example. One farmer whose entire harvest can fail in a bad year faces a non-ergodic gamble: a single ruinous season can end the farm. Now let a thousand farmers with uncorrelated harvest risks pool into a crop-insurance mutual. Each pays a small, additive premium; the pool absorbs individual disasters; and because the risks are uncorrelated, the pool’s aggregate outcome converges to the ensemble mean by the law of large numbers. Each farmer has swapped a non-ergodic personal gamble for a small, bounded, additive cost — and the collective now genuinely lives the average that no single farmer could. (The catch, as ever, is the word uncorrelated: pool a thousand farms hit by the same drought and you haven’t diversified at all.)

The correction. Non-ergodicity isn’t only a warning; it’s also a design brief. Once you see that pooling can move a problem from the non-ergodic column to the ergodic one, you understand why so much of civilisation is built out of shared-risk institutions. The model doesn’t just tell you which bets to fear — it tells you how to build structures that make fearsome bets survivable.

The net: help vs mislead

Analogy. Ergodicity is a scalpel, not a hammer. A scalpel is extraordinary for the incision it’s shaped for and useless — worse, dangerous — swung at everything. Knowing where it cuts is the whole of the skill.

The precise point. Here is the honest ledger of the model.

Where the model HELPS — reach for the time average whenever wealth compounds and ruin is possible:

• Position sizing — how much of your bankroll to stake, where full-staking silently guarantees decay.
• Leverage — which sharpens multiplicativity and pulls the margin-call barrier within reach.
• Insurance — where a rare catastrophe is an absorbing state, and paying a small additive premium to avoid it is the geometrically-optimal move even at a “negative expected value.”
• One-shot life gambles — the all-in career, the undiversifiable business, the bet you only get to make once.

Where the model MISLEADS if over-applied — do not invoke it for:

• Small, bounded, repeatable bets — the fixed-dollar side-stake with pocket change, where the ensemble mean is exactly right.
• Confusing it with a mood — using “non-ergodic” as a synonym for pessimism instead of a claim about dynamics.
• Confusing it with mere risk preference — mistaking a fact about which average governs your future for a taste for or against variance.

The correction. A model held honestly is a model you can point away from. The test you now own — multiplicativity plus a reachable absorbing barrier — is exactly what tells you which column you’re in. Use it in both directions: to sound the alarm when the fire is real, and to stay silent when it’s only toast.

Sort the cases

Apply the test — multiplicativity plus a reachable absorbing barrier — to each case, and drop it in the right bucket.

Recap

That is the whole course, held honestly — from the first divergence of the two averages to the boundaries of where the idea applies. Next up is the Final Exam: a graded, one-shot run across everything you’ve built. No Back, no retry — fitting, for a course about paths you only get to walk once.