When Expected Value Lies: Fat Tails, Ruin, and the Ergodicity Trap

Expected value is the hero of this whole latticework. Multiply each outcome by its probability, add them up, take the biggest number — it’s how you weigh a bet, price an insurance policy, and choose between options. So this lesson is going to feel like betrayal, because it’s about the three places that hero turns on you. Not because the arithmetic is wrong — it’s flawless — but because of the assumptions it smuggles in when the future is deeply uncertain. By the end you’ll still love expected value. You’ll just know exactly when not to trust it with your life.

Lie #1: the point estimate hides the spread

The first betrayal is compression. An expected value is a single number, and a single number can only ever be a summary — it throws away the shape of what could happen. Two bets can share an identical expected value of “+5” while being wildly different animals: one pays you between +4 and +6 every time (all outcomes hug the average), the other pays +6 almost always but −1,000 once in a blue moon (a rare catastrophe lurking under a friendly-looking mean). The average is the same. The experience is not remotely the same.

Under fat tails — the world you met earlier, where one rare extreme dominates everything — this is deadly. The mean is dragged around by tail events so rare you may never have seen one in your data, so a confident-looking expected value can be resting entirely on a catastrophe you haven’t sampled yet. The point estimate says “+5, relax.” The distribution says “…unless the once-a-century world shows up, in which case you’re finished.” A number cannot warn you about a shape it has already averaged away.

Lie #2: the odds were invented in the first place

The second betrayal we met last lesson: expected value requires probabilities, and under deep uncertainty you don’t have real ones. So you guess — and then the machine launders your guess into a decimal-pointed verdict that looks like knowledge. This is false precision. A detailed model with invented inputs is not more reliable than a rough guess; it’s a rough guess wearing a lab coat. The danger is that the elaborate spreadsheet feels rigorous, so you trust it more and hedge less — the exact opposite of what deep uncertainty demands.

The failure has a name in the trade: garbage in, gospel out. The more elaborate the model, the more authority it seems to carry, and the easier it is to forget that every “precise” output is only as sound as the made-up odds at the bottom. A seven-decimal-place answer built on a shrug is still a shrug.

Lie #3: the ergodicity trap — the average that isn’t yours

This is the deepest and least intuitive betrayal, and it’s the one that reframes the entire course. Expected value is an average across many parallel worlds — a kind of “if a thousand copies of you each took this bet once, what would the crowd average?” But you are not a thousand copies. You are one person, taking your bets one after another, through time. And when there’s any chance of ruin — of hitting zero, of going bust, of being knocked out of the game — those two averages are not the same. This gap has a name: ergodicity (a system is ergodic when the average across the crowd equals the average across time for one individual — and betting-with-ruin is not ergodic).

Here’s the classic gut-punch. A wager pays +50% of your wealth on a coin’s heads and −40% on tails. The ensemble average — across many parallel bettors — is positive: gain 50, lose 40, mean +5% per round, a “great” bet by expected value. So you should keep playing forever, right? Now trace one person playing it round after round. Win then lose: 1.5 × 0.6 = 0.9 — you’re down to 90% of where you started. Every heads-then-tails pair multiplies your wealth by 0.9, so over time, almost surely, one real player grinds toward zero. The crowd’s average soars while the typical individual goes broke, because the crowd average is inflated by a handful of insanely lucky paths that a single person will never live.

The moral flips a cornerstone of naive expected-value thinking: a bet with a positive average can still be a near-certain path to ruin for the person actually living it. Survival isn’t one consideration among many — it’s the precondition for all the future bets that were supposed to make you rich. Zero is a trapdoor, not a low number, because after it there is nothing left to compound.

See it: the many-worlds explorer

Enough words — go drive it. Below, each strategy faces a whole spread of possible worlds: most ordinary, a few rare and ruinous (catastrophes on the left) or rare and wonderful (booms on the right). You don’t know which world you’ll get, so watch how each strategy does across all of them at once. Start on Optimise — the expected-value champion that bets everything on the single most-likely world — and drag the turbulence slider from calm toward deep uncertainty.

Notice what the numbers do. In a calm future (turbulence low), Optimise posts the best average by a mile — exactly why it’s so seductive. Now crank turbulence up. Its average may still look respectable, but its worst world plunges below the ruin line and its worlds survived count starts dropping: there are now futures in which this strategy simply blows up. Meanwhile the boring Robust strategy never touches the ruin line — it trades away the flashy average for the one thing that matters when the future is unknown: it comes home alive in every world. That trade is the entire course, and you just watched it happen.

The through-line

The three lies share one root: expected value quietly assumes a well-behaved, knowable, survivable world — thin tails, real probabilities, no trapdoor at zero. Under deep uncertainty all three assumptions wobble. The tails are fat and hide catastrophes you haven’t sampled; the probabilities are invented; and ruin turns the comforting average into a lie about your actual fate. That doesn’t mean throw the tool away — it means cap its authority. Use expected value as one input, never the verdict, and put an unbreakable rule above it: first, don’t get wiped out. That rule is the doorway to the next lesson.

Next up: Robustness Over Optimality — having seen why optimising a fragile average can kill you, we build the alternative goal from the ground up: satisficing, robust decision-making, minimax-regret, and the ruin-avoidance principle that sits above every calculation.