Ruin & the Real World

Four lessons in, you own the theory: two averages, the flagship coin, additive versus multiplicative dynamics, and the Kelly rule for how much to bet. This lesson spends that theory. It takes the coin’s cold arithmetic and turns it into decisions you can actually make on Monday morning — about how much to stake, whether to borrow, why to buy insurance you “expect” to lose money on, and which parts of your one life you must never gamble. It all falls out of a single, unglamorous fact: you can only lose everything once.

Ruin and the absorbing barrier

Analogy. Think of zero wealth as a black hole with no escape velocity. Every other level of wealth can bounce — you can fall to $10 and climb back to $1,000 by multiplying up. But $0 is sticky: multiply zero by anything, however lucky, and you get zero. There is nothing left to compound. In the language of physics this is an absorbing barrier — a state that, once entered, is never left.

Precise definition. An absorbing barrier is a level (here, zero wealth, or “ruin”) such that once your path touches it, the sequence terminates — no round comes after ruin. This is what makes wealth non-ergodic: your personal path can be permanently removed from the game, so it stops sampling the crowd’s spread. Every glowing expected-value calculation silently assumes you never touch the barrier. Erase that assumption and the whole calculation loses its meaning.

The formula. Suppose each period carries an independent probability $p$ of total ruin. The chance you survive one period is $1 - p$; the chance you survive $n$ periods in a row is $(1 - p)^n$. So the probability of being ruined at least once over $n$ periods is:

$P(\text{ruin in } n \text{ periods}) = 1 - (1 - p)^n$

Worked example. Take a risk so small it feels negligible: $p = 0.001$, a 1-in-1,000 chance of blowing up each period. Play for $n = 1{,}000$ periods and the survival probability is $(0.999)^{1000}$. That is not close to 1 — it is about 0.368. So the probability of ruin is:

$1 - (0.999)^{1000} \approx 1 - 0.368 = 0.63$

A 63% chance of ruin — worse than a coin flip — assembled entirely out of a risk you dismissed as trivial. This is Russian roulette for money: any single pull is probably fine, and that is exactly the trap. Small ruin risks don’t add up gently; they compound into near-certainty.

Pitfall. “The odds of disaster are tiny” is a per-round statement, and you don’t live one round — you live thousands. A tail risk that is negligible once is near-guaranteed if you re-run it enough. The question is never “how unlikely is ruin this time?” but “how unlikely is it that I never hit ruin across every time I play?”

When to use it

Any time a decision repeats and carries even a small chance of a game-ending loss, compute $1 - (1 - p)^n$ before you trust the per-round odds. If that number is uncomfortably large, the “rare” catastrophe is actually your base case, and no expected value can save you from it.

Gambler’s ruin: why the small player is eventually absorbed

Analogy. Two people flip a fair coin for $1 a round: you with $100 in your pocket, the casino with effectively unlimited money. Every session your balance wanders up and down like a drunkard on a cliff edge. The casino’s edge of the cliff is so far away it might as well not exist; yours is right beside you at $0. Given enough wandering, the drunkard falls off the near edge — not because he’s unlucky on average, but because he only has one edge to fall off.

Precise definition. Gambler’s ruin is the classical result that a bettor with a finite bankroll making repeated bets against an opponent with far deeper pockets is eventually absorbed at zero — even at fair odds (50/50), and often even at slightly favourable odds — because the deep-pocketed side can outlast the small player’s swings. The binding constraint is survival, not edge. You can be right on average and still be removed from the table before the average pays out.

Worked example. In a fair (50/50, even-money) game, the probability that a player starting with $i$ dollars reaches a target of $N$ dollars before hitting zero is exactly $i / N$. Start with $100 and dream of turning it into $10,000 (so $N = 10{,}000$): your chance of getting there before busting is $100 / 10{,}000 = 1\%$. The other 99% of the time, the near edge — zero — claims you first. The coin was fair; the bankroll was not.

Pitfall. People conflate “fair odds” with “safe to play forever.” Fairness governs the average drift (zero, here); it says nothing about which absorbing barrier you hit first. With one nearby barrier (your ruin) and one impossibly distant one (the house’s), the near one wins almost every long game.

When to use it

Whenever you are the small player against a much larger, patient counterparty — a retail trader against the market, a startup against incumbents, a bettor against the house — assume the constraint is lasting long enough, not being right on average. Size and pace your bets so the near edge stays out of reach.

Why the ensemble mean misleads across fat tails

Analogy. Imagine a lottery whose average prize is enormous because one ticket in a billion pays a trillion dollars. The arithmetic mean is huge — but almost no one ever touches that one ticket, and to “collect” the average you’d have to still be solvent across a billion draws. You won’t be. Ruin removes you long before the outlier arrives.

The point. In a fat-tailed world — what we called Extremistan in the fat-tails course — the arithmetic (ensemble) mean is dominated by rare, unrepeatable outliers. A single monstrous value can carry the entire average. But an average you have to survive to collect is an average you can’t collect, because the very fatness of the tail that inflates the mean is also what wipes you out on the way there. The ensemble average counts the jackpot path; the time average — your path — almost never rides it.

Worked contrast. Return to the flagship coin (+50% heads / −40% tails on your whole balance). The ensemble mean is a cheerful +5% per round, propped up entirely by a shrinking sliver of paths that keep hitting heads. The typical, median path multiplies down at about −5% per round toward the barrier. The mean is real; it’s just describing a lucky ghost, not you.

Pitfall. In fat tails, “the average” and “what typically happens” are different creatures — and the gap grows with how ruinous the tail is. Reporting the mean as though it were the typical outcome is not a rounding error; it’s a category mistake that hides exactly the outcome (ruin) that ends your ability to average anything.

When to use it

Distrust any headline average drawn from a fat-tailed, compounding, or ruin-capable process. Ask for the median path and the survival probability instead. If the mean is being held up by a handful of outliers you must survive to reach, the mean is not your forecast.

The real-world faces of the same fact

Everything above is one idea in four costumes. Here is where the coin’s arithmetic becomes advice.

Position sizing: never bet the game-ender

Concrete illustration. Two traders each have a genuine +5% edge per trade. Alice risks 2% of her capital per trade; Bob, sure of his edge, risks 60%. A run of a few bad trades barely dents Alice — she keeps compounding for years. The same run takes Bob past a ×0.6 sequence he can’t recover from, and he’s out. Both had the same edge; only one survived to compound it. Over a career, the trader who’s still trading beats the trader who was briefly brilliant.

The rule. Position sizing means never staking so much that a plausible bad run ends the game. This is exactly what Kelly (lesson 04) formalises — bet a fraction proportional to your edge — and why practitioners bet fractional Kelly (half-Kelly or less): the growth cost of betting a little too small is mild, but the ruin cost of betting too big is permanent. Survival is the multiplier on every future return.

When to use it. On any repeated, compounding bet, size from the downside first: pick a stake small enough that the worst realistic streak leaves you still in the game, then let the edge compound. Never size from the upside.

Leverage: the ergodicity-destroyer

Concrete illustration. Borrow to double your exposure and you double every swing. The flagship coin becomes +100% / −80%. The ensemble mean rises — $0.5 \times 2.0 + 0.5 \times 0.2 = 1.10$, a shiny +10% — so the brochure looks better. But the time average falls: $\sqrt{2.0 \times 0.2} = \sqrt{0.4} \approx 0.632$, a brutal −37% per round. And a single −80% is a long stride toward the barrier. Leverage made the crowd’s number prettier and your number lethal.

The definition. Leverage (borrowing to amplify a position) sharpens multiplicative swings: it raises the ensemble expected value while lowering the time-average growth rate and raising the probability of ruin. In the language of this course, over-leverage is voluntary non-ergodicity — you deliberately widen the gap between the crowd’s average and your own fate, on the wrong side. The cautionary monument is Long-Term Capital Management (LTCM), a fund of Nobel laureates whose models had a real edge but whose enormous leverage meant one fat-tailed move in 1998 pushed them straight to the barrier. Edge without survival is just a slower way to blow up.

When to use it. Treat leverage as a survival decision, not a return decision. Before borrowing, recompute the time average and the ruin probability at the levered swings — not the ensemble EV, which will always tempt you upward. If leverage raises your chance of touching zero, it is destroying growth even as it flatters the average.

Insurance: buying back ergodicity

The reframe (Peters). Here is the most beautiful idea in the lesson. You pay an insurer a small, certain premium every year. Actuarially, insurance is a negative-expected-value bet for you — the insurer prices in their profit, so on the ensemble average you lose money. Textbook expected-value logic says: don’t buy it. Textbook expected-value logic is wrong, and ergodicity explains why.

Precise definition. Buying back ergodicity means paying a small certain cost to remove a rare, ruinous tail from your personal wealth process. Your house burning down is an absorbing-barrier event — it can zero you out and end your ability to compound. Insurance amputates that tail. In exchange for a guaranteed small drag, your wealth process can no longer hit the barrier, so it becomes ergodic-enough to keep compounding. You are not buying a positive-EV bet; you are buying the right to keep playing — and the time-average value of staying in the game can vastly exceed the ensemble-average cost of the premium. This is rational even at negative expected value.

The asymmetry (why it works for both sides). The insurer and you face different processes. The insurer pools millions of uncorrelated policies; no single house fire moves their fortune, so their book is effectively ergodic — their time average converges to the ensemble average, and they collect their positive edge like the casino in lesson 01. You are one house, one path. The same event that is a rounding error to the pool is ruin to you. The negative-EV premium you pay funds the insurer’s ergodic edge, and in return it makes your path ergodic-enough to survive. Both sides win because they’re playing different games.

When to use it. Insure against any loss that is small in probability but ruinous in consequence to your single path — and skip insurance on losses you can absorb and shrug off. The test isn’t “is this bet positive-EV?” It’s “does this event threaten to end my sequence?” If yes, pay to remove it even at a loss.

One-path life assets

Concrete illustration. A brilliant executive takes an aggressive, reputation-risking gamble because the “expected” payoff looks great across a hundred hypothetical careers. But there aren’t a hundred careers — there is one. The 5% chance of a disqualifying scandal isn’t a manageable cost averaged over parallel selves; it’s a possible absorbing barrier for the only self that exists.

The point. Your career, reputation, health, and relationships are non-ergodic in the deepest sense: you get one trajectory, and several of these have absorbing barriers of their own (a ruined reputation, a fatal illness, a burned relationship). Decline-and-recover isn’t always on the menu. So a gamble that looks fine “on average” can be catastrophic on the one path you actually walk — because the average is again computed over a crowd of parallel yous who don’t exist.

When to use it. For anything you have exactly one of and can’t rebuild, refuse ruin gambles regardless of their advertised average. Protect the base — health, name, key relationships — the way you’d protect a bankroll you can never re-deposit.

The through-line: survive first, optimise second

Strip away the costumes and every section says the same sentence. “On average it works out” is cold comfort to the one path that busted. The ensemble average lives in a world of parallel yous; you live on a single line through time with a trapdoor at zero. Because that trapdoor is absorbing, survival isn’t one goal among many — it’s the precondition for every other goal. A strategy has to survive before it is allowed to compound. Growth that risks ruin isn’t fast growth; it’s a lottery ticket wearing growth’s clothes.

That is why the order is fixed and non-negotiable: survive first, optimise second. Position sizing, avoiding over-leverage, buying ruin-removing insurance, guarding your one-path assets — these aren’t cautious add-ons to a return-maximising strategy. They are the strategy, because they’re what keep you on the board long enough for compounding to happen at all.

Vocabulary recap: match the core terms

Before the final check, lock in the five load-bearing terms of this whole course.

Pitfall and when to use all of this

The pitfall to retire for good. The master error, one final time: computing an expected value and treating it as your destiny, without checking whether a single step can end the game. Every real disaster in this lesson — the compounding 1-in-1,000 risk, gambler’s ruin, LTCM’s leverage, the executive’s scandal — is the same omission. Nobody wrote “we assume we never hit the barrier”; they just quoted the average and moved on.

The one question to ask, always. Before any repeated or compounding bet: “Can a single step end the game?” If no — the loss is bounded, non-ruinous, additive — go ahead and trust the ensemble average; it’s ergodic and expected value is your friend. If yes — a step can take you to an absorbing barrier — then price survival before you price the average. Size the position for the worst run, refuse the ruinous leverage, buy the insurance that removes the tail, and protect the one-path assets you can’t rebuild. The average can wait; you can’t afford to.