It's Just Unknown

When you know the odds, more data is enough

Start with the tame case. A fair die has six faces and you know all of them; the long-run average is 3.5 and nothing can change that. This is what Frank Knight called risk in 1921: a situation where the odds are known, or at least knowable. Here, data is your friend. Roll a few times and your estimate jumps around. Roll a thousand and it settles, tightly, onto the truth.

Below, each faint line is one person rolling the same die over and over, plotting their running average as they go. Early on they sprawl all over. Push the slider right and watch the whole crowd funnel toward 3.5 — the spread shrinks like clockwork. This is the law of large numbers, and it is the entire reason planning works when you know the distribution.

This is the world that spreadsheets are built for: insurance tables, casino edges, factory defect rates. The dice are honest and you've seen them. Take enough samples and the average is not a guess — it's a measurement. The mistake is to assume every important decision lives here.

“Uncertainty must be taken in a sense radically distinct from the familiar notion of risk, from which it has never been properly separated.”

Frank Knight · Risk, Uncertainty and Profit, 1921

Knight's whole point was that most consequential choices are not dice you've seen. You don't know the faces, you don't know how many there are, and no amount of past data tells you. That's Game Two — and there, the comforting average turns into a trap.

When you don't know the distribution, the average is a rumour

Now swap the honest die for a black box. It pays out numbers, and you have no idea how it's built. You do what any sensible person does: you sample it, keep a running average, and watch it appear to settle. It looks like it's converging — just like the die did. So you write the number down and plan around it.

Then a single draw arrives that's larger than everything before it combined, and your tidy average lurches. Press play a few times: sometimes the giant lands early, sometimes it hasn't landed yet. Watch the bottom readout — the share of the whole total owned by the single biggest draw. In the die's world that share melts toward nothing. Here it stays stubbornly large. One observation keeps dominating all the others. That is the fingerprint of a fat tail, and it means your sample mean is not a measurement — it's a hostage to the largest thing you happen to have seen so far.

This is the engine room of bad forecasts. Markets, pandemics, wars, hit books, viral posts, the downside of a new technology — they are black boxes with fat tails, not dice you've measured. The number you confidently averaged from the calm years is exactly the number that betrays you in the loud one. Knight's heir, Nassim Taleb, hammers the same nail: in these domains the past is a deceptive sample, because the events that matter most are precisely the ones that haven't shown up in it.

“About these matters there is no scientific basis on which to form any calculable probability whatever. We simply do not know.”

John Maynard Keynes · 1937

So if the average can't be trusted and the odds can't be known, what's left to do? You stop trying to predict the box — and start deciding how much it's allowed to cost you when you're wrong. That's Game Three.

So size for survival, not for the average

Here is the move that separates people who last from people who don't. When you genuinely can't know the odds, you don't optimise for the expected outcome — because there's a hidden way to lose that you can't see, and you only get one life, not the average of a thousand. Cross the river that's on average four feet deep and the six-foot middle still drowns you.

Below are a thousand lives compounding wealth through a world with an unseen crash lurking in it. The optimiser bets big, sized as if the rosy estimate were real; the slider sets how aggressive. The survivor caps exposure and keeps a floor. Run it. The optimiser posts a gaudy median for whoever lives — and piles a clump of lives onto the ruin line at the bottom, wiped to nothing and unable to come back. The survivor gives up some upside and is simply never there.

Notice what the median hides. Averaged across a thousand parallel lives, aggression can look brilliant — but you don't get a thousand lives. You get this one, walking forward in time, and a single visit to the ruin line ends the story permanently. Once you're at zero, there is no comeback, no matter how good the next bet would have been. The survivor isn't timid; the survivor has simply noticed that staying in the game is the precondition for every gain that follows.

“Never cross a river if it is on average four feet deep.”

Nassim Nicholas Taleb