Why
Poker had to invent a word for judging a decision by its outcome, and the word is "resulting". It needed one because the correct play loses often enough that a player who learns from results learns the wrong lesson — folds a hand that was right to call because the last one lost, calls a hand that was wrong because the last one hit. Every field with noisy feedback has this problem; poker is unusual in having named it, priced it, and built its whole training culture around separating process from outcome. In a sample of one, a good bet that lost and a bad bet that lost are indistinguishable, and the only thing that separates them is a record of the reasoning made before the result was known.
The first transferable idea is that the threshold is arithmetic on costs rather than a level of confidence. Calling is correct when the probability of winning exceeds the amount to call divided by the pot plus that amount — a number the price hands you before you know anything about the hand. This catalogue keeps arriving at the same structure from other directions: lead-time-is-the-product finds the alert threshold already stated in the cost ratio on the motivating slide, and the-70-has-to-be-wrong argues that a threshold is a statement about costs and not a property of the forecast. Poker's contribution is that it writes the number down and everyone else leaves it implicit.
The second is that expected value is not the objective, and this is the correction most people need. A strategy that maximises expected value per decision, sized without regard to the bankroll, goes broke with probability one given enough time — because ruin is absorbing and no later edge can be collected from zero. Kelly turns a probability into a size, and the constraint that binds is not the size that maximises growth but the size that survives a bad run of the length the variance actually produces. Calibration gets you the right p, Kelly turns p into a stake, and risk of ruin says the stake is capped by survival rather than by edge. A trading agent with a well-calibrated model and no bankroll rule is a solved forecasting problem attached to an unsolved sizing problem.
The third is that a record means nothing until it is long enough, and how long is computable rather than a matter of opinion. Winning players lose over stretches long enough to destroy confidence, so "am I any good" is not answerable from a month of results. Given a claimed edge and a variance, the number of trials needed to distinguish that edge from zero at a stated confidence is arithmetic, and it is routinely far larger than the sample anyone has. Computing it first converts an argument into a schedule — the same move as the observation window in monad-last-general-purpose-l1, where the useful act was recording a baseline before it became unrecoverable.
And then the part that matters most, which is where the analogy stops. Poker is a game of risk: the distribution is known exactly — fifty-two cards, fixed rules — and only the realisation is hidden. Forecasting is a problem of uncertainty: the distribution itself is unknown and is the thing being estimated. So a poker player's calibration is arithmetic while a forecaster's is an empirical claim that has to be earned against reality, and poker intuitions systematically overrate how well calibrated one can be about the world. Borrow poker's decision hygiene and not its confidence. The second half of the disanalogy does transfer, though, and people forget it: poker is adversarial, so an edge decays as it is used and the opponent adapts. Markets do the same thing, which is why a backtest describes a world that no longer contains you once you trade in it.
How it works
Three ideas that transfer, and their form here
| Poker | The general form | Where this catalogue already has it |
|---|---|---|
| Pot odds — call if p > c / (pot + c) | The threshold is arithmetic on costs, available before you know anything | lead-time-is-the-product, the-70-has-to-be-wrong |
| Bankroll management | Expected value is not the objective; survival caps the stake | Kelly, mentioned but not built out, in the-70-has-to-be-wrong |
| Variance and sample size | A record means nothing until N is large enough, and N is computable | The observation window in monad-last-general-purpose-l1 |
| Resulting | Judging a decision by its outcome, which noisy feedback punishes | — this card |
The threshold, written out
| Pot | To call | Break-even p | Read as |
|---|---|---|---|
| 100 | 50 | 33% | The price is demanding a one-in-three |
| 100 | 20 | 17% | Cheap; a weak read still calls |
| 100 | 200 | 67% | Expensive; needs a strong read |
The number comes from the price, not from the hand. A decision that skips it is choosing a threshold anyway — just not on purpose, which is the same failure the-70-has-to-be-wrong finds in scoring a probability with accuracy.
Why expected value is the wrong objective
| Objective | What it maximises | How it fails |
|---|---|---|
| EV per decision | Average outcome, ignoring path | Ruin is absorbing. With enough time, probability of ruin → 1 |
| Growth rate (Kelly) | Long-run compounding | Assumes the edge estimate is right; overbetting is punished asymmetrically |
| Survival first | Time in the game | Leaves growth on the table, and that is the price of staying |
A well-calibrated model with no bankroll rule is a solved forecasting problem bolted to an unsolved sizing problem.
Where the analogy stops, which is the useful half
| Poker | Forecasting the world | |
|---|---|---|
| The distribution | Known exactly — 52 cards | Unknown, and is what is being estimated |
| Kind of unknown | Risk | Uncertainty |
| Calibration is | Arithmetic | An empirical claim to be earned |
| Adversarial | Yes — the opponent adapts | Markets too, which is why a backtest describes a world that no longer contains you |
The first three rows are why poker intuitions overrate achievable calibration: borrow the decision hygiene, not the confidence. The last row is the one that does transfer and gets forgotten.