Numeracy · 0 connections

Probability

mathematical probability

Consistent reasoning under uncertainty takes exactly one form: probability. When we are not sure what will happen, using clear numbers for probability is the best and fairest way to show our guesses. It helps everyone understand the same thing.

Explanation

Imagine you have a bag with 10 marbles: 9 blue and only 1 red. You reach in and pick one without looking. Even if you really hope for the red marble, the chance of picking it is still small because there is only one red marble out of ten. Probability helps us remember to look at the real numbers instead of just hoping. We start with what we already know about the bag. If we learn something new, we can change our number. Three simple rules keep everything working: Probability numbers are never less than zero. If something is sure to happen, the number is 1. If two things cannot happen at the same time, we add their numbers together. We should not say the chance is exactly zero or exactly 1 unless it is truly impossible or completely sure. That way we can still change our mind if we learn something new.

Watch

No video has yet earned its place here.

When it applies

When you are about to forecast something uncertain and a score will eventually exist — the outcome will be known — probability applies. Not because certainty is possible, but because a number anchored to evidence performs better than a verbal hedge when you can check. It applies when you are making the same judgment repeatedly: whether a candidate will succeed, whether a project will finish on time, whether a customer will return. Across enough cases, patterns are trackable. A consistent model beats unguided expert intuition more reliably than most experts expect. And it applies when you are pooling risk across many cases and the cost of mispricing it will land on you. Insurance, lending, and clinical triage are all exercises in turning probability into prices that hold up.

Where it stops

Probability does not promise how a short run will go. The law that averages converge toward the true probability is a statement about limits. It says nothing about the next ten flips. A fair coin with five heads in a row is no more likely to land tails on the sixth; the framework has no memory. "Fair coin" is itself a model. A coin flipped with consistent technique lands the same way it started roughly 51% of the time. Fairness is a useful idealization, not a measured property. Where the distribution is unknowable — rare financial events, extreme disasters, most geopolitical outcomes — assigning a precise probability is not careful reasoning. It is false precision. Some domains simply don't supply enough data, or have outcomes spread too far from the center, for a number to reliably describe the rare event. Three live disputes among people who understood probability fully: whether measured overconfidence is a real cognitive bias or partly an artifact of how the questions are framed; whether it is meaningful to assign a probability at all to a question that will only ever be tested once; and whether calibrated forecasting tournaments measure what matters or mostly reward the frequent and inconsequential while missing the tail event that determines everything. These tensions are unresolved. Hold them rather than adjudicate.

The misuse

An intelligence analyst at the CIA kept running into the same problem. An estimate would cross his desk describing some outcome as a "serious possibility." He asked several colleagues, separately, to put a number on those words. Answers ranged from roughly 20% to 80%. The same phrase, read by trained estimators, meant something different in every head. The estimate read as precise. It contained none. This is the most visible form: a verbal hedge doing the job of a probability. The word does none of the work — no base rate, no update, no score — but reasoning proceeds downstream as if a number had been fixed. The gap is invisible until the decisions start diverging. The deeper error is treating an assigned number as a property of the event itself rather than a statement about your own information. A probability you believe belongs to the event stops updating when you learn something new. It gets defended instead. What looks like rigor is a guess that has been promoted to a fact. The third form is grading the call by its outcome. Probability says a choice is scored against the distribution it was made against, not against the single result that happened. Revising your process based on one draw reverses the logic while keeping the vocabulary.

A worked example

There are many more bus drivers than librarians in most towns. You hear about a quiet person who loves reading books and spending time in quiet places. Many people quickly guess that this person is a librarian because the description sounds like what they imagine a librarian to be. But because there are far more bus drivers than librarians, it is actually more likely that the person drives a bus. The description makes people forget to look at the real numbers first. The best step is always to start with how many of each kind there really are.

Push

Replace the verbal hedge with a number. Find the base rate — across all cases like this one, how often does the outcome you are anticipating actually happen? Start there. State the number. Write it down, because an unrecorded estimate cannot be improved.

Veto

Do not hold a set of probability estimates that contradict each other — where the parts add up to more than one, or where you judge a joint event more likely than either event alone. A set of estimates like this can be turned into a combination of bets you are guaranteed to lose, whatever happens. This is not a risk to manage; it is a proved theorem. The number does not have to be wrong to be dangerous. It only has to contradict itself.

Connects to

Connections appear as the deck grows.

Go deeper

Mastery question

**Question:** You said an event was 30% likely. It happened. A friend says, "You were wrong." Were you? And if not, what would it take to show that your 30% was actually a bad number? **The answer:** A single outcome cannot show that one probability was wrong. 30% does not predict "won't happen"; it predicts that across many calls like this one, the thing happens roughly three times in ten — and this was one of those three. The number is a statement about your information, not a forecast of this one draw, so the realized result neither confirms nor refutes it. The only way to show the 30% was bad is to gather many of your "30%" calls and check whether the rate they come true lands near 30%. The number is graded against a distribution and a track record, never against one sampled outcome. **The answer that misses it:** "It happened, so I was wrong — I should have said something higher" (or, defending it: "No, low-probability things happen sometimes, so I was still basically right"). Both treat the single outcome as the thing that scores the number — one concedes to it, one explains it away. Both are arguing on the friend's terms. **Why the difference matters:** A person who has memorised probability can recite that "30% doesn't mean it won't happen" — but the moment a real outcome lands, they slide into grading the call by that outcome. Someone who has internalised it refuses the frame: the number lives in their information state, the realized draw is the one variable the framework says is uncontrollable, and the only valid grader is the rate across many calls. Both the surrender and the defence sound like understanding while quietly conceding the central point.