A director walks out of one good demo convinced a struggling project is fixed. A board hears one bad quarter and writes off a strategy that took three years to build. Both made the same mistake, in opposite directions: they let the latest piece of evidence overwrite everything they knew before it. Bayesian reasoning is the discipline of doing neither, weighing new evidence against your existing beliefs, in proportion, and landing somewhere sensible in between.

The quick version

  • A prior is your honest belief before new evidence, usually best expressed as a probability, not a yes/no.
  • When evidence arrives, you update to a posterior: a blend of the prior and how strongly the evidence points one way.
  • Weak or unreliable evidence should move you a little; strong, surprising evidence should move you a lot. Most people get this backwards under pressure.
  • The leadership skill isn't doing the algebra, it's stating a number you'd actually bet on, then changing it when the world tells you something new.

The idea in depth: where the rule comes from

The maths is older than most countries. The Reverend Thomas Bayes, an English nonconformist minister, worked out a way to reason backwards from observed evidence to the probability of its cause. He never published it. After his death, his friend Richard Price found the paper, revised it, and read it to the Royal Society; it appeared in the Philosophical Transactions in 1763 as "An Essay towards solving a Problem in the Doctrine of Chances." Two and a half centuries later it underwrites spam filters, weather forecasts, clinical diagnosis and a good chunk of machine learning.

You can skip the formula and keep the logic. Bayes' theorem says your belief after seeing evidence (the posterior) depends on two things: your belief before (the prior), and how much more likely that evidence is in a world where you're right than in a world where you're wrong (the likelihood). Strong evidence is evidence you'd almost never see if your hypothesis were false. Weak evidence is the kind you'd see either way, it shouldn't move you much, however vivid it feels.

So the move is: before you look at any new data, say out loud what you already believe, as a probability. "I think this hire works out, call it 70%." Naming the prior does two things. It stops you treating the newest fact as the only fact, and it gives you something concrete to update instead of a vague mood.

flowchart LR
  A(["Prior belief
(what you think now)"]) --> C(["Update with
Bayes' rule"])
  B(["New evidence
× how reliable it is"]) --> C
  C --> D(["Posterior belief
(what you think next)"])
  D -. "becomes the prior for
the next piece of evidence" .-> A
The loop: today's posterior is tomorrow's prior. Belief is something you revise, not something you hold. Leaders Loop

Why people get it wrong: base rates

The most expensive Bayesian error is ignoring the prior altogether, what psychologists call base-rate neglect. Daniel Kahneman and Amos Tversky documented it across a run of studies in the 1970s, beginning with "On the Psychology of Prediction" (Psychological Review, 1973), which showed that people judge probability by how closely a case resembles a stereotype and quietly discard how common the category actually is. The sharpest demonstration is the taxicab problem. A cab is in a hit-and-run at night. 85% of the city's cabs are Green, 15% Blue. A witness says the cab was Blue, and tests show the witness is right about colour 80% of the time. How likely is it the cab really was Blue?

Most people answer around 80%, they anchor on the witness. The Bayesian answer is roughly 41%: because Blue cabs are so rare to begin with, the witness's 80% reliability isn't enough to make "Blue" more likely than not. The vivid testimony fights the base rate, and loses. The practical habit follows: when a striking piece of evidence lands, ask "how common is this thing in the first place?" before you act on it. A glowing reference, a single user complaint, one stellar interview, each has to be read against a base rate, or you'll over-react to noise.

Strong evidence is evidence you'd almost never see if you were wrong. Everything else should move you only a little.

The idea in depth: updating like a forecaster

Knowing the rule isn't the same as having the temperament. The largest study of real-world judgment we have is Philip Tetlock's Good Judgment Project, written up with Dan Gardner in Superforecasting (2015). Across roughly a million forecasts on world events, the people who consistently beat the field, and beat intelligence analysts with access to classified material, shared a signature habit: they updated often, and in small steps. They nudged a forecast from 0.6 to 0.65, not from "yes" to "no." Their accuracy, measured by Brier score (lower is better), averaged around 0.166 against roughly 0.259 for ordinary forecasters.

That is Bayesian reasoning as a personality trait rather than a calculation. A superforecaster treats each new fact as a small force acting on a belief, not a verdict that replaces it. So: resist the urge to either dismiss new evidence or capitulate to it. When a project update arrives, don't ask "good news or bad news?" Ask "by how much should this shift my estimate?", and let the answer usually be "a little." This is the same incremental, evidence-first habit behind decision theory and expected value: hold beliefs as probabilities so you can compute with them, not as flags you plant and defend.

flowchart TD
  E(["New evidence arrives"]) --> Q(["Would I see this evidence
if my belief were WRONG?"])
  Q -->|"Yes, easily, weak signal"| S(["Nudge belief a little"])
  Q -->|"Almost never, strong signal"| L(["Move belief a lot"])
  S --> R(["State your new probability"])
  L --> R
One question separates a strong signal from noise: how surprised would you be to see this evidence in a world where you're wrong? Leaders Loop

The honest limitation: garbage priors, garbage posteriors

Bayes' rule is only as good as the prior you feed it, and priors are where bias hides. If you start from a badly miscalibrated belief, overconfidence, a stereotype, last year's strategy mistaken for a law of nature, the maths will faithfully carry that error forward. Two people with different priors can see the same evidence and update to different conclusions, both perfectly "rationally." That's not a flaw to paper over; it's the reason calibration matters. The honest version of this skill includes interrogating where your prior came from, writing your estimate down before the result so you can't quietly rewrite history, and seeking out the evidence that would move you against your current view rather than the evidence that confirms it. Bayes tells you how to update. It does not tell you that your starting point was any good.

A worked example

Your team ships a redesigned checkout flow. Your prior, from past releases, is that a redesign like this improves conversion about 40% of the time, most ship neutral or slightly worse (illustrative figure, drawn from a typical release history). Day one, conversion is up 6%. The room wants to declare victory.

Run it Bayesian. The prior says "probably not an improvement", only 40% of these work. Now weigh the evidence: a single day of data is exactly the kind of swing you'd expect from normal traffic noise whether or not the redesign helped. So it's weak evidence, it should nudge your belief up, maybe from 40% to 50%, not slam it to "it works." A week later the lift holds at 5% across far more sessions and through a weekend. That's evidence you would not expect if the redesign were neutral, strong evidence, and it's right to move you hard, say to 80%. Same direction of news, two very different updates, because the strength of the evidence differs. The leader who shipped a celebration email on day one has to walk it back; the Bayesian leader bought a week of patience for the price of one sentence: "Good early sign, let's see if it survives more data."

Frequently asked questions

Do I actually have to use the formula?

Almost never, in a meeting. The value is in the habit, not the arithmetic: state a probability for your belief, then ask how reliable the new evidence really is before you let it move you. When a decision is big and the numbers are knowable, by all means write out prior × likelihood, but most days the discipline is verbal.

Isn't "changing your mind" just indecision?

The opposite. Indecision is having no position; flip-flopping is overwriting your whole position with the latest input. Bayesian updating is holding a clear, stated probability and adjusting it in proportion to evidence. It's the most disciplined form of changing your mind there is, and it lets you explain exactly why you moved.

Where does my prior come from if I have no data?

From base rates and honest experience. How often do projects like this hit their date? How often does a candidate this impressive turn out to be a great hire? Even a rough prior beats none, because it forces you to weigh the new evidence against something rather than treating it as the whole truth. Just flag it as rough, and update faster as real data lands.

How is this different from gut feel?

Gut feel is a prior you won't say out loud. Making it explicit, "I'm at 70%", is what lets you check it, update it, and be held to it later. The intuition is welcome; the refusal to name it is the problem.

Can this be gamed by someone with a stubborn prior?

Yes, and that's the real limitation. If someone sets their prior at 99% and refuses to budge, the maths can't save you. The fix is social, not mathematical: ask people to write estimates down beforehand, and to name in advance what evidence would change their mind. A prior nobody can falsify isn't a belief, it's a position.

Related in the Toolkit

Where to go next