Bayesian Updating
Start with what is reasonably plausible, examine how strongly new evidence fits competing explanations, then revise your confidence instead of starting from zero or refusing to move.
Update your confidence as new evidence arrives.
How much should this evidence change my view?
Forecasts, diagnoses, strategy, investigations, hiring and uncertain decisions.
Fast overview
Bayes’ theorem connects what was plausible before the evidence, how expected the evidence would be under a hypothesis and what becomes plausible after seeing it. In formal Bayesian inference, the result is a posterior probability.
What was plausible before?
How diagnostic is it?
What should you believe now?
As a mental model, exact numbers are not always necessary. Make the starting assumptions visible, judge whether evidence truly favors one explanation and update by an appropriate amount.
Mental Model Compass
Learn it progressively
1. Set a prior
Estimate what was already plausible from relevant background information or base rates.
2. Name competing explanations
Ask what else could reasonably produce the same observation.
3. Judge the evidence
Evidence matters more when it is much more expected under one explanation than another.
4. Update proportionately
Strong diagnostic evidence should move you more. Ambiguous evidence should move you less.
5. Repeat
Your updated view becomes the starting point for the next meaningful piece of evidence.
See it in practice
Hiring
A polished interview is evidence, but a relevant work sample may better distinguish strong candidates from skilled interviewees.
Business
One strong sales month should raise confidence in a new offer, but the size of the update depends on sample quality and other explanations.
Learning
One low quiz score may reflect weak understanding or temporary conditions. Repeated evidence across tasks supports a larger update.
Worked application
Is a new customer segment genuinely promising?
Prior: Similar segments have converted poorly, so initial confidence is modest.
New evidence: A small pilot converts unusually well.
Alternatives: The segment is better, the sample was unusually favorable or the pilot conditions will not scale.
Update: Raise confidence, but not to certainty. Run a larger pilot under normal conditions. If performance holds, update again.
Numerical intuition
Suppose a condition affects 1 in 100 people. A test detects 99 percent of affected people, but also returns positive for 5 percent of unaffected people. A positive result is important evidence, but it does not imply a 99 percent chance of having the condition. The low starting rate matters.
This is a teaching example only. Real medical decisions require validated clinical interpretation.
Common mistakes
Ignoring the base rate
A rare event can remain relatively unlikely after evidence that sounds impressive.
Reversing conditional probabilities
P(evidence | hypothesis) and P(hypothesis | evidence) are not interchangeable.
Updating too much
Anecdotes, small samples and ambiguous clues usually deserve modest weight.
Refusing to update
A prior is a starting point, not a permanent verdict.
When not to use it
Do not force numerical probabilities when the numbers would be invented. A qualitative Bayesian check can still help. In high-stakes settings, use the model to organize questions, not replace validated evidence or domain expertise.
Related mental models
Circle of Competence
Judge whether you can assess the evidence yourself or need deeper expertise.
Inversion
Search for evidence that could disconfirm your preferred explanation.
Second-Order Thinking
Extend an updated belief into likely downstream consequences.