Tool/Decision Theory/No. 0087

Bayesian Updating

The fraud alert looks damning, until you learn how often it flags perfectly ordinary purchases.

Also called Bayesian Belief Updating

a tool: pick it up

Evidence: Well established
Read: 6 min
Links: 23 connections
Useful when: Deciding under uncertainty · Evaluating a claim · Forecasting · Reading data and statistics · Risk and safety

01You've seen this when…

in life
A friend is twenty minutes late for dinner. You suspect she forgot, then a transit alert reports delays on her route. Your suspicion eases.
at work
A software team treats a latency spike as a monitoring glitch. Customer complaints arrive from two regions, and the incident lead raises the probability of a real outage.
out in the world
An election forecast shifts after a new poll. It barely moves when ten news sites repeat that same poll.

02The idea

An alert should change your mind. How much depends on what you believed beforehand and how often that alert appears when nothing is wrong.

Bayesian updating makes that adjustment explicit. You begin with a probability before seeing the new evidence: the prior. Then you compare how likely the evidence would be if your claim were true with how likely it would be if it were false. That comparison tells you how far to move. The revised probability is the posterior.

The key distinction is between how likely the evidence is given the claim and how likely the claim is given the evidence. These are different conditional probabilities. A detector can catch most fraud and still flag more legitimate purchases than fraudulent ones.

For a yes-or-no claim, the calculation is compact:

Updated odds = prior odds × likelihood ratio.

Odds compare the probability of the claim with the probability of its alternative. A 20% probability gives odds of 20:80, or 1:4. The likelihood ratio compares the probability of the evidence under those two possibilities. If the evidence is three times as likely when the claim is true, multiply the odds by three.

Bayes’ theorem supplies the arithmetic. Bayesian updating is the practice of applying it as new evidence arrives. Today’s posterior can become tomorrow’s prior—but yesterday’s evidence must not get counted again.

03How to use it

Specify the claim. Choose something clear enough to check, such as whether a purchase is fraudulent. Identify plausible alternatives to compare with your favored explanation.
Start before this evidence. Estimate the probability using only information available before the observation you’re about to add. Look at comparable cases. Fraud rates among all purchases may differ from rates among purchases from a particular channel.
Compare both likelihoods. Estimate how often this evidence appears when the claim is true and when it is false. Evidence that fits your explanation but fits its alternatives equally well should barely move you.
Calculate the change. Convert probability to odds, multiply by the likelihood ratio, then convert back. For odds of a:b, the probability is a divided by a+b. Alternatively, count expected cases in a group of 100 or 10,000, as in the example below.
Check for recycled evidence. Multiply separate likelihood ratios only when the observations are independent given each hypothesis. Otherwise, ask how much the later observation adds after knowing the earlier one. Two reports drawn from the same witness count as dependent evidence.
Turn the estimate into a decision. Decisions depend on both probability and consequences. Compare the cost of acting with the cost of waiting, accounting for the cost of being wrong. A further check may be worthwhile when its value of information exceeds its cost.

If your inputs are rough estimates, calculate a range that reflects their uncertainty. Record the estimate and outcome. Over repeated judgments, check your calibration: do events you rate at 70% happen about seven times in ten?

04A worked example

Imagine an online store where audits estimate that 1% of purchases are fraudulent. Its screening system flags 80% of fraudulent purchases and 5% of legitimate ones. For this illustration, assume those rates apply to the purchase being checked.

What it looks like A flagged purchase seems almost certainly fraudulent. The system catches 80% of fraud, so a manager treats the alert as an 80% probability of fraud.

What’s actually going on Picture 10,000 comparable purchases. About 100 are fraudulent, and the system flags 80 of them. The other 9,900 are legitimate, and the system flags 495 of those. Of the 575 alerts altogether, 80 involve fraud. A flagged purchase therefore has about a 14% probability of being fraudulent, not 80%.

The odds calculation reaches the same result. The prior odds are 1:99. The alert is sixteen times as likely for fraud as for a legitimate purchase: 80% divided by 5%. Updated odds are 16:99, giving a probability of 16 divided by 115, or about 14%.

The alert raises the probability from 1% to 14% while leaving fraud uncertain. Ignoring it and treating it as a verdict are both mistakes.

What made it work Starting with the base rate and counting both false and true alerts. Those steps separate detection accuracy from the probability of fraud. The store now has a basis for comparing verification, approval and rejection using the estimated probability of fraud.

05When to reach for it

A result sounds conclusive. A positive test, fraud flag or suspicious behavior needs interpretation against the background rate and the false-alarm rate.
Evidence arrives in installments. Forecasts, investigations and troubleshooting benefit from keeping track of what each observation actually adds.
People disagree about the same evidence. Separate disagreements about starting probabilities from disagreements about what the evidence predicts. Those are different arguments.
You’re testing a favored explanation. Ask whether competing explanations predict the same observation. This can interrupt confirmation bias, where merely finding compatible evidence feels like proving your case.

06When it misleads

The alternatives are incomplete. You can correctly favor one of two explanations while both are wrong. Comparing equipment failure with a cyberattack can miss a mistaken configuration. Leave room for explanations you haven’t considered.
The rates come from the wrong population. A test’s performance in a research sample may differ from its performance among current users. Distribution shift can make yesterday’s carefully measured likelihoods unsuitable today.
The evidence is selected. Updating on the most striking result while ignoring how many results were searched can exaggerate its force. The relevant evidence includes both the observation and the process that produced it.
Precision disguises guesses. An exact calculation with poorly estimated inputs gives an exact answer to a poorly specified problem. Check plausible alternatives for the prior and likelihoods before trusting the final decimal.
You assign impossibility too casually. A prior probability of exactly zero stays zero under ordinary Bayesian updating. Reserve it for what your model rules out. Merely surprising events still need a probability above zero. When reality contradicts the model, revise it to reflect the evidence.

07Roots

After Thomas Bayes died in 1761, his friend Richard Price prepared an unpublished paper for the Royal Society. Bayes had been a minister in Tunbridge Wells, England, with a serious interest in mathematics. The usual probability exercise starts with a known chance and calculates outcomes. His problem ran backward, inferring an unknown chance from observed outcomes.

The paper used balls thrown onto a table to connect an unknown position with an unknown probability. Repeated throws supplied information about that probability. Published in 1763, the argument solved a particular version of the problem. The general-purpose updating recipe used today developed later.

Pierre-Simon Laplace independently developed a broader approach in 1774. He pursued the problem of inferring causes from observed events and later applied probability methods to astronomy and population questions. Bayes supplied an influential special case; Laplace helped turn inverse probability into a working method.

The approach spread through science, though its dependence on prior assumptions remained controversial. Much later, computers made complicated Bayesian models practical. The modern technique retains the original reversal: begin with possible explanations, work out what observations they predict, then use what actually happens to revise their probabilities.

08How solid is this?

The update rule follows from probability theory, assuming the model is appropriate. Experiments support using natural-frequency counts to improve Bayesian reasoning, but formal arithmetic does not guarantee sound priors, likelihoods or everyday judgments.

10Origin and sources

Thomas Bayes’s paper, published posthumously by Richard Price in 1763, addressed inference about an unknown probability. Pierre-Simon Laplace independently developed a broader treatment in 1774.

[1]Sivia, D. S., & Skilling, J. (2006). Data Analysis: A Bayesian Tutorial (2nd ed.). Oxford University Press.
[2]Gigerenzer, G., & Hoffrage, U. (1995). How to improve Bayesian reasoning without instruction: Frequency formats. Psychological Review, 102(4), 684–704.
[3]McGrayne, S. B. (2011). The Theory That Would Not Die: How Bayes' Rule Cracked the Enigma Code, Hunted Down Russian Submarines, and Emerged Triumphant from Two Centuries of Controversy. Yale University Press.

Suggest an edit· Updated 2026-10-02