Suppose you think rain is unlikely, then notice dark clouds. How much should that observation change your mind? Bayes’ theorem updates the probability of each possible hypothesis using the data you observed. section 2.1
For a hypothesis H and observed data D, the rule is Pr(H | D) = Pr(D | H) × Pr(H) / Pr(D). The prior, Pr(H), is your starting probability. The likelihood, Pr(D | H), says how probable the data would be if H were true. The posterior, Pr(H | D), is your updated probability. The evidence, Pr(D), makes the posterior probabilities sum to 1. section 2.2
To update a set of competing hypotheses:
- Assign each hypothesis a prior probability. section 2.2
- Find the likelihood of the observed data under each hypothesis. section 2.2
- Multiply each prior by its likelihood, then divide each product by the sum of all the products. section 2.2
Why a positive test need not mean disease
Suppose a disease affects about 1 in 150 people. A test detects it 98% of the time but returns a false positive for 5% of people without it. For a person with no related prior health conditions, a positive result raises the probability of disease to about 12%, not 98%. The disease’s low starting probability still matters; overlooking it is the base-rate fallacy. section 2.5
Updating beverage preferences
To estimate the proportion of people who prefer tea, consider eleven possible proportions from 0 to 1. Start with equal prior probabilities, then use a survey in which three of five people prefer tea to calculate each likelihood. Multiplying and normalizing produces a posterior that peaks near 0.6. NumPy arrays let the same calculation handle all eleven hypotheses at once. section 2.11
unnormalized_posterior = prior * likelihood
posterior = (
unnormalized_posterior
/ unnormalized_posterior.sum()
)
These two lines turn the tea survey’s prior and likelihood arrays into posterior probabilities: multiply matching entries, then divide by their total. section 2.11
The starting choices matter. A plausible hypothesis given zero prior probability cannot gain posterior probability, even when data favor it; zero likelihood also rules it out. Include the possibilities you care about rather than excluding one prematurely. With a lot of data, many different priors yield similar posteriors. section 2.15 section 2.14


