Can a prior use past data?

Yes, provided the update handles the evidence coherently.

Is a prior necessarily subjective opinion?

It can incorporate substantive evidence and formal modeling choices; its justification should be explicit.

Can evidence overcome a small prior?

Sometimes, if the likelihood ratio is sufficiently strong.

AI Glossary / Statistics and probability / Intermediate

Hypothesis prior

Definition

A hypothesis prior is the probability assigned to a candidate hypothesis before conditioning on the particular evidence being analyzed. In Bayesian learning, it expresses initial relative plausibility over candidate models, explanations, or prediction rules. Combined with the likelihood of the evidence under each candidate, it determines posterior probabilities through Bayes' rule.

What it means

Evidence does not begin from a blank slate. A rare explanation and a common explanation may both predict an observed pattern, but their starting probabilities affect how strongly the pattern changes the comparison.
Prior means before the evidence used in this update, not necessarily before all experience. Earlier observations, domain knowledge, and explicit modeling assumptions can inform it. A previous posterior can become a new prior when genuinely new evidence arrives without double-counting the earlier evidence.

How it works

Define the candidate hypotheses clearly. For a finite set of mutually exclusive and collectively exhaustive alternatives, assign nonnegative prior probabilities that sum to one. If the alternatives omit important possibilities, posterior comparisons remain conditional on that restricted set.
Specify how each hypothesis generates the evidence. Its likelihood is P(D | H), whereas its prior is P(H). These are different quantities: a hypothesis can predict the observed evidence well while initially being considered unlikely.
Bayes' rule gives P(H | D) = P(D | H)*P(H)/P(D), where the denominator sums the weighted likelihoods across the alternatives. This normalizes posterior probabilities. In odds form, posterior odds equal prior odds multiplied by the Bayes factor comparing the alternatives.
For a model with unknown internal parameters, its model-level likelihood is generally a marginal likelihood that integrates those parameters against their prior distribution. Substituting the best-fitting likelihood is not the same calculation and can favor flexible models without accounting for how broadly their prior predictions are spread.
A prior can express preferences for simpler hypotheses or regularize a learning problem, but simplicity must be defined. Assigning equal probabilities to listed candidates is not inherently assumption-free: changing the candidate list can change the total mass allocated to a broader class.
For continuous parameterized hypotheses, use a probability distribution or density rather than assigning a positive mass to every individual parameter value. Under an ordinary continuous prior, a single exact value has probability zero even though neighborhoods around it have positive probability. A point-null hypothesis requires explicit point mass if it is to receive positive posterior probability.
Examine how the prior affects predictions and conclusions. Sensitivity analysis compares defensible alternatives, while prior predictive checks reveal outcomes implied before the new data are used. Strong evidence may reduce prior influence, but it need not do so when models are weakly identified or evidence is limited.
Avoid using the same data first to construct a concentrated prior and then again as if they were new independent evidence. Empirical-Bayes approaches can use data to estimate prior settings, but their procedure and uncertainty treatment should be identified explicitly.
A posterior probability is not a universal truth score. It depends on the hypothesis set, prior, likelihood, and evidence-selection process. Good probability calculations cannot rescue an incomplete set of explanations or an implausible data model.

Three practical examples

An AI diagnostic system combines a prior probability of a logging fault with the likelihood of an observed error pattern.
A forecasting team assigns prior probabilities to competing demand models and updates them with new observations.
A Bayesian learner places a distribution over candidate prediction functions, then combines that prior with training-data likelihoods.

Think of it this way

Think of evidence as changing the balance between two explanations rather than placing each on an empty scale. The prior gives the starting balance. The likelihood tells you how strongly the new observation pushes toward either side, and the posterior records the updated balance.

Worked example

Suppose two exclusive explanations are H-A and H-B. Assign priors P(H-A) = 0.9 and P(H-B) = 0.1. An observed event E has probabilities P(E | H-A) = 0.2 and P(E | H-B) = 0.8.
The weighted likelihoods are 0.9*0.2 = 0.18 and 0.1*0.8 = 0.08. Therefore P(E) = 0.26, and the posterior probabilities are P(H-A | E) = 0.18/0.26 = 9/13, approximately 0.692308, and P(H-B | E) = 4/13, approximately 0.307692.
The evidence favors B relative to A by a factor of four, but A remains more probable because its prior odds were nine to one. Posterior odds for B versus A are (1/9)*4 = 4/9.
With equal prior probabilities, the same likelihoods would instead give B posterior probability 0.8. This difference illustrates why the prior must be reported. Neither prior assignment is justified merely by producing the conclusion someone prefers.

Illustrative case study

In this fictional case, Spruce Operations evaluates whether an unusual AI-service log pattern comes from routine variability or a logging fault. Its analysts use documented historical information for the starting probabilities and update them with a newly observed pattern. They show how the result changes under alternative plausible priors and retain an unmodeled-cause review when neither explanation predicts the data well. The calculation supports investigation rather than certifying a cause.

Common misconceptions and mistakes

A prior is not the likelihood of the observed evidence. A strong likelihood ratio does not necessarily make a rare hypothesis the most probable. Equal priors are a modeling choice rather than a universal default dictated by ignorance.
A zero prior probability for a discrete hypothesis prevents ordinary Bayesian updating from giving it positive posterior probability. This differs from zero probability at individual points under a continuous density. A favorable posterior among poor candidates does not establish that the winning candidate is adequate.

Business use

Hypothesis priors make starting assumptions reviewable in forecasting, anomaly investigation, and model comparison. Record where they came from, which alternatives are covered, and whether sensitivity changes the decision. Distinguish selecting the most probable model from averaging predictions across models; the latter can retain uncertainty that a single-model decision discards.

Check your understanding

Why does H-A remain more probable in the example?

Answer: Its nine-to-one prior advantage exceeds the four-to-one evidential advantage for B.

What changes if an important third explanation is added?

Answer: The normalization and posterior probabilities can change, so the original result was conditional on its two-candidate set.

Related terms

Prior probability
Posterior probability
Bayes factor
Bayes' rule
marginal likelihood
Bayesian model averaging

Learn next

Learn probability versus likelihood and Bayesian odds updating. Then study parameter priors, model comparison, and sensitivity checks for incomplete or weakly supported hypothesis sets.

Sources

Michael Rescorla, Bayesianism, Open Encyclopedia of Cognitive Science: https://oecs.mit.edu/pub/98iya9su/release/1/ Stan User's Guide, Posterior and Prior Predictive Checks: https://mc-stan.org/docs/stan-users-guide/posterior-predictive-checks.html

Previous: Hypothesis class approximation error Next: Hypothesis space