Is a prior probability always a historical frequency?

No. It can incorporate historical evidence, contextual information or a stated modeling assumption.

Can a posterior become a prior?

Yes. Sequential updating is possible when new evidence is incorporated with the correct conditional model.

Does a 50% prior mean complete ignorance?

Not necessarily. It assigns equal probability to a particular binary partition, which may or may not be justified by the information and framing.

AI Glossary / Statistics and probability / Beginner

Prior probability

Definition

Prior probability is the probability assigned to an event or hypothesis before incorporating a specified new piece of evidence. It represents the starting information for a Bayesian update. For example, a prior probability of a fault can summarize what is known before a new sensor result arrives. Prior does not mean uninformed: earlier measurements, background rates and relevant contextual information may already be included. The word only identifies the probability's position relative to the evidence currently being considered.

What it means

The same clue can lead to different conclusions when it occurs in different starting contexts. An unusual signal among systems with frequent faults may be interpreted differently from the same signal among systems where faults are rare. Prior probability makes that starting context explicit. It should refer to a clearly defined event and relevant population or model. An arbitrary historical rate from an unrelated setting is not automatically a suitable prior for a new situation merely because it is numerical.

How it works

Define the possible hypotheses and the information available before the current observation. Assign probabilities that are coherent: for an exhaustive set of mutually exclusive hypotheses they must sum to one. Combine each prior with the likelihood of the new evidence under that hypothesis and normalize to obtain posterior probabilities. For continuous unknown quantities, use a prior distribution and derive probabilities for the events of interest. Document the source and relevance of the prior, assess sensitivity to plausible alternatives and avoid counting the same information both in the prior and as supposedly new evidence.

Three practical examples

Equipment diagnosis: start with fault probabilities appropriate to the equipment class and operating history before evaluating a new alert.
Quality review: use relevant background information about a production process before incorporating the latest inspection results.
AI classification: represent class frequencies or other prior information in a probabilistic model, while checking whether the training population matches the population in which predictions will be used.

Think of it this way

Think of setting the initial weights on several possible explanations before receiving the next clue. A strong clue can substantially rearrange those weights, but it does not erase the need to know where they started. The weights should reflect the available context rather than simply treating every imaginable explanation as equally plausible.

Worked example

Consider an original artificial diagnosis with three mutually exclusive and exhaustive explanations A, B and C. Their prior probabilities are 0.50, 0.30 and 0.20. Suppose a particular observed signal has likelihoods 0.20, 0.60 and 0.10 under those explanations. Multiplying gives unnormalized posterior weights 0.10, 0.18 and 0.02, which sum to 0.30. The updated probabilities are therefore 1/3, 0.60 and 1/15, approximately 0.3333, 0.6000 and 0.0667. Explanation A began with the largest prior, but B becomes most probable after the signal. The prior probability of B was 0.30, while its prior odds against all other explanations were 0.30/0.70 = 3/7. This calculation distinguishes probabilities from odds and shows that a prior is a starting point, not a requirement that the initial leading explanation must remain the final one.

Illustrative case study

In a fictional case, Cedar Monitoring applies a diagnostic model developed on frequently serviced equipment to a newer fleet. Analysts notice that the baseline fault frequency differs. They review both the starting probabilities and the signal behavior rather than copying the original interpretation unchanged. The revised documentation explains which history enters the prior and which new measurements enter the likelihood. Sensitivity analysis shows how uncertain background rates influence the final probabilities used for inspection planning.

Common misconceptions and mistakes

Prior probability is not the probability of the observed evidence under a hypothesis; that is a likelihood-related quantity. It is not necessarily a personal guess and can be informed by earlier data. Equal priors are a substantive assumption, not automatically the absence of assumptions. Assigning exactly zero prior probability to a hypothesis prevents ordinary Bayesian updating from giving it positive posterior probability when the model is unchanged and likelihoods are finite. Finally, a posterior from one stage can become a prior at the next stage, but only with careful treatment of dependence and information reuse.

Business use

Businesses should make prior assumptions visible when probabilities affect decisions. Identify the reference population, observation period and any contextual adjustments. A rate estimated from historical data may itself be uncertain, which can be represented in a larger model rather than treated as perfectly known. Compare results under reasonable alternatives when the prior is weakly supported. Avoid using the word prior to disguise an unexplained judgment or to imply that a calculation is automatically objective. Clear documentation allows decision makers to separate evidence-driven conclusions from assumptions that deserve further investigation.

Check your understanding

Three exhaustive, mutually exclusive hypotheses are assigned prior probabilities 0.5, 0.4 and 0.3. Is that coherent?

Answer: No. They sum to 1.2 rather than one, so the probabilities must be revised before a valid update.

Related terms

Learn next

Learn how probabilities differ from odds and likelihoods, then practice updates that explicitly distinguish old information from genuinely new evidence.

Sources

Frank Harrell, Glossary of Statistical Terms, prior probability and posterior probability: https://hbiostat.org/glossary/ Stan User's Guide, Posterior and Prior Predictive Checks: https://mc-stan.org/docs/2_36/stan-users-guide/posterior-predictive-checks.html

Previous: Prior-independent mechanism Next: Prior (regularizer)