Prior Probability

Prior probability is your starting belief about something before you look at new data. It’s the “what you’d reasonably expect” part of a decision, based on background knowledge, history, or domain expertise.

What it means technically
In Bayesian statistics, a prior probability (often just called a prior) is a probability distribution that represents uncertainty about a quantity before observing the current dataset. After you see data, Bayes’ theorem combines the prior with the likelihood (how compatible the data are with different possibilities) to produce the posterior probability—your updated belief.

Quick intuition
Think of the prior as a “default setting” for your beliefs. Data can move you away from that default, but how far you move depends on how strong and informative the data are. With lots of high-quality data, the posterior is driven mostly by the data; with sparse or noisy data, the prior can matter a lot.

Practical examples
Priors show up whenever you have sensible expectations before measuring:

Medical testing: Prior probability of a disease is the base rate in the population (or for a risk group). A positive test result means something very different when the prior is 1% versus 20%.
House prices: Before seeing a specific home’s features, you might expect prices in a neighborhood to cluster around a typical range; that expectation can be encoded as a prior over the average price.
Spam detection: The prior probability that an email is spam can reflect historical spam rates for a user or organization.

Why it matters in AI/ML
Priors help stabilize learning, especially with limited data, by regularizing models and preventing extreme estimates. Many common ML ideas have Bayesian interpretations: L2 regularization corresponds to a Gaussian prior on weights, and L1 regularization to a Laplace prior. Priors are central in Bayesian neural networks, probabilistic programming (Stan, PyMC), and calibration tasks where base rates matter.

Prior Probability is the probability assigned to a hypothesis, parameter value, or class label before observing the current data, reflecting existing knowledge or assumptions. In Bayesian inference it combines with the likelihood via Bayes’ theorem to produce a posterior probability, influencing estimates especially with limited or noisy data. Example: in spam detection, a prior might encode that only 5% of emails are spam, shaping early predictions.

Imagine you’re guessing what’s inside a wrapped gift before you shake it. If most gifts at this party are books, you’d naturally think “probably a book” even before getting any clues. That starting guess is like a prior probability.

In Bayesian statistics (often used in AI), a prior probability is your initial belief about how likely something is before you look at new data. Then, when you observe evidence (like shaking the gift or seeing part of the label), you update that belief to get a better, more informed probability.