Every forecast starts before the data arrives. When our model says a team has a 49% chance to win tonight, part of that number existed yesterday — a starting belief about how strong each side is. In Bayesian forecasting that starting belief has a name: the prior.
## What a prior actually is
A prior is your estimate of how likely something is BEFORE you look at the newest evidence. Not a guess pulled from the air — a structured summary of everything you knew up to this moment: long-run results, ratings, base rates, the boring history that doesn't change overnight.
The word comes from "prior probability" — literally, the probability prior to the new data. Its counterpart is the posterior: the updated probability AFTER the evidence is folded in. Bayesian forecasting is nothing more mysterious than a disciplined loop: prior in, evidence weighed, posterior out. Tomorrow, today's posterior becomes the new prior.
## A football example you can check
Take a World Cup knockout match. Before kickoff our engine sets expected-goal rates for each team — numbers like "1.45 goals' worth of chances for the favourite, 0.95 for the underdog". Those two numbers ARE a prior: they compress team strength, form, injuries and history into a starting belief, before a single minute is played.
Then the match starts producing evidence. A goal, a red card, twenty minutes of one-way pressure — each event updates the belief. The 49% favourite who concedes early might drop to 30% in minutes. That live number is the posterior, recalculated over and over as evidence arrives.
The same structure runs everywhere forecasts exist: a weather model starts from seasonal base rates before reading today's satellite data; a market maker starts from historical volatility before reading this morning's order flow.
## Why priors matter more than people think
**A good prior keeps you calm.** Without one, every headline drags your forecast around. One shock result and a prior-free forecaster rewrites everything; a forecaster with a well-built prior moves a little, because one match is one data point against years of history.
**A bad prior poisons everything downstream.** If your starting belief says an underdog has no chance at all — a prior of zero — then no amount of evidence can ever move you. Zero times anything is zero. Seasoned forecasters never assign true zero to things that are merely unlikely; the world punishes that arrogance regularly.
**Priors are where honesty lives.** Two forecasters can watch the same match and disagree, and the disagreement usually traces back to their priors. Making the prior explicit — writing the number down before the event — is what separates forecasting from storytelling after the fact.
## Prior vs. the market price
Prediction markets offer a fascinating mirror. The price of a contract is, in effect, the crowd's posterior — thousands of participants' priors, updated by whatever each of them knows, compressed into one number. When an independent model's prior-driven forecast disagrees with the market price, one of three things is true: the model's prior is off, the crowd knows something the model doesn't, or the crowd is leaning on a bias. Watching which one it turns out to be, match after match, is one of the most instructive exercises in applied probability.
## How to build a sensible prior (plain-English rules)
- **Start from base rates.** How often do favourites at this level actually win? How often do underdogs hold a draw? History first, story second. - **Weight slow evidence over fast.** Years of results beat last week's highlight reel. Recent form matters — as an adjustment, not a foundation. - **Never use zero or one.** "Almost impossible" is 2%, not 0%. Certainty is not a probability, it's a resignation letter. - **Write it down before the event.** A prior recorded after the result isn't a prior; it's a story. Timestamps keep you honest. - **Let evidence move you — proportionally.** Strong evidence, big update; weak evidence, small update. The math of that proportionality is Bayes' rule, but the discipline matters more than the formula.
## The takeaway
A prior is simply your starting point, stated as a number and owned in public. It is the difference between "I always knew it" and "here is what I believed before, here is the evidence, and here is what I believe now". Good forecasting isn't about having no assumptions — it's about knowing exactly what your assumptions are, writing them down before reality arrives, and letting reality grade them.