"AI predictions" has become a sticker that tipster sites put on the same old guesswork. This page explains what a real prediction pipeline does, step by step, using ours as the worked example. No magic anywhere, which is rather the point.
Step 1: the data
A prediction is only as good as what goes into it. Our pipeline pulls, every thirty minutes:
- Odds from the market, which compress thousands of opinions into a price
- Team form, weighted so last week matters more than last season
- Expected goals (xG), which measures chance quality rather than just results. A team that lost 1-0 while creating 2.4 xG of chances is usually better than its scoreline
- Lineups and injuries, because a missing first-choice striker moves a match more than most narratives do
- Head-to-head history, used cautiously; old fixtures fade fast
One source of truth, refreshed on a schedule. No scraping Twitter rumours.
Step 2: the model weighs signals against a baseline
For each fixture, the model compares today's signals with a historical baseline: in matches that looked like this one, what actually happened? The output isn't a verdict, it's a probability for each outcome and market.
This is where "AI" earns its keep. A human can hold five factors in mind; a model holds every one of them across thousands of past matches and doesn't get sentimental about big clubs.
Step 3: calibration, the unglamorous bit that matters most
Raw model probabilities are usually overconfident. Calibration corrects them against reality: if picks the model called 70% only land 60% of the time, the model's 70s get pulled down until the numbers tell the truth. You can check how well this is working on our performance page, where the 30-day calibration curve is public.
Picks that end up below the confidence threshold get dropped. Nothing ships just because the fixture is famous.
Step 4: publish, with the reasoning attached
Every pick that clears the threshold goes out with its confidence score and the reasoning behind it: the form delta, the xG picture, the injury news that moved the number. After the match, the result is graded automatically, win or lose, and joins the public archive.
That last part is the real difference between a pipeline and a tipster. A tipster's bad week disappears from their Telegram channel. A pipeline's bad week is in the archive, plotted on the curve, forever.
What no model can do
Worth saying plainly:
- No model predicts a red card in the 12th minute, a penalty given by VAR, or a goalkeeper having the game of his life.
- No model turns a 65% probability into a certainty. A 65% pick loses more than a third of the time, and that's when it's priced correctly.
- No model removes the need for staking discipline. Probabilities guide decisions; they don't guarantee outcomes.
Anyone selling AI predictions without those caveats is selling the sticker, not the pipeline.
See it running
The full slate, with confidence scores and reasoning on every pick, is on the predictions page, and the graded record is on performance. Read next: how accurate are football predictions, really?