The word “AI” promises more than it should. It is more useful to think of the model as a calculator: it takes historical numbers and returns a probability estimate for each outcome. There is no magic in it, and that is good news — a calculator can be checked.
What it computes
The model estimates team strength from played matches, accounts for home advantage and league characteristics, and converts that into probabilities for results, totals and other markets. The AI predictions section gathers estimates across upcoming matches, and the match page shows them broken down by market.
The estimates are calibrated: if the model says “30%”, the outcome should occur in roughly 30% of all such predictions. Calibration is checked on matches the model did not see during training — otherwise it would be grading its own homework.
What the model cannot know
- News. An injury to the first-choice striker announced an hour before kick-off is invisible to it.
- Motivation. That the match means nothing because a European tie follows in three days.
- Circumstances. The pitch, the weather, the travel, a falling-out with the manager.
This is not an implementation flaw but the boundary of the method: none of the above is expressed in historical numbers.
How to use it
The model agreeing with the odds gives you nothing: paying the margin to have a price confirmed is pointless. All the interest is in disagreement — and here honesty is required. A disagreement means one of two things: either the model found something the market missed, or it does not know something the market has already priced in. The second happens more often.
So the order is: you see a disagreement, then look for what the model might not know. Find it (injury, rotation, a dead rubber) and the disagreement is explained, with no bet in it. Fail to find it — then it is worth considering.
More on which models specifically run inside verified picks is in the article about the verifier’s models.
About model agreement
When several independent estimates converge on one outcome, that carries more weight than a single one. But the catch is known too: models trained on the same data make the same mistakes. The analysis is in the article on consensus.
Next
The verifier — a section where people make the picks and a machine checks them.
The model is a probability calculator over historical numbers. Its value lies in disagreeing with the odds — and a disagreement more often means the model is missing something.