How to read our projections

Every projection on this site is the model's central estimate of a statistical total for a single match — projected throw-ins, projected corners, projected shots on target, and so on. Around that number sits a probability and a confidence rating. None of those three pieces is decorative; together they tell you what the model thinks and how much weight to put on it.

• High — the historical signal is strong, the 2018 and 2022 sub-samples agree closely, and contextual filters reinforce the projection rather than fighting it. Throw-ins land here most often.
• Medium — a real signal with caveats. The market is reasonably reliable historically but one or two filters pull in opposing directions.
• Low — the data is thin, the sub-samples disagree, or the match itself is hard to read (extreme heat, first 48-team-format fixture, etc.). A Low rating is a warning, not a hidden recommendation.

A 62% probability does not promise the outcome. It is the model's estimated long-run rate: across many matches with the same statistical profile, the threshold would land on the "over" side roughly 62 times in every 100 if the model is well calibrated.

That word — calibrated — is the one that matters. We do not chase the highest probability; we chase the right one. When we say 62%, our job is to be right at 62%, not to be loud at 62%.

• Read the confidence rating before the number. A 70% Low looks bolder than a 58% High, but the High is the one with the firmer ground under it.
• Treat the projection as one input to your understanding of the match — alongside lineups, fitness, weather and your own watching. It is not a verdict.
• Check the market hub before reading the per-match line. It explains why a market is rated the way it is in our framework.
• If you disagree with a projection, the methodology page exists so you can disagree precisely. That is healthier than agreeing for the wrong reason.

The next essays in this series will be projected-versus-actual reviews, posted after each matchday of the tournament. They will be honest: where the model nailed it, where it missed, and what — if anything — we are changing. The point of publishing the record is to be held to it.

Pair this with the methodology, the glossary, or the live reliability ladder.