How our predictions are made

Step 1, we read every result

The model keeps a year of finished games for every league we cover. Each game counts less as it gets older, so what a team did last month matters more than what it did last autumn.

From those results it works out two numbers for every team: how much it scores and how much it lets in, at home and away separately. Playing at home is worth more in some leagues than others, so that is measured per league instead of using one number for the world.

Step 2, we turn strength into a score line

For football the two strengths give an expected goal number for each side. Those two numbers feed a Poisson calculation with a low score correction, which gives the chance of every score line from 0-0 up to 10-10.

Every market comes out of that one grid. Home win, draw, away win, both teams to score, over and under, correct score and handicaps all agree with each other because they come from the same place.

For the NFL, college football, MLB, NBA and NHL the same idea runs on points or runs instead of goals, using a normal curve around the expected margin and the expected total.

Step 3, Eagle 2.0 adds what the results cannot see

A table of results does not know that a team played 48 hours ago, or that its main striker is suspended. Eagle 2.0 folds in four more readings before anything is priced.

• Recent form. Each game arrives with the form both teams carry from every competition, including cups that sit outside our league pool.
• Rest days. A side playing its second game in three days creates less. A side that has not played for three weeks starts slowly.
• Team news. Our news desk reads injury and suspension reports and marks the team. Several reports move the number, one rarely does.
• League home edge. The size of home advantage is measured from that league own results rather than assumed.

Every one of these is deliberately small. Together they can never move an expected goal number by more than 12 percent, because a model that swings on one headline is a model you cannot trust.

Step 4, we compare our number with the price

The bookmaker price carries a margin. Remove the margin and you get what the market really thinks. For the markets the model is not fitted to, such as totals and handicaps, we lean on the market, because it usually knows more than we do about a single line.

Value is the gap between our chance and the price. If we make it 50 in 100 and the price is 2.30, the price pays more than it should. Only bets in that direction reach the site, and only between 3 and 15 percent, because a 40 percent edge on a real market almost always means our number is wrong, not the bookmaker.

Step 5, we grade every result and learn from it

Each game gets three picks: a safe one, a value one and a bold one. When the game finishes they settle on their own and go on the public record page, win or lose.

Those settled bets feed back into the model. Our first batch showed that we were too hopeful on long shots: we called them at 35 in 100 and they landed 27 times. Eagle 2.0 builds a correction curve from what really happened and applies it to every new price, so the same mistake gets smaller every week.

What we will never do

• Hide a losing week. Every settled bet is on the record page the moment the game ends.
• Invent a number. Nothing on the site is written by a language model pretending to be a model.
• Sell a guaranteed win. Anyone selling one is selling you something that does not exist.
• Change a pick after kick off. Picks freeze two hours before the game starts.

Note: If you ever see a number here that does not match the page it came from, tell us on the contact page.