How predictions are made

A team's rating moves only when a real score arrives: a result credited to a named source, a result a reviewer accepted, or a final that several people at the game agreed on. Fixtures the model generates so there is something to predict are never learned from. A generated score is not a fact about either team, and a rating trained on one would only be predicting the generator.

Every result is checked each week: no game counted twice, no school stored under two names, nothing outside the football season, no impossible score.

A team starts each season from its most recent rating in any earlier season, pulled a tenth of the way back toward average for every year that has passed. That tenth was measured, not assumed: established programmes stay strong from one year to the next far more than roster turnover suggests, and a bigger pull made the predictions worse. A programme's record is never thrown away because one season had few results on file.

The gap between two ratings becomes a predicted margin in points, adjusted for home field and for how far the schools travel. The model then plays the game thousands of times with realistic football scoring — touchdowns, field goals, the occasional safety — and the win chance is simply how often each side came out ahead.

How many points a game will produce comes from each team's real scoring history: points scored and allowed per game across every earlier season, with last season counting most and each season before it counting half as much. That history is worth about two games; once a team has played this season, its own results take over. Two high-scoring offences get a high total, two defences a low one, and both are measured against the league's real scoring level.

Two different checks, kept apart on purpose.

Calls made before kickoff. Every line the board posts is recorded before the game and graded when the real final is in. So far that is 127 games this season: the model's favourite won 72 in every 100, and its predicted margin missed by about 23 points on average. That is a small sample, it grows every Saturday, and this paragraph will say so as it does.

Replaying the past. The model's settings are tuned by replaying earlier seasons game by game, each prediction using only the games before it. That tells us which settings work better. It is a test of the method, not a record of calls anybody could have seen in advance, so it is not quoted here as the model's accuracy.

Its points totals are, so far, no better than the league average, and its margins still run short in mismatches. Those are the next things being fixed.

Injuries, weather, a transfer who arrived in August, a coach who changed the offence. The older part of the record leans toward ranked teams, which makes the very best programmes look closer together than they are; this season's comes from whole states' schedules, so the picture is getting wider every week. Where a state has no real results yet, the model knows almost nothing and its win chances sit near even, which is the honest answer. That is the gap your own knowledge of your team fills, and it is why this is a game of skill.

Pick games, set how confident you are, and see how your calls compare with the model's and everyone else's once the real score is in. Confidence is measured in units, which are a score and not money: they cannot be bought, cashed, transferred or redeemed, and there are no prizes. Nothing on this site takes money or accepts a wager.

Picks open for everybody on the same day: the Thursday a game's week comes into the two-week window.