Model B
A goal-margin rating for high-school soccer, tested against each state's official ranking, or its standings where there is none, on games the methods had not yet seen.
What this is
State associations rank teams to seed their playoffs. Most rank by wins and losses in some form: a formula such as RPI or power points, or the standings, a coaches' vote or a draw. A handful count goal margin, among them the CIF Southern Section, Massachusetts and Colorado, and the states that seed by the MaxPreps rating. Model B is a margin rating. It counts goal margin and strength of schedule together, the way the Southern Section's Massey ratings and the MIAA's power rating do.
Model B comes in two versions. One starts each season from last season's standing, which fades as games are played. That version is the one registered for the 2026 tests, and both versions are in every frozen file and both are scored. The other uses only the current season's games. The figures on this page are for the version without the head start unless they say otherwise. The head start helps most in the first weeks of a season; by the playoffs it adds under a hundredth of log-loss and no measurable number of games called right.
Each page on this site compares a team's standing under its association's official ranking with its standing under Model B. It shows the playoff bracket as posted beside the bracket Model B's order would give, and it gives Model B's probabilities for the team's next game. Every page is made from ratings frozen before the playoffs, and the frozen file is not changed afterward.
How Model B is tested
Where an association publishes a rating, it was rebuilt from the association's own rule. At about half the sites the rebuilt table matches the posted one exactly; at the rest it matches in rank order, and the report lists each site's gap and its cause. Where an association ranks by standings or a vote, the comparison is to its standings rule or to winning percentage, and the seeds the coaches actually chose are tested separately on the bracket games.
Each method is recomputed as of every game date from the games before it and judged on the games that came next, so no method is judged on a game it has already seen. Two measures are reported, and each number below says which it is. Games called right is the share of games whose result was the one the method called most likely. Log-loss scores the probability the method gave the actual result, and it is the measure used to decide whether a difference is too close to call. The error bars allow for the fact that one team's games are related to each other.
The settings were chosen by leaving each of Oregon's 24 boys' and girls' seasons out in turn and keeping what did best on the rest, then checked on 31 seasons in eight other states, Massachusetts and the Southern Section among them. On September 30, 2026, before any 2026 result was scored, three settings changed: the cap on margin was removed, the pull toward the class average was lightened, and the head start was reweighted. The choice comes out the same whichever Oregon seasons, or whichever whole states, are left out of the test. No 2026 game entered it. The old settings are scored beside the new ones on every 2026 freeze and shown in the Oregon and multi-season tables of the report.
What the tests found
In Oregon, rebuilt from the OSAA's own pages for the twelve boys' and twelve girls' seasons since 2013 (the 2020 season had no rankings), Model B called 66 of 100 6A boys' games right and the OSAA rank 64, Model B ahead in all twelve seasons; in 6A girls' games, 72 to 68, ahead in ten of twelve. Across 6A, 5A and 4A together, Model B called 3.2 more boys' games per hundred (error bar 2.4 to 4.0) and 3.6 more girls' games (2.8 to 4.4): 72 to 69 and 75 to 72. In regular-season games between teams that both made the playoffs, 56 to 52 for boys and 62 to 58 for girls. In the playoff games themselves the two methods call the same number right, within a point either way; on log-loss Model B is ahead there too, by less than in the regular season.
In 45 other associations and sections, Model B beat the official ranking or standings on log-loss in 89 of 93 boys' and girls' seasons and lost in none. The four too close to call were all on 33 to 101 games; when the eight smallest sites were given two or three seasons each, every one but the District of Columbia came clear. On games called right, over about 85,000 regular-season test games, the official methods called 69 of 100 and Model B 76. Against standings and winning percentage the gain is about 8 games in 100; against RPI-style formulas about 5; against the MaxPreps rating, which already counts margin, about 3 or 4. In two of the 93 seasons, and at one pooled site, Model B called slightly fewer games right; all three are too close to call.
Against the ratings that already count margin, Model B is close, not clearly better. On the Southern Section's 261 boys' playoff games the two are equal to the third decimal on log-loss, 0.921 to 0.920, with an error bar of about plus or minus 0.05; on the girls' 225 games Model B is ahead of the official rating by about 0.05, with an error bar that just clears zero. Both comparisons use playoff games only, because the section posts Massey ratings only at its seeding lock. On four regular seasons of Massachusetts games, both sexes, with no venues and no playoff games in the MIAA's set, Model B is ahead of the MIAA's rating by two to three hundredths of log-loss under the September 30 settings; under the earlier settings it tied for boys and trailed for girls, and those seasons were among the 31 that confirmed the new settings.
The seeds that coaches and committees choose themselves were tested on bracket games at about twenty sites, and they are the strongest comparison found. Most comparisons are too close to call. Model B is clearly ahead at Wisconsin boys on log-loss (229 games), at Iowa on the committee's ranking order (93 games), and at two small sites on games called right, Hawaii boys (17 games) and Wyoming girls (28). It is clearly behind at CIF North Coast boys on recorded results.
The differences reach the playoff field. At the 20 sites where a rating decides who qualifies, applying each association's own rule with Model B's order in place of the official one changes 188 of about 3,600 qualifiers in one season: about one in twenty of the whole field, and about one in thirteen of the places the rating itself fills, since automatic bids take the rest. Several of those rules had to be inferred from the posted fields rather than read, and the report says which.
Pages
Oregon School Activities Association: North Salem, boys 2026, 6A; Sherwood, boys 2026, 6A; St. Mary's Acad., girls 2026, 6A; Tigard, boys 2026, 6A