How the ratings work
What the rating and its range mean
A team's rating is how many goals a game it would win or lose by, on average, against a typical Division II team on a neutral field. A rating of +0.50 means about half a goal a game better than that typical team. A few games cannot pin a team's level down exactly, so every rating has a likely range: there is an 8 in 10 chance the team's true level sits inside it (the thick bar). The thin line is a wider range, 95 in 100. Early in the season, and for teams that have mostly played their own conference, the bars are wide because there is less evidence. Where two teams' bars overlap a lot, the results so far cannot tell them apart, so the order between them in the list is not settled.
The model
Goals in soccer behave like random counts, so the model treats each team's goals in a game as a random number whose average depends on three things: how good the team is at scoring (its attack), how good the opponent is at stopping goals (its defense), and whether the team is at home. Working backwards from every final score between two D2 teams this season, it finds the attack and defense values that fit the results. It does this with a statistics program called Stan. Because a few games cannot pin down each team's level exactly, the model produces 1,000 sets of ratings that all fit the results so far. Every range on this site shows how much those 1,000 disagree.
Some details matter for how it behaves:
- Teams are compared within their conference first, and conferences are compared through the games between them. That keeps a team from a weak conference with a perfect record from being placed above everyone.
- Each team starts the season close to where it finished last season: it keeps 90% of its gap from its conference's average, with plenty of room to move as results come in. New D2 teams start at their conference average with even more room. Testing on past seasons found that D2 teams' strength carries over strongly from one year to the next.
- Every game this season counts equally. Testing on past seasons found no gain from giving recent games more weight.
- Goals in overtime count. A game decided on penalty kicks counts as a tie.
- Games against teams outside Division II are shown on team pages but left out of the ratings.
- A team's likely rank range comes from ranking all teams in each of those 1,000 sets of ratings and keeping the middle 8 in 10 of the ranks it gets.
On a team page, "Rating +0.05" next to a result is how much that result changed the team's rating: its rating with that game included, minus its rating without it. Refitting the model once per game would take too long, so this reweights the 1,000 sets of ratings by how well each one predicted that game. A small ≈ marks rougher estimates, where few of the 1,000 sets fit the result well.
NCAA tournament chances
The NCAA field is 40 teams: ten from each of four super regions. From 2026 the NCAA picks them by formula. Each of the 19 conference champions gets an automatic bid, and the rest of each super region's ten places go to its highest teams in the NCAA Power Index (NPI). NPI is 25% winning percentage and 75% strength of schedule, measured as the average NPI of a team's opponents. Away wins count for a little more, home wins a little less, and there is a bonus for beating strong opponents. Only games against D2 opponents count, and a team needs ten of them to be eligible.
To estimate each team's chance, the site plays out the rest of the season 1,000 times (the "simulated seasons"), once with each set of ratings. Each time it plays every remaining scheduled game and each conference tournament, works out every team's NPI, and picks the 40 teams. A team's NCAA chance is the share of those 1,000 seasons in which it is picked.
Three parts of this are approximations:
- The NCAA has published NPI's ingredients and settings but not every step of the calculation. Ours follows the published description. We will check it against the NCAA's own NPI table once that appears on ncaa.com from October 12.
- Conference tournaments are simulated as a simple knockout among the top teams in the conference standings, sized from last season's tournament, until the real brackets appear in the schedule.
- We assume each super region's remaining places go strictly in NPI order. That is our reading of the 2026 handbook, which does not say so in one sentence.
How well it predicts
We tested the model only on games it had not seen: for every week of a past season, we fitted it using only the games played before that Monday, predicted that week's games, and scored the predictions once the results were in. Settings were tuned on the 2024 season, and the results below are for 2025, a season the tuning never saw.
In plain terms, the model predicted the 2025 season a little better than a well-tuned Elo rating, though the gap is not yet large enough to rule out chance.
| Method | Games | Prediction error (log-loss, lower is better) | Prediction error (Brier, lower is better) | Goal margin inside the likely range |
|---|---|---|---|---|
| This site's model | 1,773 | 0.938 | 0.554 | 89% |
| Simple Elo rating (the chess-style rating many sites use) | 1,773 | 0.946 | 0.559 | – |
| Home win 50%, draw 25%, away win 25% | 1,773 | 1.072 | 0.648 | – |
Log-loss and Brier score both measure how much probability a method put on what actually happened (win, tie or loss); lower is better. Across the same 1,773 games, the model's log-loss was 0.008 lower than Elo's (standard error 0.005). The last column checks the ranges: the actual goal margin should land inside the likely range about 8 times in 10. It runs a little above that because margins are whole goals.
Both are far better than guessing from home advantage alone. The model's advantage is that it gives every team a range and explains each result's effect, which Elo cannot.When the model says 30%, does it happen 30% of the time?
Each point groups predictions with a similar chance and shows how often those results actually happened. Points close to the diagonal mean the percentages can be taken at face value.
Picking the 40 teams
Applying the 2026 selection rules to the final results of past seasons (actual conference champions plus the NPI at-large rule):
- 2023: 27 of 40 selected teams matched. NCAA's data for this season starts in October, so the formula saw only half the season.
- 2024: 38 of 40 selected teams matched.
- 2025: 38 of 40 selected teams matched.
Those seasons were picked by a committee using older criteria, so a perfect match is not expected; this mainly shows how far the new formula and the old committee agree.
We would like to compare against Massey Ratings, the best-known computer ratings for D2 soccer. Its terms do not allow its numbers to be copied or used in building models, so we have not done that without permission.
Data
Scores and fixtures come from NCAA.com's public scoreboards, read every morning at about 6 a.m. Eastern, after the last West Coast games. Each run re-reads the previous three days so late and corrected scores are picked up, and once a week the whole season is re-read. The feed has no neutral-site flag, so a game is treated as neutral when it is played away from both teams' usual home city; some neutral games without a listed venue will be missed.
This is an independent project and is not affiliated with the NCAA.