How we rank D2 volleyball

The short version

• Points beat wins. How many rallies a team wins says more about how good it is than whether it won. Systems built on wins alone (the RPI, the NCAA's NPI, Colley) predicted worst in every season we tested: log loss 0.500 against our 0.430 over 9,254 matches.
• Conference strength has to be measured. D2 conferences mostly play each other, so a 17–1 record in one conference and a 17–1 record in another can mean very different things. We judge each team against its own conference, and each conference by how its teams do against everyone else.
• Recent form counts more, but conference strength keeps its evidence. The non-conference matches that tell conferences apart are played early, so they never fade.
• Last season is the starting point, then fades as this season's matches arrive. That's why our early-season numbers are steadier than most.
• Two questions, two rankings. Résumé ranks who has earned the most; Power ranks who'd win tomorrow. We show both, plus the NCAA's NPI, because they answer different questions.

How we test

• Walk forward, no hindsight. For every match day, each system is fitted only on matches played before that day, then predicts that day's matches. Nothing it predicts was in the data it learned from.
• Same matches for everyone. Every system predicts every match between two D2 teams that have each played at least once that season.
• Score the confidence, not just the pick. Our main score is log loss: calling a team a 90% favorite and losing costs far more than calling it 55% and losing, so a system can't look good by being overconfident. Lower is better; flipping a coin scores 0.693. We also track the share of winners picked and the subsets that matter most: cross-conference matches and the NCAA tournament.
• Is the difference real? We test it two ways. For win chances, a paired bootstrap: resample the matches thousands of times and count how often the other system comes out ahead. For rankings, McNemar's test on the matches where the two rankings disagree. "p < .001" means a gap that size would happen by luck less than one time in a thousand.
• Tuned on one slice, checked on others. The model's settings were chosen on the first weeks of 2026, and the win-chance calibration on 2024 alone; both were then confirmed on full seasons they never saw, so the results aren't fitted to the test.

The results

All 3 seasons together (2024, 2025, 2026): 9,254 matches

Averages are weighted by matches. The ranking test pools every match where the two rankings disagreed. The coaches' poll only ranks 25 teams and covers far fewer matches, so it takes several seasons of disagreements to separate it from ours; this test grows every week.

Season by season

2026 so far 2025 2024

This season's results update every day as matches are played (last run Oct 8, 4:56 AM CT).

The systems

How the Power rating works

• The evidence. For every match, the log of the ratio of rallies won (49 to 38 is ln(49/38) = 0.25). Weighted by how many rallies were played.
• The model. That number ≈ rating(home) − rating(away) + a home edge, fitted across all matches at once so every result is judged against the strength of who it came against.
• The hierarchy. Rating = conference strength + standing within the conference. A team with little evidence is assumed to be like its conference, not like the average D2 team. Conference strength comes from cross-conference matches, which hold full weight all season.
• Recency. A team's standing within its conference weighs a match half as much after 30 days.
• Robust weights. A result far from what was expected (a surprise blowout) counts a little less, so one strange night can't swing a team.
• Last season. 90% of each conference's strength and 70% of each team's standing in it carry over, worth about one match of evidence, and fade as this season's matches come in.
• Serve and tempo adjustment. Not every point repeats. Teams that win points with service aces, or by forcing passing errors, do a little worse next time than their point totals suggest; teams that get aced a lot also do a little worse; teams with more attacks per set (long rallies, defense into transition) and more setter assists per set do a little better. Each team's season-to-date numbers adjust its rating slightly. We found it in 2025 and 2026 and confirmed it on 2024, which played no part in finding it; each season below is scored with weights trained without it.
• Win chances. Rating differences give the chance of winning a rally; the scoring rules (sets to 25, a fifth to 15, win by 2) turn that into set and match chances. The ± on the rankings page is one standard deviation: teams whose ranges overlap are close to even.

Résumé

For each team: how likely would the 25th-best team be to win at least as many matches against the same opponents, at the same sites? The less likely, the more impressive the record, and the higher the Résumé rank. Only wins and losses count; the power ratings judge how hard each match was. It's the same idea as ESPN's Strength of Record and Bart Torvik's Wins Above Bubble, which the NCAA's basketball committee now uses.

What we tried that didn't help

• Sideout and serving as separate ratings (the volleyball version of KenPom's offense and defense). It predicted worse than one rating: two numbers per team means twice as much to estimate from the same matches, and sets are decided almost entirely by the overall share of rallies won.
• Counting wins or sets on top of points. Predicted worse in every version we tried.
• Capping blowouts outright. Worse than down-weighting surprises.
• Recency on everything. Fading old matches across the board weakens conference strength, because the matches that measure it are played early.
• Travel, time zones, start times and rest. Over three seasons, how far a team traveled, which way, how many time zones, its body clock and days of rest made no difference beyond what the rating already expects.
• "Clutch." Winning rallies late in close sets, close-set records and recent form didn't carry over from match to match.
• Box-score stats in bulk. Of 31 team and position stats, only the serve and tempo ones above added anything; the rest are already in the points.

Keeping it honest

• This season's test reruns every day. If the numbers slip, you'll see it here.
• Every change to the model is tested the same way before it ships, and logged.
• Our match predictions, saved before each match, are graded on the accuracy page.

Sources: Pablo FAQ · Massey (1997) · NCAA D2 NPI overview · KenPom methodology · LRMC (Kvam & Sokol) · Ranking and schedule connectivity (Osting et al.) · RPI