The Sunday before March Madness produces a familiar ritual. ESPN reveals the bracket. Bracketologists explain what went right and wrong. Inside an hour, the model crowd has published predictions, the upset-finder columns are live, and a national audience tries to make sense of a tournament that, mathematically, defeats most of its own models within two rounds.
This is the fundamental tension of NCAA Tournament prediction. The single-elimination format compresses an extraordinary amount of variance into a handful of games. The best public models — KenPom, Bart Torvik, ESPN’s BPI, the official NET ranking — all describe the regular season honestly. None of them can fully predict the bracket, because predicting it would require eliminating the variance that defines tournament basketball.
The piece below is the working guide to bracketology in the modern era. What each major model actually measures, where they outperform polls, where they fail in predictable ways, and the short framework for reading bracket predictions without taking any single one as gospel.
Quick read: NCAA Tournament models in 60 seconds
- KenPom and Bart Torvik are the leading public efficiency models. They rank teams by points per possession adjusted for opponent strength.
- ESPN’s BPI incorporates similar inputs with proprietary tweaks; ESPN also publishes its own bracketology team.
- NET ranking is the NCAA’s official selection tool, blending game results, strength of schedule, and location.
- Models agree on the top: The favorites are usually obvious. They diverge sharply on the 4-12 seed range.
- What models cannot predict: Single-elimination variance, hot-shooting nights, the foul trouble that swings a Sweet 16 game.
How the major bracketology models actually work
Four public models drive most serious NCAA Tournament conversation. Each uses different inputs and produces slightly different rankings, but the underlying logic shares more than it differs.
KenPom (Ken Pomeroy’s site) ranks teams by offensive and defensive efficiency per 100 possessions, adjusted for opponent strength. The methodology has been public since 2002 and has been the de facto standard for college basketball analytics for nearly two decades. Subscriptions unlock detailed splits; the top-line rankings are free.
Bart Torvik publishes a similar opponent-adjusted efficiency model with slightly different methodology and several useful additional tools (matchup predictors, tournament simulations). Torvik’s model and KenPom’s tend to agree on the top 30 teams and disagree more meaningfully outside it. Bart Torvik’s site publishes the full ranking and a popular bracket-prediction tool every March.
ESPN’s BPI (Basketball Power Index) uses similar efficiency-based inputs plus proprietary adjustments. The model’s public-facing version on ESPN appears with bracket projections during the season.
NET ranking is the NCAA’s official ranking system, used by the selection committee. It blends efficiency-style team rating, game results weighted by location and opponent quality, and a “quadrant” system that categorizes wins by opponent quality and home/away context. The NET is what the committee actually uses to seed the tournament; KenPom and Torvik are what analysts use to argue about the seeding.
What each model is best and worst at
The table below maps the four models against the kinds of predictions they handle well and the kinds they struggle with.
| Model | Strongest at | Weakest at |
|---|---|---|
| KenPom | Season-long team quality rankings; tempo-adjusted efficiency | Single-game variance; late-season hot streaks |
| Bart Torvik | Matchup-specific predictions; tournament simulations | Low-major teams with limited data |
| ESPN BPI | Mainstream-friendly summaries; bracket projections | Methodology opacity (proprietary inputs) |
| NET ranking | Capturing what the selection committee actually weights | Predictive accuracy in early rounds |
| Polls (AP, Coaches) | Reflecting public consensus | Adjusting for opponent quality; lag behind reality |
The pattern is consistent. Efficiency-based models (KenPom, Torvik, BPI) outperform polls in predicting individual game outcomes and out-of-conference performance. The NET ranking outperforms all of them at predicting selection committee behavior. None outperforms a coin flip dramatically in single-elimination first-round games between 7-seeds and 10-seeds, where the seeding implies one team is meaningfully better than the other and the data often disagrees.
The patterns the models surface every March
Several patterns recur across NCAA Tournament seasons, and the better public models tend to flag them consistently.
The “good-by-record, bad-by-efficiency” team. A team that wins 25 games in a weak conference often ranks much lower in KenPom and Torvik than its record suggests. These teams tend to underperform their seeding in the tournament, frequently losing in the first or second round despite favorable bracket placement. The companion read on how schedule quality affects this conversation lives in our forthcoming piece on strength of schedule.
The “high-variance offense” team. A team that lives on three-point shooting (above 38% of attempts from beyond the arc) is mathematically more likely to experience a bad-shooting game at the worst possible time. The models flag these teams as high-variance regardless of season-long ranking. The pattern explains several memorable Sweet 16 upsets over the years.
The “elite defense, average offense” team. Teams with top-15 defensive efficiency and middle-of-the-pack offensive efficiency tend to outperform their seeding in tournament games. Defense travels better than offense in single-elimination basketball. The model rankings often place these teams higher than the committee’s seeding implies they deserve.
The “transitioning star” team. A team whose primary creator transferred mid-season or was acquired late often ranks higher in efficiency models than the committee’s eye-test seeding. These teams have stabilized statistically but lack the win record that the committee values. The framework on how role context affects evaluation applies here — see our context problem piece.
A reading framework for bracket predictions
The table below is the workflow we use before quoting any bracketology projection. The job is to identify which model’s prediction the writer is using and whether that prediction has earned its place against alternatives.
| Question to ask | What it reveals | What it changes |
|---|---|---|
| Which model is the projection based on? | Whether the predictor used efficiency or polls/records | Efficiency models earn more analytical weight |
| Do KenPom and Torvik agree on this matchup? | Whether the underlying ratings produce consensus | Agreement = higher-confidence prediction |
| What is the historical accuracy of this prediction style? | Whether the source has earned trust over multiple tournaments | Multi-year track records carry more weight |
| Does the projection account for shooting variance? | Whether high-three-point teams are flagged as volatile | Variance-aware predictions age better |
| How does the projection treat low-major teams? | Whether the model has enough data on conference outliers | Low-major projections are inherently less reliable |
| Does the projection cite specific weakness/strength matchups? | Whether the analyst did the matchup-specific work | Matchup-aware predictions earn more credit |
| Is the sample of past tournament games big enough? | Whether the predictor is generalizing from too few brackets | See our small samples piece |
The framework is the version we run before publishing or sharing any bracketology projection. The careful writer names which model they used, why, and what the model is bad at predicting. The lazy version cites a single bracket as if the tournament were already decided.
Where the eye test still beats the models
Bracketology models capture team-level inputs better than any other public tool, but several factors remain harder to quantify.
Late-season injuries. A team’s best player tearing an ACL in the last week of the regular season often produces a tournament team that statistically resembles the healthy version but plays like a different squad. Public models update slowly to this kind of change. Beat writers catch it immediately.
Coaching adjustments between games. A first-round opponent provides 40 minutes of fresh tape. Coaches who scout aggressively can install scheme adjustments that the second-round model has not yet priced in. This is most visible when an inexperienced staff faces a well-coached underdog and the model’s “by talent” prediction breaks against the “by coaching” reality.
Senior leadership effects. Teams with significant senior experience tend to handle tournament pressure better than freshman-heavy rosters of similar talent. The data captures some of this through cumulative production, but the eye-test version is more reliable than the small-sample analytical version.
The companion read on balancing data with direct observation lives in our match-reading workflow piece.
Frequently asked questions
Which model should I trust most for filling out my bracket?
KenPom and Bart Torvik are the strongest public efficiency models. Use both. When they agree, you have a high-confidence prediction. When they disagree by significant margins, the matchup is genuinely uncertain and either side could win. Treating any single bracket as definitive misses the variance that defines tournament basketball.
Why does the NET ranking sometimes disagree with KenPom?
Because they measure different things. NET incorporates the NCAA’s “quadrant system” weighting of game results based on opponent quality and location. KenPom is a pure efficiency model. The NET rewards winning quality games more directly; KenPom rewards underlying performance regardless of result. Both have merit. Selection committee uses NET; predictive accuracy slightly favors KenPom.
How often does the model favorite actually win the championship?
About 35-40% of the time across the modern era. The remaining majority of tournaments produce a champion ranked between #2 and #10 in the leading models, with occasional deeper runs by teams ranked outside the top 25. The bracket’s single-elimination structure creates more variance than the regular season’s quality differences can fully overcome.
Where can I read bracketology coverage without becoming insufferable?
The Athletic, ESPN’s Joe Lunardi, and CBS Sports all publish bracketology that explains its methodology. Ken Pomeroy’s site and Bart Torvik’s site provide the underlying ratings most of these writers reference. Combining a serious analytics source with one mainstream bracketology column tends to produce the most honest read.
The takeaway, in one paragraph
NCAA Tournament prediction is the cleanest case study in sports analytics of a discipline that has improved dramatically without fully solving its core problem. KenPom, Bart Torvik, BPI, and the NET all describe the regular season more honestly than the polls ever did. None of them can fully predict the bracket because the bracket compresses too much variance into too few games. Reading bracketology well in 2026 means using the models, naming their limits, and saving room for the upsets the models cannot price. For the broader vocabulary this conversation sits inside, our sports analytics field guide is the natural companion read.
