How the BracketIQ Models Work

How Everyone Is Scored

Every bracket on this site, human or model, is scored the same way. We use the standard ESPN Tournament Challenge points: 1 for a correct Round of 64 pick, then 2, 4, 8, 16, and 32 for the Championship. A perfect bracket is worth 192 points. First Four games are not scored.

Picks cascade, so a team you eliminated in the Round of 32 can't earn you points in the Sweet 16. Every model on this page fills out a complete 63-game bracket the same way a person does.

The leaderboard ranks by a recency-weighted average of seasonal points. 2026 counts 35%, 2025 counts 25%, and 2024 counts 20%. The weights are rescaled over whichever seasons a bracket has, so an expert with only two seasons isn't penalized for the missing year.

Four of the brackets we track are our own models, described below. Bart Torvik's bracket is his own most-likely bracket from T-Rank. We track it as an outside benchmark, not as one of ours.

The Four Models at a Glance

Season columns show bracket points out of 192. Pick accuracy is the share of all 63 games called correctly, averaged across seasons.

Model 1: Consensus

What it is: a weighted vote. It was the first model we built, in March 2026, before that year's tournament started.

Who votes: eight human experts (Gary Parrish, Matt Norlander, David Cobb, Cameron Salerno, Jay Bilas, Jeff Borzello, Chip Patterson, and Andy Katz) plus the KEN POM model. Bill Bender, Mike DeCourcy, Seth Davis, and Bart Torvik were added to the site in August 2026, so they aren't part of the vote. Neither is Vegas/KenPom or my personal bracket.

How votes are weighted: each voter gets a track-record score equal to their 2025 bracket points plus 0.8 times their 2024 points, divided by 1.8. Their vote weight is the square root of that score relative to the lowest score in the group:

We tried linear weights first. Jay Bilas ended up deciding nearly every game, which defeats the point of a crowd. Taking the square root compresses the spread, so the best voter counts about 1.3× as much as the worst one.

How the bracket is filled out: for each game, the team with more weighted votes gets the pick. Winners carry forward round by round through the Championship.

What we learned: a crowd removes the worst picks, but it also removes the bold ones that win pools. Consensus beats 8 of the 11 human experts, but it trails the other three and all three of our ratings-based models.

Model 2: KEN POM

What it is: our first purely statistical bracket, built in March 2026. It uses Ken Pomeroy's pre-tournament ratings for every team in the field: AdjO (points scored per 100 possessions, adjusted for opponents), AdjD (points allowed per 100 possessions, adjusted), and AdjT (possessions per 40 minutes).

The formula:

The tempo penalty is zero for teams between 68 and 71 possessions. Slower teams lose 0.5 per possession below 68, and faster teams lose 0.3 per possession above 71. The idea was that extreme pace adds risk in a single-elimination game.

How the bracket is filled out: in every game, the team with the higher score advances. The model walks the bracket from the Round of 64 to the Championship.

What changed in September 2026: the original weights were 0.55 × AdjO + 0.20 × (−AdjD) + 0.25 × tempo, chosen by comparing a few splits against the 2024 and 2025 tournaments only. Divide that through by 0.55 and the offense/defense tilt is 0.36 — a 73/27 lean on offense, on two ratings that share the same points-per-100-possessions scale. That was an accident of the weights, not a finding.

We then separated the two things the formula does, across all 1,133 tournament games from 2008 to 2026, changing one at a time:

The tempo penalty was never the problem. Removing it is worth five bracket points and one game across eighteen tournaments — a rounding error. The whole deficit was the tilt, and fixing the tilt alone recovers 158 of the 175 points available (p = 0.002). So we moved the tilt to 0.80 and left the tempo penalty exactly as it was. KEN POM is still a three-input KenPom model rather than a second copy of Lake Hanson.

How it did, and the honest part: under the original weights it picked the champion in 2024 (UConn) and 2025 (Florida), which is why it rated so well early. The rebalanced version picks UConn in 2024, Duke in 2025, and Duke in 2026 — one champion instead of two. On these three seasons the change is worse on the leaderboard: two more games right, but 39 fewer bracket points, because a champion is worth 32 points on its own. Over eighteen tournaments the same change is worth +158. Three seasons cannot tell those two facts apart, and we are showing you the three-season number on the board above.

Model 3: Vegas/KenPom

What it is: KEN POM with a single override for games that look like a coin flip.

The rule: in the Round of 64, if the two teams' AdjO ratings are within 5 points of each other, take the team favored in the pre-tournament Vegas odds. Every other game, including every game after the first round, is decided by KEN POM's formula, and the bracket carries forward from there.

Why: when we built KEN POM, the close 5–12, 6–11, 7–10, and 8–9 games were costing it the most points. Across 2024 and 2025, the betting market picked those close games better than any stat combination we tried.

What the override actually changed: across three tournaments, Vegas disagreed with KEN POM on 9 first-round picks, all in 2024 and 2025. It was right on 6 and wrong on 3. That's +3 games in 2024, and 2025 was a wash. In 2026 the market agreed with KEN POM on every close game, so the two brackets were identical. From the Round of 32 on, the two models have made the same picks in all three years.

What we found later: we pulled closing lines for every tournament game and compared them to a plain efficiency-ratings model. The market went 137 of 183, and the ratings went 138 of 183. The two correlate at 0.981 and disagreed on only 9 picks, which split 5–4. The market is basically the ratings with extra steps, so the Vegas layer is a small tweak to KEN POM, not a separate source of edge.

Model 4: Lake Hanson

What it is: our current best model, built in August 2026. Instead of tuning weights on two tournaments, we tuned them on every NCAA tournament game from 2008 through 2026: 1,133 games in all (there was no 2020 tournament).

Where the data comes from: Bart Torvik's T-Rank adjusted offense and defense, pulled as of the Monday after Selection Sunday for each season. That date matters. Pulling ratings any later mixes tournament games into the ratings used to predict them. The pipeline refuses to run if any team shows more than 36 games. We use Torvik because it's free, open, and has pre-tournament ratings going back to 2008. Over 25,000 games, Torvik and KenPom are equally accurate, so this is not a claim that Torvik is better.

Why these weights: KEN POM originally used a 0.55 / 0.20 split, which works out to 73% offense and 27% defense. Both ratings are on the same points-per-100-possessions scale, so that lopsided weighting was an accident, not a finding. We tested a range of weights across all 1,133 games:

We picked 0.80 with leave-one-season-out testing: hide one tournament, choose the weight using the other seventeen, then check it on the hidden one. All 18 rounds of that test chose 0.80. Compared with KEN POM's original weighting, it gets 19 more games right out of 1,133 (p = 0.04). That same test is why KEN POM's tilt was moved to 0.80 in September 2026.

Why there's no tempo term: tempo is a style, not a measure of how good a team is. KEN POM's penalty on slow teams bets against the kind of team that wins titles: since 2001, champions have averaged about 134th nationally in adjusted tempo. Removing the tempo term made the model better at every offense/defense weight we tested — though only barely, which is why KEN POM keeps its penalty and this model does not have one.

How the bracket is filled out: the same greedy walk. The higher score advances every game, and First Four games are played through by the model rather than assumed.

Honest limits: any defense weight between 0.7 and 1.0 performs about the same. The real finding is that KEN POM's original 0.36 was measurably wrong, not that 0.80 is magic. Lake Hanson ranks above KEN POM largely because it picked Michigan in 2026, the season carrying the most leaderboard weight; the two models now share the same offense/defense tilt and differ only in the tempo penalty and the rating source.

What We Tested and Left Out

Every idea below was tested against the same 1,133 games and didn't make it into Lake Hanson. The baseline gets 826 games right.

How Much Any Model Can Do

Over 1,133 tournament games, picking the team with the better efficiency rating wins 72.4% of the time. Just picking the better seed wins 71.4%. Every rating system is worth about one percentage point over chalk. The edge mostly disappears in the second weekend: ratings get 64.6% of Sweet 16 games and 62.5% of Elite 8 games.

Single seasons swing wildly, from 63.5% in 2011 to 82.5% in 2025. So three tournaments of leaderboard results can't prove one model is better than another. The 1,133-game test is the evidence; the leaderboard is the scoreboard.