How We Built BracketIQ
Using Math to Outsmart March Madness
It Started With a Simple Question
Who's actually good at picking brackets?
Not who sounds good on TV. Not who has the best highlight reel takes. Who, when you score every single pick across an entire 63-game tournament, consistently gets it right?
We went looking for the answer. We pulled bracket picks from nine of the most prominent college basketball analysts in the country — names you'd recognize from ESPN, CBS Sports, and NCAA.com — and tracked every selection they made for the 2024 and 2025 NCAA tournaments. All 63 games. Every round. Every pick scored.
Finding nine experts who performed above average to genuinely elite was the first challenge. We wanted analysts with public, verifiable brackets who took the exercise seriously — not celebrities filling out picks for content. We landed on a group that includes Jay Bilas, Jeff Borzello, Andy Katz, Gary Parrish, Matt Norlander, David Cobb, and Chip Patterson, among others. Honestly, we wish we had more consensus-level great performers in the pool. But the model we've built from this group is testing as well as Jay Bilas has performed over the past two years — and Bilas is the best individual bracket picker we've tracked. So we'll take it.
One important caveat before we go further: this model doesn't backtest well prior to 2024. The expert pool, the data pipeline, and the analytical framework were all built around the 2024 and 2025 tournaments. We're not going to pretend this system would have dominated 2018 or 2021. But if 2026 is anything like 2024 and 2025 — and the structural dynamics of the tournament suggest it will be — this model is basically guaranteed to finish near the top of any bracket pool you enter.
Here's the road map of how we got there.
Phase 1: Building the Consensus
We scored every expert using the standard ESPN Tournament Challenge system: 1 point for a correct Round of 64 pick, 2 for Round of 32, 4 for Sweet 16, 8 for Elite 8, 16 for Final Four, and 32 for the Championship. A perfect bracket is 192 points. Nobody gets close to that. The best experts land between 100 and 160 in a given year.
To separate skill from luck, we applied recency weights — the 2025 season counts at full value (1.0x), 2024 at 0.8x. An expert's weighted score is the weighted average of their seasonal totals. This penalizes one-year wonders and rewards consistency.
With the experts ranked, we built the first version of our model: a weighted consensus bracket. The idea is straightforward — instead of following one expert, aggregate all nine, but give more influence to the ones who've proven they know what they're doing.
The weighting formula uses square-root compression:
We tried linear weighting first. It was a disaster — Jay Bilas effectively made every pick, which defeats the entire purpose of aggregation. Square root compression flattens the distribution. Top experts get roughly 1.3x influence. Bottom experts get 1.0x. The gap exists, but it doesn't let any single voice steamroll the group.
For each game, every expert casts a weighted vote for their predicted winner. The team with the higher total weighted vote gets the consensus pick. And critically, the bracket cascades — Round of 64 winners feed into Round of 32 matchups, Sweet 16 matchups flow from Round of 32 winners, and so on all the way through the Championship. Without cascading, you'd end up with impossible brackets, teams showing up in rounds they'd already been eliminated from. We found that bug the hard way.
The cascaded consensus scored 139 points in 2024 and 127 in 2025, landing at 5th on our leaderboard with a 132.3 weighted score. It outperformed 7 of 9 individual human experts.
Not bad. But not enough. The consensus was conservative by design — it smoothed out the terrible picks, but it also smoothed out the brilliant ones. We needed something with more edge.
Phase 2: Adding KenPom to the Mix
Human experts bring knowledge, but they also bring bias. Alma mater loyalty. Narrative-driven reasoning. Overreaction to a bad loss in the conference tournament. We wanted a purely data-driven “expert” to sit alongside the humans and see how it stacked up.
We built one using KenPom efficiency metrics — the gold standard in college basketball analytics. KenPom rates every team on three adjusted metrics: AdjO (offensive efficiency, points per 100 possessions adjusted for opponent strength), AdjD (defensive efficiency, same concept for the other end), and AdjT (tempo, possessions per 40 minutes).
We constructed a composite score after extensive backtesting:
The 55/20/25 split wasn't arbitrary. We tested dozens of configurations. 40/30/30. 50/25/25. 60/20/20. We tested versions that incorporated momentum — a team's last-10 game record heading into the tournament — at weights of 10%, 15%, 20%, and 25%. Every single momentum-inclusive version performed worse than the baseline without it. Recent form, it turns out, is noise. The efficiency numbers already capture everything momentum is trying to tell you, and adding it just muddied the water.
The tempo adjustment penalizes extreme paces. Teams below 68 possessions per game get a volatility penalty: -(68 - tempo) × 0.5. Teams above 71 get a sloppiness penalty: -(tempo - 71) × 0.3. Between 68 and 71 is a neutral zone. The penalties are asymmetric on purpose — playing too slow introduces more variance per possession than playing too fast introduces through turnovers.
We added KenPom as a virtual expert on our leaderboard, weighted it into the consensus alongside the human analysts, and backtested it.
The results were excellent. The KenPom model scored 155 points in 2024 — correctly picking UConn as national champion — and 152 in 2025, correctly picking Florida. It sits at 3rd on our leaderboard with a 155.4 weighted score, beating every human expert except Jay Bilas. When we folded it into the consensus, the whole model tightened up. The aggregate picks got sharper. The late-round accuracy improved.
But there was still a problem.
Phase 3: Solving the Coin-Flip Problem
When we started analyzing where the model was losing points, a pattern emerged immediately. It wasn't the 1 vs 16 games. It wasn't the Sweet 16 or Elite 8. It was the first-round upset zone — 5 vs 12, 6 vs 11, 7 vs 10, and 8 vs 9 seed matchups.
We isolated all of these matchups across 2024 and 2025 and cross-referenced each game's KenPom AdjO gap — the difference in offensive efficiency between the two teams.
We found 23 games where that gap was less than 5 points.
The data was unambiguous: 47.8% of those games were upsets. Not a slight lean. A genuine coin flip. The selection committee's seeding carried no predictive value. Expert intuition carried no predictive value. Our consensus went 9 for 23 on these games — worse than if we'd flipped a coin.
This was the leak in the model. These 23 games were bleeding points, and every analytical tool we'd built was essentially guessing on them.
So we went looking for a signal that actually worked. We tested everything we had:
| MODEL | RECORD | WIN % |
|---|---|---|
| Vegas alone | 15/23 | 65.2% |
| Vegas + AdjD + L10 (majority) | 14/23 | 60.9% |
| Vegas + Net Rating (60/40) | 14/23 | 60.9% |
| Net Rating alone | 14/23 | 60.9% |
| AdjD alone | 10/23 | 43.5% |
| Last-10 record alone | 6/23 | 26.1% |
| 5-signal ensemble | 9/23 | 39.1% |
Pre-tournament Vegas futures odds — just the raw betting market — outperformed every statistical combination we could build. And it wasn't close. Every additional signal we layered on top of Vegas degraded performance. The five-signal ensemble, which was supposed to be our most sophisticated creation, performed dead last. Worse than the consensus. Worse than random.
The reason is straightforward once you see it: Vegas odds already incorporate defensive efficiency, momentum, tempo, and a hundred other factors we can't even measure — coaching tendencies, injury reports, locker room dynamics, sharp money from professional bettors. Layering our own versions of those same signals on top is double-counting information the market has already priced in. More data doesn't always mean more signal. Sometimes it just means more noise dressed up in a spreadsheet.
This was the breakthrough. For any first-round matchup where the AdjO gap is under 5 points, we defer entirely to Vegas. For every other game — where one team has a clear efficiency advantage — we use our KenPom model. The hybrid scored 156.8 weighted points across 2024-2025, good for 2nd on our leaderboard, just 0.4 points behind Jay Bilas and ahead of every other human expert and model we tested.
That moved us up a notch. We went from a model that was losing points on coin-flip games to one that captures those games well above industry average. The 5-point rule is the single most impactful finding of this entire project.
The Tempo Wrinkle
While Vegas dominated the aggregate numbers on close games, we found one narrow anomaly worth noting — three specific upsets that no model, no expert, and no Vegas line identified. The only thing that flagged them was isolated tempo analysis:
The pattern: the higher-seeded favorite played below 67 possessions per game, and the AdjO gap was under 5 points. All three favorites had strong offensive efficiency numbers — Texas Tech top 25, Saint Mary's top 30, Clemson top 15. On paper, they looked like safe picks. But efficiency per possession only matters if you generate enough possessions.
Saint Mary's at 61.9 was the extreme case. One of the slowest teams in Division I, they were producing fewer scoring opportunities per game than almost anyone. Their margin for error was razor-thin. One bad 3-4 possession stretch in a 60-possession game is catastrophic. That same stretch in a 72-possession game is recoverable. The underdogs in all three cases pushed pace just enough to create the kind of variance that single-elimination tournaments reward.
Tempo alone doesn't work as a full-bracket model — it scored just 115 points across a 63-game bracket. But as a targeted upset detector in close-efficiency, slow-pace matchups, it caught things nothing else could see.
2026: The Picks Are In
For the current tournament, we identified 8 first-round matchups with AdjO gaps under 5 points — our coin-flip alert games:
For these 8 games, the model defers to Vegas. For the other 24 first-round games, KenPom drives the pick. The full bracket cascades from there through the Championship.
We can't tell you to follow the consensus, or Jay Bilas, or KenPom, or the Vegas/KenPom hybrid. What we can tell you is this: anything at or above our consensus line — 132 weighted points — would put you comfortably in the top 5 of virtually any bracket pool you enter. The models above that line have been right about two consecutive national champions and are picking these coin-flip games at 65% in backtesting when the public picks them at 39%.
So pick your comfort level, trust the math as much or as little as you want, and get wild with it. That's kind of the whole point of March.
What We Learned
Averaging smart people makes you smarter than most smart people, but not smarter than the smartest one. The consensus beat 7 of 9 experts and lost to 2. We're choosing to focus on the 7.
Every additional variable we added to a model made it worse. At some point, data science is just the art of knowing when to stop adding data. We learned this after building a five-signal ensemble model that performed worse than checking the Vegas odds on our phone. Live and learn.
Vegas knows more than you. More than us, too. We spent real time constructing weighted multi-signal frameworks that a sportsbook intern could have outperformed by reading a line sheet. Somewhere, a quant at DraftKings is laughing at our scatter plots.
The slowest team in the tournament is not the safest pick. It might be the most dangerous one. Saint Mary's at 61.9 possessions per game looked like a lock right up until Grand Canyon ran them off the floor. Turns out playing fewer possessions in a single-elimination tournament is less “controlling the game” and more “giving yourself fewer chances to recover from a bad stretch.” We'll remember that one.
The real edge in March Madness isn't knowing who's going to win — it's knowing which games nobody knows. Half the battle is identifying the coin flips before tipoff. Once you know which games are genuinely unpredictable, you stop wasting confidence on them and start putting it where the data actually has something to say. That's not really a basketball insight. It's just math wearing a foam finger.
My 2026 Brackets
So after all this analysis, what did I actually do with it? I filled out three brackets.
I followed the weighted consensus of all 9 experts plus the models almost entirely — except for one pick where I broke from the group. The safe play with one moment of conviction. We'll see if that one deviation costs me or saves me.
Our hybrid model — KenPom efficiency for clear mismatches, Vegas odds for the coin-flip games. The one that backtested at 156.8 weighted points across 2024-2025 and sits second on the all-time leaderboard. This is the one the data says to trust.
I copied my buddy Jay Bilas's bracket pick for pick — the man's been the best bracket picker we've tracked for two years, so why fight it? But I followed my heart on one pick where the data and my gut disagreed. Sometimes you just have to trust the instinct over the spreadsheet. We'll see if that one call makes the difference.
We can't tell you which approach is right. But anything at or above our consensus line would put you comfortably in the top 5 of virtually any bracket pool you enter. So pick your comfort level, trust the math as much or as little as you want, and get wild with it. That's kind of the whole point of March.
BracketIQ is live. The 2026 bracket is locked. Let's see if the math holds up.