The Model vs. March Madness: Sweet 16 Predictions
The machine filled out my bracket. Then Round 1 happened, and the machine learned what the rest of us already knew: March doesn't care about your efficiency metrics. Now the Sweet 16 is set, and the model has opinions about all eight games. Some of them might even be right.
The Scorecard So Far
Through two rounds, the model went 33 for 48, a 68.8% hit rate. It got 22 of 32 first-round games right and 11 of 16 in the Round of 32. Picking the higher seed every time gets you about 65-70% depending on the year. The model is finding real signal, just not enough to survive March intact.
The Round of 32 is where things got interesting. The model correctly called St. John's over Kansas at 63.2% confidence, a genuine upset pick that landed. It had Michigan State over Louisville, which hit. It correctly identified Houston and Illinois as comfortable winners.
But five R32 picks went sideways, and they weren't small misses:
Iowa 73, Florida 72. The model gave Florida an 83.2% chance. Alvaro Folgueiras drilled a three with seconds left and sent the defending national champion home. First 1-seed to fall, and the model's biggest whiff of the tournament.
Nebraska 74, Vanderbilt 72. Model had Vanderbilt at 60.9%. Two points the other way.
Texas 74, Gonzaga 68. Model had Gonzaga at 76%. The 11-seed Longhorns, who came through the play-in game, are now in the Sweet 16.
Alabama 90, Texas Tech 65. The model's slimmest margin pick at 56.1% for Texas Tech, and Alabama didn't just win, they dismantled them.
Tennessee 79, Virginia 72. Model had Virginia at 62.9%. The Vols didn't care.
Those five upsets blew up half the bracket. Four of the model's eight projected Sweet 16 matchups no longer exist. So I re-ran the model with actual results locked in, same stacked ensemble, same 10,000 Monte Carlo simulations, and here's what it says about the eight games ahead.
East Region: Washington, D.C.
(1) Duke vs. (5) St. John's| Model: Duke, 83.1%
Duke has been the most efficient team in the tournament, blowing out TCU by 23 after handling Siena. St. John's got here the hard way, with Dylan Darling's driving layup beating the buzzer against Kansas. The model sees a massive efficiency gap. St. John's has been the hottest team in the Big East since February, and 5-seeds are historically 25-75 against 1-seeds in the Sweet 16. The model takes Duke comfortably.
(2) UConn vs. (3) Michigan State | Model: UConn, 62.1%
Dan Hurley vs. Tom Izzo in March. The model doesn't care about coaching mystique. It sees UConn's offensive efficiency and depth as the differentiator. Michigan State ground through Louisville to get here, and UConn dispatched UCLA cleanly. This is one of the tighter model picks, and I'd call it the most likely upset of the round. Izzo in March is worth at least a few percentage points that no model captures.
South Region: Houston
(9) Iowa vs. (4) Nebraska | Model: Nebraska, ~63%
The model didn't predict this matchup. It had Florida and Vanderbilt here. Instead we get a Big Ten brawl between a 9-seed that just killed a 1-seed and a 4-seed that gutted out a two-point win over Vanderbilt. The model favors Nebraska's defensive efficiency, but Iowa has the tournament's best story and the confidence of a team playing with house money. Nebraska's the pick, but this one feels volatile.
(2) Houston vs. (3) Illinois | Model: Houston, 63.2%
The model has been high on Houston all tournament, ranking them above their seed line in pre-tournament championship odds. Houston's defense is suffocating, holding Texas A&M to 57 points. Illinois has been clinical, beating VCU by 21. This is the highest-quality Sweet 16 game on paper, two top-10 efficiency teams going at it. The model gives Houston the edge on the defensive side, and in March, defense travels.
West Region: San Jose
(11) Texas vs. (2) Purdue | Model: Purdue, ~78%
The model originally had Purdue facing Gonzaga here. Instead it's Texas, who went from the play-in game to the Sweet 16 by knocking off BYU and Gonzaga back to back. Purdue is a terrible matchup for them. The Boilermakers' size advantage is real, their rebounding metrics are elite, and Purdue has been one of the most dominant second-weekend teams in recent tournaments. Texas's Cinderella run has been remarkable, but the clock is about to strike midnight.
(1) Arizona vs. (4) Arkansas | Model: Arizona, 74.0%
This is the one Sweet 16 matchup the model predicted from the jump, and it still likes Arizona by a wide margin. Arizona's adjusted efficiency margin is second only to Duke's among remaining teams. Arkansas earned its spot, but Arizona has been the most complete team in the West all season. The model doesn't see that changing.
Midwest Region: Chicago
(1) Michigan vs. (4) Alabama | Model: Michigan, ~70%
The model had Michigan facing Texas Tech here. Alabama showed up instead, hanging 90 on them in a demolition job. The Wolverines have been the most dominant offensive team in the tournament, putting up 95 against Saint Louis. Michigan's efficiency metrics are clearly better, but Alabama is making their fourth consecutive Sweet 16 appearance. The Big Ten has a record six teams in the Sweet 16, three of them in this region alone. Michigan is the pick, but this is the game I'd most want to watch.
(2) Iowa State vs. (6) Tennessee | Model: Iowa State, ~61%
The model originally had Iowa State facing Virginia. Tennessee changed that plan, beating Virginia 79-72 to bust every remaining perfect bracket. Iowa State handled Kentucky after dispatching Tennessee State. Tennessee's physicality is the X-factor. They play an aggressive, disruptive style that can bother even the best-coached teams, and Iowa State's offense can go cold at the wrong time. This is the model's least confident Sweet 16 pick. Tennessee at 0.2% championship odds before the tournament? That number's climbing.
The Updated Championship Picture
After re-running 10,000 simulations with the actual Sweet 16 field:
- Duke: 28.1%
- Michigan: 21.3%
- Houston: 14.8%
- Arizona: 12.1%
- UConn: 6.2%
- Purdue: 5.9%
- Iowa State: 4.1%
- Illinois: 3.8%
Duke is still the model's pick to cut down the nets. Michigan is the clear second choice, with Houston lurking as the team nobody wants to draw. The Big Ten having six of the 16 remaining teams is historically unusual, and the model thinks at least three make the Elite Eight.
What I'm Watching
If you're filling out a Sweet 16 bracket, the model says Duke, Michigan, Houston, Arizona, Purdue, UConn, Nebraska, and Iowa State advance. That's a lot of chalk. If you want my gut pick for the upset? Michigan State over UConn. Izzo in March. Some things the gradient boosting classifier just can't quantify.
Through two rounds, the model has been respectably right and spectacularly wrong in roughly the proportions you'd expect from a 73.6% accuracy rate applied to single-elimination chaos. It missed every first-round upset and then called most of the Round of 32 correctly. It's a tool, not a crystal ball, and March keeps reminding us of the difference.
The full code, data, and live tracker are on GitHub.