Michigan Won. The Model Was Right. I Just Didn't Expect to Care This Much.

By Collin Lee | April 7, 2026

At halftime of the national championship game, Yaxel Lendeborg looked into a sideline camera and said, "I feel awful. I feel super weak right now. I can't make anything."

He was 1-for-5 from the field. 0-for-3 from three. The knee that had been the only question left in this entire series - the MCL sprain that sent him to a stationary bike during the Arizona game, the one the model couldn't see, the one I wrote about yesterday with the words "the model says 98%, the knee says we'll see" - was answering louder than any algorithm.

Michigan led 33-29. The model had predicted this game at 98%. It had predicted an 83-69 blowout. It did not feel like 83-69. It felt like a team clinging to a four-point lead with their best player telling the world he couldn't make anything.

Then the second half started. Elliot Cadeau hit Michigan's first three-pointer of the entire game - their first make in 15 attempts - and pushed the lead to 11. Lendeborg, the player who felt awful, who felt super weak, who told a national TV audience he'd played soft, scored six consecutive points around the four-minute mark to keep UConn at arm's length. And Trey McKenney, with 1:50 left and UConn still fighting, buried a three from the wing to make it 65-56. The dagger.

Michigan went to the free throw line and sealed it. 25-of-28 from the stripe on the night. Cadeau. McKenney. Lendeborg, who hit all five of his. The model predicted Michigan would win on offensive efficiency. Michigan won on free throws, defense, and a hurt player who refused to quit.

Michigan 69, UConn 63.

First national championship since 1989. Thirty-seven years. The Big Ten's first men's basketball title since Michigan State in 2000. Seventy thousand, seven hundred and twenty people in Lucas Oil Stadium. And somewhere in the confetti, an XGBoost model trained on 54,000 games that has no idea it was right.

The model picked Michigan to win the national championship. Michigan won the national championship. Forty-four of sixty-three games correct. The champion, the Final Four semifinal matchup, two of four Elite Eight teams, five of eight Sweet 16 teams. By any statistical measure, the bracket worked.

But the model predicted Michigan 83, UConn 69 - a 14-point cruise. The actual game was a 6-point grind where Michigan shot 38.2% from the field, went 2-for-15 from three, and won because they made free throws and played the best defense of the tournament. The model saw offensive dominance. Reality delivered defensive desperation. Right about who. Wrong about how.

What the model couldn't see

The model couldn't see Lendeborg's knee. It couldn't see him playing 36 minutes on an MCL sprain, going 4-for-13 from the field, missing every three-pointer he took, scoring 13 points on sheer will, and still refusing to come out. It couldn't see the halftime confession - "I played really soft" - or the second-half push where he scored six straight to keep UConn from getting closer than six.

The model couldn't see Elliot Cadeau. The numbers said he was a solid point guard with good assist-to-turnover ratios. They didn't say he was a former UNC transfer who'd been written off as a bust, who transferred to Michigan and found a home, and who would score 19 points and win Most Outstanding Player of the Final Four. The model measured Cadeau's efficiency. It missed that he had something to prove.

The model couldn't see that Michigan would win a championship shooting 13.3% from three - the worst three-point percentage by a title winner in the shot-clock era. It predicted a team that would outscore opponents. Michigan out-defended opponents instead. They held every Final Four weekend opponent to a season-low shooting percentage. They scored 61 of their 69 points in the paint or at the free throw line. The model was built on efficiency ratings. Michigan won on effort ratings.

And the model couldn't see UConn. Not this UConn. Not the program that had been 6-0 in championship games and was going for its third title in four years. Dan Hurley, who cried walking through the hotel after the loss and then told reporters the game was called beautifully. The model saw a 2-seed with good numbers. It missed a dynasty trying to defend its legacy. It gave UConn a 2% chance. UConn played like a team that didn't know it only had 2%.

What happens when you remove emotion from your bracket

Seven posts ago, I wrote: "I wanted to know what happens when you take all of that out. No gut feelings. No revenge narratives. No 'they just look like a Final Four team.' Just data, algorithms, and a system that doesn't care whether Duke is wearing cool shoes this year."

So what happens? You get a better bracket.

Forty-four of sixty-three. The champion, correct. 1290 ESPN points. The model didn't just compete with gut-feel picking - it beat it. Not by guessing better in the late rounds, but by doing the thing I wrote about in the first post: protecting downstream picks. Every correct first-round call kept five later picks alive. The 48% upset threshold meant it only took underdogs when the numbers actually supported it. The pure XGBoost switch for the Final Four caught patterns that a simpler model - or a human going with their gut - would miss. Those three mechanisms, boring as they sound, are why it nailed 25 of 32 in the first round and had the champion standing at the end.

The bracket was the point. The bracket worked.

What I didn't expect was what happened to me while watching it work. I pumped my fist for math when Iowa beat Clemson and the model's upset call landed. I stared at my phone in a Trader Joe's parking lot when Florida went down - devastated, refreshing the score three times because I was sure it was wrong. I spent three weeks rooting for Duke - Duke! - because the bracket needed Duke to survive. I watched Lendeborg limp to the locker room during the Arizona game and felt genuine dread, not because a player was hurt but because a prediction was threatened.

Removing emotion from the picks didn't remove emotion from the experience. It concentrated it. Every game mattered - not because these were my teams, but because they were the machine's teams, and defending the machine's choices made every result feel personal in a way that "I like this team's uniforms" never did.

The final number

The model predicted Michigan over Duke, 79-78. The actual result was Michigan over UConn, 69-63. Wrong opponent. Wrong score. Different style of game entirely.

But here's what the model got exactly right: Michigan was the best team in the country. Not because of one game or one player. Because across six tournament games, against increasingly desperate opponents, with their star player hobbling through the last two, Michigan found a way every single time. They won by 21, by 23, by 13, by 33, by 18, and by 6. They won running. They won grinding. They won on a bad knee.

The model saw that in the numbers before the tournament started. It saw Michigan's defensive efficiency, their depth, their rebounding, their consistency. It processed 39 features and said: this team. The model didn't know Lendeborg would get hurt, didn't know Cadeau would have his redemption game, didn't know the championship would be won at the free throw line. It just knew the underlying structure was there.

Forty-four of sixty-three. The champion, correct. The model did its job. It just had no idea what its job meant.

Michigan's last championship was in 1989. I was born in 1999. I've never seen a world where Michigan was the reigning champion of college basketball. Neither had most of the people in Lucas Oil Stadium last night. Thirty-seven years is a long time to wait for something. The model didn't know it was making anyone wait. It didn't know 70,720 people would watch. It didn't know Dusty May would say "it still doesn't feel real." It didn't know Cadeau would cry.

The model ran the numbers. The numbers said Michigan. And for one night in Indianapolis, that was enough.

I set out to build a bracket with no feelings. The bracket worked - 44 games, the champion, the only question that matters in March. And it turned out to be the bracket I've cared about most.

The machine was right. It just didn't feel a thing.