48 Games In, the Emotionless Bracket Just Got Emotional.

By Collin Lee | March 23, 2026

There's a moment in every bracket where you stop checking scores and start checking damage. For us, that moment was Saturday afternoon, when Iowa's Alvaro Folgueiras pulled up from three with 4.5 seconds left and ended our South bracket in a single arc of the ball.

Iowa 73, Florida 72. Our model had Florida at 91%.

I built this system to remove emotion from my bracket. I'd like to report that watching a 91% confidence pick die on a late three-pointer felt like a calm statistical observation. It did not. What it felt like was watching someone knock over the first domino in a chain you spent weeks building.

Florida was our South regional champion. Our Final Four pick. A load-bearing wall in a bracket that depends on structural integrity. And now it's gone - taken out by the same Iowa team our model correctly identified as an upset threat in the Round of 64. The model called Iowa over Clemson at 84% and was right. Then it called Florida over Iowa at 91% and was wrong. Our own upset pick came back to haunt us.

That irony is not lost on me.

10 out of 16 Sweet 16 teams match our original bracket. One pick was already dead on arrival - Wisconsin's R64 loss to High Point meant we were playing with 15 live picks, not 16. Of those 15, five more fell in the Round of 32.

But the number that matters: three of our four Final Four picks are still alive. The championship prediction - Michigan over Duke, 79-78 - is still breathing. Barely.

The full results

Here's every original bracket prediction for the Round of 32 against what actually happened. When a first-round upset changed the opponent, I've noted it. Strikethrough means we got it wrong. Bold green means we correctly called an upset.

* Predicted opponent was eliminated in R64; our pick faced a different team.

10 out of 16. Not as sharp as the R64, but the Round of 32 is where the bracket starts getting hard. Seed advantages narrow. Every game is between teams that just won. Here's the breakdown of our six misses:

• Florida at 91% - The painful one. More on this below.
• Gonzaga at 85% - A real miss. Texas's Cinderella run from the First Four is legit.
• Virginia at 73% - Tennessee's fourth consecutive Sweet 16. The Vols are built for March.
• Vanderbilt at 72% - Nebraska won on a layup with 2.2 seconds left. Their first Sweet 16 ever.
• Wisconsin at 62% - Already gone from the R64 High Point upset. A dead pick walking.
• Texas Tech at 51% - A coin flip. Alabama won by 25. The model shrugged and the scoreboard screamed.

What went right

The East is perfect

I need to talk about the East region, because it's doing something remarkable. Through two full rounds - eight games - our bracket has not missed a single pick in the East. Not one.

Duke handled TCU 81-58. St. John's edged Kansas 67-65 on a buzzer-beating layup by Dylan Darling, who had been scoreless the entire game - the model had this as a 73% upset pick and it came through on the most dramatic shot of the tournament so far. Michigan State dispatched Louisville 77-69. UConn rolled UCLA 73-57.

The Sweet 16 matchups in the East are exactly what our original bracket predicted: (1) Duke vs. (5) St. John's and (2) UConn vs. (3) Michigan State. Exactly. Not adjusted, not re-routed around upsets. The original pre-tournament bracket, published before a single game was played, nailed all eight East games and produced the exact Sweet 16 field.

If you had told me one region would go 8-for-8 through two rounds, I would have guessed it was a fluke. But the East is also the region where the model was most confident on average. It saw Duke and UConn as dominant, saw St. John's as underseeded, and rated the mid-seeds correctly. The model didn't get lucky in the East. It was right.

The worry list was wrong (in a good way)

Last week I published four games that "keep me up at night." Let me grade my own anxiety:

• Duke vs. TCU (97%) - I worried Duke looked vulnerable after a close R64 win over Siena. Duke won by 23. The model was right, my gut was wrong.
• Michigan vs. Saint Louis (97%) - I worried SLU's momentum from a 25-point R64 win made them dangerous. Michigan won 95-72. Momentum, it turns out, doesn't show up in the data because it doesn't show up in the results.
• Houston vs. Texas A&M (96%) - I called the Aggies "the most dangerous replacement opponent in the bracket." Houston won 88-57. A 31-point demolition. The model treated them like a routine mid-seed because they were one.
• Iowa State vs. Kentucky (88%) - I said this was "the Round of 32 game most likely to derail our Midwest bracket." Iowa State won 82-63. Not close.

Four games I was worried about. Four model wins. The model was right about every single one of them, most by comfortable margins. It turns out the things that keep you up at night aren't the things that actually go wrong. It's the ones you weren't watching. I wasn't worried about Florida.

What went wrong

The Florida problem

Let's talk about it.

Florida entered the Round of 32 as the defending national champion, a 1-seed, and a team that had just beaten Prairie View A&M 114-55 in the most dominant R64 performance of the tournament. Our model gave Florida a 91% chance of beating Iowa. The line wasn't close. The data wasn't ambiguous. Florida was, by every metric the model tracks, one of the best teams in the country facing a 9-seed that had barely beaten Clemson by six.

And then Alvaro Folgueiras hit a three with 4.5 seconds left. Iowa 73, Florida 72.

Here's what makes this one sting differently than the Wisconsin loss in R64. Wisconsin at 96% was our most confident miss, but Wisconsin wasn't carrying downstream weight. They were a 5-seed we had losing in the Sweet 16 anyway. Florida was a 1-seed. Our South regional champion. A Final Four pick. One of the four pillars holding up the entire bracket structure.

When Wisconsin lost, it cost us one branch. When Florida lost, it cost us the whole South side of the bracket - the regional champion, the Final Four semifinal, and potentially the championship game matchup. One three-pointer just killed at minimum three future picks. The cascade we've been writing about since the very first post is no longer theoretical. It's our bracket now.

A 91% probability means roughly 1 in 11. Over the course of a tournament, you'd expect one or two of these to flip. Getting it in the Round of 32 on a Final Four pick is the worst possible timing. The model would tell you that's just variance, that a 91% pick failing doesn't mean the model is broken, that this is what single-elimination tournaments do to even the strongest predictions.

The model would be right. And to be clear: there's nothing to fix here. The model's calibration is working - its 90%+ picks are hitting at better than 9 in 10 through two rounds. Changing the model to avoid 91% misses would mean distrusting every high-confidence pick, which would break the system that's gotten us 35 out of 48 correct. The right response to a 1-in-11 event is to acknowledge it happened and move on.

But watching it happen felt like getting punched in a dream - technically not real damage, but your body doesn't know that.

Texas's Cinderella run

Texas just became the sixth team in tournament history to go from the First Four to the Sweet 16. They beat NC State in the play-in, upset BYU in the R64, and now they've taken down 3-seed Gonzaga 74-68 on the strength of 17-point games from Jordan Pope and Matas Vokietaitis and a game-sealing corner three from Camden Heide.

Our model had Gonzaga at 85%. This wasn't a coin flip - the model genuinely liked Gonzaga. But 11-seed Texas, playing under first-year coach Sean Miller, is on a run that transcends the numbers. The model evaluates season-long metrics. It can't see that a team has found something in the last week that wasn't there in the first 30 games. That's not a bug - it's a structural limitation of any system that relies on historical data.

Texas in the Sweet 16 means our West bracket now has an 11-seed facing 2-seed Purdue instead of the 3-seed Gonzaga we predicted. In a weird way, this might actually help our bracket downstream - Purdue was always our pick to reach the Elite Eight, and Texas is a weaker opponent than Gonzaga would have been.

Nebraska makes history

Nebraska 74, Vanderbilt 72. Braden Frager drove to the basket and scored with 2.2 seconds left to send the Huskers to their first-ever Sweet 16. Before this tournament, Nebraska had been to the NCAA Tournament eight times and won a total of zero games. Now they've won two in a row and they're playing in the second weekend.

Our model had Vanderbilt at 72%. That's the kind of confidence level where you feel okay about the pick but you wouldn't bet your house on it. And that's roughly what happened - a two-point game decided in the final seconds. The model was on the wrong side of a razor-thin margin. It happens. But Nebraska being in the Sweet 16 means our South bracket is now completely off-script: instead of Florida vs. Vanderbilt, it's Iowa vs. Nebraska. A 9-seed against a 4-seed with zero March Madness pedigree.

The South region was supposed to be the boring one. Florida running through everyone. Instead, it's the most chaotic region in the tournament.

The dominoes keep falling

In the original post, we found that 76.5% of late-round bracket misses come from the wrong team being in the game, not from picking the wrong winner. In the R64 retrospective, we watched seven first-round misses start to ripple through the bracket.

Now the ripples have turned into waves.

Here's the damage report. Our original bracket predicted 16 teams in the Sweet 16. Here's which ones actually made it:

10 out of 16. The East is pristine. The West lost two (Wisconsin's cascade from R64, plus the Texas Cinderella). The South lost two (the Florida earthquake and the Vanderbilt/Nebraska flip). The Midwest lost two (Alabama over Texas Tech and Tennessee over Virginia).

And here's the cascade math: of our six wrong Sweet 16 picks, only one (Wisconsin) was a direct R64 cascade. The other five teams were all in their R32 games and lost. These aren't inherited failures from the first round - they're new misses. The Round of 32 fought back.

What's still alive

Our champion pick, Michigan, just steamrolled Saint Louis 95-72 with five starters in double figures. Yaxel Lendeborg had 25 points. Michigan is 33-3 on the season, matching a program record for wins. They look exactly like a team that's going to win six games in March. The model's most important prediction is in no danger.

Our runner-up pick, Duke, crushed TCU 81-58 after I spent last week worrying about their close R64 win over Siena. Whatever rust Duke had in the first game is gone. They're playing like the team the model sees.

Our West Final Four pick, Arizona, beat Utah State 78-66. Clean, professional, boring. The model's favorite kind of win.

Our South Final Four pick, Florida, is on a plane home. RIP.

That leaves us at 3/4 Final Four picks alive and the championship prediction intact. In ESPN scoring terms, the champion pick is worth 320 points, each Final Four pick is worth 160. We've lost one 160-point pick (Florida), but the 320-point pick (Michigan) and the 160-point runner-up (Duke) are both healthy. The bracket's highest-value bets survived.

What the model says about the Sweet 16

The bracket is set, and some matchups look different than what we predicted. Three of our eight Sweet 16 matchups are exactly what the original bracket had. Five changed because of upsets along the way. I fed the actual Sweet 16 field into the same model - same training data, same blend, same everything - and asked it what it sees. Here's the full path from here:

What changed

The championship prediction hasn't moved. Michigan over Duke, 79-78. Same score. Same one-point margin. Two rounds of upsets, twelve games that went differently than expected, an entire Final Four pick eliminated - and the model's view of the endgame is unchanged.

That's either the deepest conviction I've ever seen from a statistical model, or the most spectacular failure waiting to happen.

What did change:

• The South bracket is unrecognizable. Florida and Vanderbilt are both gone. The model sees Houston emerging from the South instead of Florida - which means Duke's Final Four opponent shifts from Florida (73% confidence in our original bracket) to Houston (89%). The model actually thinks Duke has a slightly easier path to the championship with Florida out. Cold.
• Purdue vs. Texas (95%) replaces Purdue vs. Gonzaga (75%). The Cinderella matchup is significantly more favorable for Purdue according to the model - a 20-percentage-point confidence jump.
• Michigan vs. Alabama (90%) replaces Michigan vs. Texas Tech (93%). Alabama is a tougher draw, but the model barely flinches - 90% is still overwhelming confidence.
• Nebraska vs. Iowa (75%) is the tournament's wildest Sweet 16 game - no top-3 seed in sight. The model favors Nebraska, but at 75% in a matchup between a 4-seed and a 9-seed who just killed a 1-seed, nothing feels safe. The South bracket has gone from "Florida walks through this" to "anything could happen."

The uncomfortable question

Here's what I keep thinking about.

After the Round of 64, I wrote: "The model would tell you not to get emotionally invested in that prediction. But between you and me, it's getting harder to take that advice."

It got harder.

When Folgueiras's three went through the net, the correct analytical response was: "91% means 1 in 11, this is within expected variance, the model's framework is intact." The actual response was me staring at my phone in a Trader Joe's parking lot, refreshing the score three times because I was sure it was wrong.

The system was designed to remove emotion from the bracket. It did. But it didn't remove emotion from me. And the further we get into the tournament - the more the bracket matters, the more each game carries ESPN points and downstream implications - the more the gap between what the model says and what I feel grows wider.

The model says Michigan 79, Duke 78. It said it before the tournament, after the R64, and again now after the R32. Three times, same answer. The model doesn't know that Florida is gone. It doesn't know I spent Saturday afternoon in a parking lot. It doesn't care that our bracket went from "boring and winning" to "cracking and stressful."

It just sees Michigan 79, Duke 78. And it's either going to be right or it isn't.

The Sweet 16 starts Thursday. Three of our eight matchups are exactly as predicted. The East is flawless. The champion pick is dominant. The runner-up is rolling.

I'm trying to remember that this is still, by any reasonable measure, a successful bracket. 35 out of 48 correct picks. Three of four Final Four teams alive. The championship game intact. If you told me before the tournament that we'd be here at this point, I'd have taken it in a heartbeat.

But you can't un-watch Florida lose. You can't un-feel that. And I think that might be the actual answer to the question this whole project set out to ask. Can you remove emotion from your bracket? Sure. The machine did it perfectly. But can you remove emotion from watching your bracket play out in real time?

Not even close.

What's next

The Sweet 16 starts Thursday. Here are the games I'm watching most closely:

• Duke vs. St. John's (83%) - Dylan Darling just hit a buzzer-beater to get here. St. John's is playing with house money. Duke is not. This is the game that tests whether the model's East region perfection was skill or a two-round heater.
• Nebraska vs. Iowa (75%) - A 4-seed that had never won a tournament game before this year against a 9-seed that just took down the defending national champion. No Florida. No Vanderbilt. The South bracket has gone completely off-script, and this game is the proof.
• Michigan vs. Alabama (90%) - Alabama just hit 19 threes and won by 25. The model is unmoved at 90%. Michigan is our championship pick. If Alabama shoots like that again, the model's confidence might not matter.

I'll be back with the Sweet 16 retrospective once the second weekend wraps. The bracket is cracked but standing. The championship prediction hasn't budged. And I'm still checking scores in parking lots.

Follow along. The machine still doesn't care. I still do.

Update: The Sweet 16 is done. Half the bracket is dead but the prediction won't budge. Read what happened.