Recapping the men's 2026 tournament and evaluating our models.
Previously: 2025
Hello again everyone! As I said in my very first post on this site, my motivation for personal projects ebbs and flows, and after the grind of watching March Madness, writing about March Madness, and developing my own website/blog for the very first time, all while working a full-time job, let’s just say there was a serious ebb in motivation.
However, with the season tipping off in a little over a week, now is as good a time as any to hop back on my horse and talk some basketball.
Since my last post:
Suffice to say a lot has happened. However, since that’s not what this post is about, and many of these topics have already been discussed ad nauseam by people far more accomplished than me, I’ll constrain myself to a short response to each:
Now, back to the matter at hand.
For those of you who are new here, I highly recommend reading last year’s post, as it goes into a lot of detail on how each of the models work, the philosophy behind each, and what each of the evaluation metrics mean, but I’ll give a brief recap here:
My name is Noah; I code a bunch of different machine learning models to make predictions about the March Madness tournaments, then I evaluate them and compare them to alternative methods and established predictive models using a host of evaluation metrics.
Those models are:
* This separates it from other resume metrics like strength of record or wins above bubble, since those metrics still evaluate the quality of the win based on advanced efficiency metrics, which typically account for margin of victory.
We compare those models to the “chalk” method (always picking the higher seed) and Torvik, a well-respected advanced analytics model coded by Bart Torvik, available at barttorvik.com. I’d also like to take this moment to thank Bart Torvik for making all of his tools on his website free and accessible. I heavily rely on his meticulously tracked data for every Division I college basketball game. This project would not have been possible without his website.
I evaluate these models using the following metrics:
For various reasons covered in more detail in the 2025 post, not every model can be evaluated by every metric.
Finally, a couple of notes for those of you who have been following since last year:
Firstly, I had to part ways with MNPI. It was my first model, so I was sad to see it go, but it was reliant on a data set that was much less user-friendly than Torvik, and while it was free to use when I first started modeling, has since begun to charge for access. I could theoretically recreate something like MNPI using the Torvik stats, but MNPI was a linear regression algorithm, which really isn’t as well suited to game predictions as these other models are.
Secondly, as I explain in this post, I made some mistakes when programming Minnow and Branchy Brackets the first time around that I only caught mid-tourney last year. I’ve since fixed the issue, but as a result, some of the scores were altered when I re-evaluated 2025 model performance. If you’re seeing a discrepancy between what I reference and the scores from the 2025 post, that’s why.
Last year, Branchy Brackets dominated in just about every category, with Resumetric close behind. I theorized that this was because the models were overindexing on the favorites in one of the chalkiest tournaments of all time. Let’s see how it goes this year.
And now, without further ado:
| Rank | Model | Accuracy |
|---|---|---|
| 1 | Torvik | 79.37% |
| 2 | Branchy Brackets | 77.78% |
| T3 | Composite | 74.60% |
| T3 | Minnow | 74.60% |
| 5 | Chalk | 73.02% |
| 6 | Resumetric | 71.43% |
| Rank | Model | Log Loss |
|---|---|---|
| 1 | Torvik | 0.4166 |
| 2 | Composite | 0.4318 |
| 3 | Branchy Brackets | 0.4324 |
| 4 | Minnow | 0.4439 |
| 5 | Resumetric | 0.4550 |
| 6 | Chalk | 2.4854 |
(For context, 0.693 is the log loss of a random classifier)
| Rank | Model | Points |
|---|---|---|
| 1 | Torvik | 990 |
| 2 | Minnow | 930 |
| 3 | Branchy Brackets | 890 |
| 4 | Chalk | 840 |
| 5 | Composite | 800 |
| 6 | Resumetric | 780 |
| Rank | Model | Expected Points |
|---|---|---|
| 1 | Resumetric | 834.71 |
| 2 | Composite | 818.21 |
| 3 | Branchy Brackets | 807.80 |
| 4 | Minnow | 789.10 |
This year Torvik beat out my models, though they still performed well overall. Torvik was the top model in all three metrics it was eligible for. This is not a surprise, given how well respected Torvik is in the field, and I’d expect over a large sample his model would rate better than mine (though over the two-year sample, Branchy Brackets is still in the lead).
Branchy Brackets was my best-performing model on average in these categories, and only trailed the composite slightly in log loss, which I’ve previously staked as the most informative metric.
This year’s tournament was pretty chalky, though not approaching the levels of chalkiness we saw in 2025, and Branchy Brackets tends to like favorites, so that could be one reason it performed well.
That being said, the top seeds the past few years, by all advanced metrics and the eye test, have been historically dominant squads, so maybe Branchy is rightly high on great teams.
The best way to tell will ultimately be to sample from many more tournaments than just the past few years to get a variety of different types of years. That way, we’ll be able to see if Branchy will get burned by being high on favorites, or if Branchy is more discerning with its confidence than we currently realize.
That project is currently in the works, so stay tuned to the blog to find out.
I think my biggest letdown from this tournament has to be Resumetric. Not because it was terrible, per se, but because to me, in real time, it felt like Resumetric was performing a lot better than the evaluation metrics are currently bearing out.
There is an interesting discrepancy with Resumetric, though. Despite placing last in the first three metrics,* Resumetric came in first for expected tourney points.
* Yes, technically chalk is last in log loss, but you can bank on that as a given in any tournament. The whole point of putting chalk through that metric is to demonstrate that giving close to 100% confidence for every outcome is going to be severely punished by a qualified evaluation metric.
Big points are available in the late rounds of a bracket, so when looking at swings in that metric, that’s always a good place to start. With that as our guide, I think the biggest contributor is that Resumetric was much higher on UConn and Michigan than the other metrics.
Consider this:
| Branchy | Minnow | Resumetric |
|---|---|---|
| 5% | 4% | 8% |
| Branchy | Minnow | Resumetric |
|---|---|---|
| 10% | 19% | 18% |
Basically all year long, the advanced metrics liked UConn a lot less than one would expect for a team that secured wins over teams such as Illinois, Florida, BYU, and Texas in the non-conference season, due to the close nature of many of their wins over inferior opponents. However, as I’ve mentioned, Resumetric doesn’t care about margin of victory. It’s solely concerned with your resume, and UConn’s resume was good enough for the committee to award it a 2-seed, despite those relatively lackluster advanced metrics, which resulted in it being higher on UConn than the other two metrics, which behave more like traditional advanced analytics.
If we isolate the expected points from UConn for each model we get:
| Branchy | Minnow | Resumetric |
|---|---|---|
| 65.5 | 51.3 | 70.1 |
That 19-point gap between Resumetric and Minnow for UConn alone accounts for over 40% of their gap on the final scoreboard.
Meanwhile, Branchy was in hindsight almost comically low on Michigan as compared to the other metrics, and that hurt it in a spot where picking the correct champion is worth 320 points.
Isolating the expected points for Michigan, we get:
| Branchy | Minnow | Resumetric |
|---|---|---|
| 171.2 | 210.9 | 220.4 |
Absolutely devastating for Branchy. Clearly Branchy paid dearly for installing Duke as a 42% favorite to win the whole thing at the start of the tournament, only to see them miss the Final Four entirely.
I think that’s why in real time Resumetric “felt” better. In the long run, it hedged its bets, not overindexing on any one particular team, which let it cash in more points in the later rounds of most simulations.
I think that’s all I have to say for this post-mortem, but I have a lot of articles and new tools planned over the course of the season, including the longer historical analysis I mentioned above, going as far back as Torvik will let me.
I also plan to publish individual team ratings and make the matchup explorer that I introduced during last year’s tournament available during the regular season (updated weekly), once we have enough games under our belt to rack up some meaningful results (so probably sometime in early December).
If you want to make sure you don’t miss any of that, good news! I just finished setting up an email subscription service. You can sign up for free, and I guarantee, the only thing I’ll be sending to your inbox will be to notify you when I post a new article or create a new tool.
You can subscribe by clicking here, or by using the buttons at the top and bottom of the page.
Anything that I missed in my analysis that you want to talk about, or have questions for me? More good news! I also just created a dedicated email for this site. Send any comments, feedback, or questions to noah@nokubrackets.com. I’d be happy to chat with you.