TL;DR - For the last 10 years, I've built a bracket optimization model to maximize my odds of winning March Madness pools. For the last few years, I've made it available to the fine folks here at MGoBlog. Email me at [email protected] or drop a comment below if you are interested in using it.

 

Happy Madness. I'm excited to make my model available once again, while also adding a few more bells and whistles. As a refresher, here are links to past years' diaries:

• Part One - A deep dive into how the model works.
• Part Two - A brief summary of how the model works, plus improvements made to last year's edition.

As a post-mortem of how the model fared for me last year, my nine brackets (yes, nine - I'm not really a gambler but investing as much time as I have into this I want to put the model to sufficient use) finished in the following percentiles of brackets submitted to ESPN: 99.3%, 97.2%, 94.7%, 91.1%, 88.2%, 85.2%, 81%, 81%, 78.4%.

Four of nine in the top 90%, and eight of nine in the top 80%. Not too shabby. But last year, I'm on the record for saying the four teams I really liked as favorable picks were Houston, UCLA, Tennessee, and UConn. UConn ended up being a great call; however, you may remember that Houston, UCLA, and Tennessee were all teams that suffered significant injuries leading into the Big Dance.

Thus, the inspiration for one of this year's additions to the model. This is very much still experimental, subjective, and more art than science; but I've attempted to quantify the overall quality of teams' stars and rotations as well as the total impact on their rosters due to injuries using offensive and defensive Porpagatu! - an advanced points-over-replacement statistic developed by Bart Torvik (and used by Seth in basketball preview lineup cards). The math I use is pretty rudimentary and relies on only three categories (probable, questionable, out) for player status, but it does yield some interesting results. An excerpt from last year's tournament:

The "Star" rating takes the average combined offensive and defensive Porpagatu! for a team's top three players, while the "Rotation" rating applies the same statistics for a team's top seven players. Last year, UCLA's injuries (Jaylen Clark - out; Adem Bona - questionable) heading into the tournament were the difference between both its Star and Rotation ratings being alone in first place (UCLA's star rating still remained in first on the backs of Jaime Jacquez Jr, Tyger Campbell, and David Singleton; but their rotation rating dropped from first to 25th among tournament teams). Houston (with Marcus Sasser even listed as probable) and Tennessee (without Zakai Zeigler) heading into the tournament also had very material impacts on their roster ratings, and these adjustments due to injury may have warranted heavier fading of those teams.

Retroactively applying these ratings to last year's tournament results:

• Teams with the greater Star rating won 61.9% of games.
• Teams with the greater Rotation rating won 69.8% of games.
• Teams with both the greater Star and Rotation ratings won 68.5% of games.

As a comparison, the favorites (according to the Kenpom/Torvik/BPI ensemble model) won 69.8% of games last year. Small sample size and all that, but there may be some justification to sprinkle in some subjective analysis in this year's picks. This year's key injuries (Marquette's Tyler Kolek, Kansas' Hunter Dickinson and Kevin McCullar, Duke's Caleb Foster, Baylor's Langston Love) appear to be a little less catastrophic than 2023's injuries; but it will be interesting to see how those teams fare relative to expectations based on the advanced team rating systems.

 

The next addition to the model is likely only relevant to those who plan to make multiple brackets and use the "Pools" tab of the model. I am now quantifying the relative allocations I have to each team based on actual selections that are made and recorded in the "Pools" tab. The higher the number, the higher the relative allocation is and the more susceptible my brackets are to a particular team losing (i.e. a "bracket buster").

This serves as a tool for measuring when it may be necessary to adjust picks and perhaps sprinkle in some diversification. I use two statistics:

• Probability Quotient: the frequency or percentage of time that a team is selected across my bracket entries / that team's probability to reach the corresponding round.
• E[PAP] Quotient: the frequency or percentage of time that a team is selected across my bracket entries / that team's E[PAP] for the round compared to all alternatives for the round.

Below is an excerpt of my picks from last year. While I did mull over plenty of high-leverage picks for my Champion, Final 4, Elite 8, etc.; I do realize now that there were several instances where I didn't vary my picks so much as I maybe should have. A few outliers from last year (combining both the probability and E[PAP] quotients together):

1. Marquette to the Elite 8
2. Arizona to the Elite 8
3. Texas to the Elite 8
4. Kansas to the Sweet 16
5. Alabama to the Final 4
6. Indiana to the Sweet 16

Obviously, it's easy in retrospect to have buyer's remorse over my selections from last year's tournament, but it is interesting that five of the six scenarios in which I had the highest relative allocations did not happen. I may have been well-justified in adding more Kansas State's, Creighton's, San Diego State's, and Miami FL's. This methodology should help serve as a way to quantify when/if diversification may be necessary before finalizing selections.

That about does it for this year's new and improved edition of the model, and I'm now diligently working on incorporating the new data for this year's tournament. As always, shoot me an email or leave a comment if you are interested.

Best of luck with your pickings!

-mgoDAB