The NerdBracket Model

I can honestly say that no one on the planet has put more time into March Madness predictions. I work on the model every day. My wife looks at me sometimes and shakes her head, especially at 11:30 at night. "Aren't you coming to bed?" she asks. "Do you want to spend time with me or the computer?" 

I have been working on a March Madness computer model for the last 10 years. The early models were pretty basic and focused on a few stats like 3 point percentage and team experience. In most years, an untrained chimpanzee could have fared better. 

I built a database of 25 years of March Madness data and began refining the model. I have tried predicting the outcomes with offensive efficiency, defensive efficiency, coaching experience, geographical data, record, conference, probability and a number of other factors. 

I conversed with some experts who are university analytics professors. They said that you can only get so far with your predictions and it depends on the year. Some years are relatively upset free and those years allow for more accurate predictions. 2019 was one of those years. That was the year Greg Nigl predicted the first 49 games of the tournament. 

In 2023, I used a very simple method to get to an 890/1920 score in the ESPN Bracket Challenge. That was a good result, but I wanted to keep working to find better solutions.

I talked to some university analytics experts to get a handle on models they had built. They said that March Madness computer models could fare well on "good years" but there was too much impredictability in the tournament that could not be explained by statistics.

I was discouraged after talking to the university experts. I wanted to find a model that would hold up year after year, not just on "good" years. I pressed on.

Just after the 2023 tournament ended, I came up with another type of model. This model uses an iterative approach with a decision tree that determines if the bracket is a viable bracket based on past year's results. This model seemed to be much more effective than the simple model from 2023. It will output thousands of brackets that will have a small chance at getting close to a perfect bracket. From these thousands of brackets, we can construct an overall bracket that would win most bracket pools, what I call the NerdBracket "Consensus Bracket." 

The 2024 model did not fare as well as I intended. One of the first operations before running the model is to select a most likeky champion. In my 2024 preselection process, I went all in with Houston as the champion. This ended up being a bad choice, as Houston was eliminated after the Sweet Sixteen, with an injury to their point guard really hitting the team hard.

I have of course been working hard to refine the March Madness computer model even further. It is a year long, never ending job that I love doing! We have a new addition to the March Madness Picks - the new March Madness Guide! It will give you a detail view of the viable brackets that the model chooses.

I have been collaborating with other March Madness model creators and this has broadened my knowledge even more. 2025 is going to be great! I can feel it.

Be on the NerdBracket team! Contact us for news and updates! 

Todd Beach

Tuscaloosa, AL