Modeling March Madness with Deep Learning

It’s officially March, and I’ve been biting my nails and having palpitations about the Purdue Boilermakers since sometime in mid-January, so I decided to try to alleviate this anxiety and see if I can do some work to guess what will happen once the tournament starts.

I’ve also gotten kind of into sports betting, but I always lose and I’d love to change that with the power of computers.

What?

When experts and advanced stats people model sporting events, they look at a number of stats and work out well-defined models that predict how a team will perform. KenPom is one such popular advanced-stats model of team skill.

However, I’m not an advanced stats guy. In fact, I think the advanced stats people are going to eventually going to have to learn the bitter lesson – whatever convenient numbers you use to boil down a team’s skill are probably going to be suboptimal, and if you can capture the right raw data and assemble a good computational model that can process it, it will do a better job than you ever did.

How?

When folks talk about how good a basketball team is, they mention a few things: the offense, the defense, maybe perimeter defending ability, post offense, etc. Why do they talk about those things? Well, eventually they show up on the score board and in the win/loss column.

These numbers are really hard to quantify. One big reason for that is that, for example, a team’s offensive output in a particular game also depends a lot on the other team’s defensive ability; conversely, a team’s defensive performance on the stat sheet depends a lot on how good their opponent’s offense is.

But it’s even more complicated than that. Let’s take this year’s Purdue team as an example, since I’m very familiar with them. Purdue has notoriously struggled to beat the press. So, a team that runs a good press defense might have inflated defensive numbers against Purdue. But, Purdue also has one of the most dominant paint-players in the country, so it’s also important to consider the opponents post defense. To further complicate things, there’s probably some complicated interaction between running a press and defending the post. You should also take into account that opposing coaches know these facts about Purdue and will try to game plan with that in mind.

And remember, at the end of the day, we have to turn these facts into numbers that can go into some equation to predict the score of the game. What you end up with is a complex system with tons of variables, all of which are difficult or impossible to measure and interact in complex ways.

So, instead, we build a deep learning network that’s capable of producing some numbers about teams, which are in turn useful for predicting the final score of the game. We train it on every game that was played in the 2021-22 season up until ~the end of February, and then we see who wins in March.

Who wins?

First of all, we’re assuming this bracket is the final one. That’s not true, but it’s the best option I have right now.

There are also some issues, and the model isn’t perfect yet. There are some improvements, to be made, and it could probably do with some more training.

That being said, I’m putting Marquette and Arkansas on upset alert.

Round of 64

West Region

• Gonzaga 92 - 58 Alcorn State
• Colorado St. 73 - 67 Xavier
• Houston 80 - 63 South Dakota State
• UCLA 87 - 69 Toledo
• LSU 73 - 66 Michigan
• Texas Tech 66 - 57 New Mexico State
• Marquette 63 - 67 San Francisco
• Purdue 92 - 69 Montana State

East Region

• Kansas 94 - 76 Cleveland State
• Seton Hall 70 - 67 Wyoming
• Arkansas 60 - 61 North Texas
• Illinois 75 - 71 Northern Iowa
• Ohio State 79 - 67 Wake Forest
• Villanova 79 - Towson 60
• Michigan State 76 - Miami 67
• Auburn 97 - Colgate 76

South Region

• Baylor 84 - 60 Long Beach State
• Notre Dame 61 - 65 Murray State
• Alabama 69 - 68 San Diego State
• UConn 74 - 63 Vermont
• Saint Mary’s 58 - 60 Indiana
• Tennessee 85 - 68 Princeton
• Iowa 80 - 76 Davidson
• Duke 90 - 64 Longwood

Midwest Region

• Arizona 91 - 63 Norfolk State
• Boise State 55 - 59 TCU
• Texas 66 - 56 Iona
• Providence 74 - 68 Chattanooga
• USC 65 - 58 Creighton
• Wisconsin 67 - 58 Texas State
• Iowa State 76 - 68 North Carolina
• Kentucky 90 - 68 Jacksonville State

Round of 32

West

• Gonzaga 87 - 70 Colorado State
• Houston 63 - 59 UCLA
• LSU 66 - 60 Texas Tech
• San Francisco 66 - 77 Purdue

East

• Kansas 79 - 75 Seton Hall
• North Texas 57 - 54 Illinois
• Ohio State 71 - 70 Villanova
• Michigan State 69 - 72 Auburn

South

• Baylor 71 - 58 Murray State
• Alabama 73 - 76 UConn
• Indiana 63 - 62 Tennessee
• Iowa 70 - 80 Duke

Midwest

• Arizona 84 - 71 TCU
• Texas 55 - 53 Providence
• USC 62 - 61 Wisconsin
• Iowa State 60 - 64 Kentucky

Sweet Sixteen

• Gonzaga 82 - 72 Houston
• LSU 76 - 74 Purdue
• Kansas 70 - 66 North Texas
• Ohio State 72 - 71 Auburn
• Baylor 75 - 67 UConn
• Indiana 59 - 66 Duke
• Arizona 84 - 74 Texas
• USC 57 - 63 Kentucky

Elite Eight:

• Gonzaga 81 - 65 LSU
• Kansas 82 - 77 Ohio State
• Baylor 68 - 68 Duke
• Arizona 78 - 75 Kentucky

Final Four:

• Gonzaga 86 - 74 Kansas
• Baylor 76 - 72 Arizona

National Championship:

Gonzaga 84 - 73 Baylor