Zoltar is my NFL football prediction system. Zoltar uses a combination of classical statistics, a neural network, and a form of reinforcement learning.

My basic Zoltar predicts game results against Las Vegas point spread data. For example, if the Chargers are playing the Titans, the Vegas line might be “Chargers -8.5” meaning the Chargers must win by more than 8.5 points. Zoltar will predict which team will win and by how many points, such as “Chargers will win by 5 points”. This would result in a recommendation to bet on the underdog Titans.

I decided to enhance Zoltar to also predict the total number of points that will be scored in a game. This can be used to place over-under bets. For example, in the hypothetical Chargers vs. Titans game, the Vegas over-under line might be 49.5 points. You can bet that the total number of points scored by both teams will either be over 49.5 points or under 49.5 points.

The goal of my initial over-under analysis was to determine what would happen if you always bet “over” or aways bet “under” for the 2024 season. To do this, I needed Vegas over-under data and game results data. I got both sets of data from the nice pro-football-reference.com Web site.

I created a data file of Vegas over-under data that looks like:

# 1
09/5 8:20 ET ravens At chiefs 46.0
09/6 8:15 ET packers At eagles 49.5
09/8 1:00 ET steelers At falcons 43.0
09/8 1:00 ET cardinals At bills 46.0
. . .
# 21
01/26 3:00 ET commanders At eagles 47.0
01/26 6:30 ET bills At chiefs 49.5
# 22
02/9 6:30 ET chiefs At eagles 48.5

The last value on each line is the Vegas over-under. My working hypothesis was that always betting over, or always betting under, would not be a winning strategy.

To cut to the chase, the analysis results were:

Number games tot pts above Vegas over-under = 151
Number games tot pts below Vegas over-under = 130
Number games tot pts exact same as Vegas over-under = 4

The 2024 season had 272 regular season games, and 12 playoff games, and the Super Bowl game, for a total of 285 games. If you bet “over” on all 285 games, you’d have won 151/281 times = ~53% of the time.

This would not be a winning strategy because you must bet $110 for a chance to win $100. This means that, in theory, you must win 53% of the time to break even. In practice, to make money, you need to predict at about 60% accuracy (to take into account your overhead, and careless mistakes, and so on).

OK, now that I have satisfied myself that a naive “always bet over” or “always bet under” prediction strategy isn’t useful, my next steps will be to look at predicting the total number of points scored in a game using a combination of Zoltar’s algorithms and two machine learning regression techniques: kernel ridge regression and neural network regression. In a previous exploration I looked at linear regression to establish a baseline prediction result.

Las Vegas believes that sports betting holds the key to the future. I’ve been going to Las Vegas for many years and I’m always amazed at how often the town reinvents itself.

Right: The pool area at the Mirage was spectacular. I vividly remember eating breakfast at the poolside cafe one morning with my good friend Ed Koolish (who worked at Disneyland with me).