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 Eagles are playing the Falcons, the Vegas line might be “Eagles -9.5” meaning the Eagles must win by more than 9.5 points. The basic Zoltar will predict which team will win and by how many points, such as “Eagles will win by 6 points”. This would result in a recommendation to bet on the underdog Falcons.

I modified Zoltar to predict the total number of points that will be scored in a game. I used a neural network. These predictions can be used to place over-under bets. For example, in the hypothetical Eagles vs. Falcons game, the Vegas over-under line might be 48.5 points. You can bet that the total number of points scored by both teams will either be over 48.5 points or under 48.5 points. (The .5 prevents pushes).

To prepare for the upcoming 2025 season, I looked at data from the 2024 season. In my initial stab at predicting the total number of points, my analysis was simple in the sense that if the predicted number of total points is greater or less than the Vegas over-under line by any amount, then a hypothetical bet would be placed. With that simple approach, Zoltar scored just under 65% accuracy:

Number correct vs Vegas OU = 181
Number wrong vs Vegas OU = 98
Number no-bet (pred equal Vegas OU) = 0
Number no-bet (neutral site) = 6
Accuracy vs Vegas OU = 0.6487

I decided to see what happens when adding thresholds. Suppose the Vegas OU for a game is 50.0 points, and Zoltar predicts the total number of points scored will be 51 points. If the over threshold is set to 2.0 points, then no hypothetical bet would be placed because the predicted total points isn’t large enough. If the predicted total points is greater than 52, then a hypothetical bet would be placed. The same logic is used for under bets.

After quite a bit of experimentation, good threshold values seem to be 4.0 points for over (in other words, for a bet to be placed, the predicted total points must be greater than 4.0 points more than the Vegas over-under) and 1.0 points for under. The results show 73% accuracy:

Over threshold = 4.0
Under threshold = 1.0

Number correct vs Vegas OU = 138
Number wrong vs Vegas OU = 51

Number correct (over) = 86
Number correct (under) = 52
Number wrong (over) = 28
Number wrong (under) = 23

Number no-bet (pred equal Vegas OU) = 0
Number no-bet (neutral site) = 6
Number no-bet (over) = 72
Number no-bet (under) = 15

Accuracy vs Vegas OU = 0.7302

These are quite impressive results. The big unknown at this time is whether or not the prediction model, based on 2024 data, will generalize to the 2025 season. We’ll see.

Predicting the total number of points that will be scored in an NFL football game is a difficult problem. But some things in my life are easy to predict.

Right: Whenever I walk my two dogs, as soon as we get back home, Riley, will go lay in her bed and take a nap. Here, Riley is missing a TV commercial that shows a Cavalier Spaniel.