The NFL model is successful enough to preclude me from detailing the model’s specifics, but it relies on both quantitative and third-party, publicly available opinion. I feel as though I am still developing this model, and it has some characteristics that give me some pause when contemplating wagering very substantive amounts of money (though I do bet on it) :

1. The model seems to much more effective when it favors the home team
2. It performs better in the earlier part of the season than the later, as if the lines improve over the course of the season
3. While the model is 23-8 overall, it is like 9-8 in the weeks that I have chosen to wager sizeable amounts of money.

In addition to these restrictions, the act of betting large amounts of money in Reno and Las Vegas (Nevada is the only state to allow sports wagering), is unsettling.  Carrying multiple thousands of dollars around on their person may feel exciting to some, but it just makes me paranoid. I used a friend's online account for about a season, but he got spooked by the federal crackdown on online gambling, so he closed it down and I share some of his concerns, so have not opened up my own.

Model Mechanics
Basically I create variables for the quantitative and subjective inputs to my model and use linear regression against the game results.  This allows me to create a predicted “net score” and compare this against the line.

             

My data (the left chart) may not look like much of an edge vs. the line (actually I have a slightly worse r-squared: .185 vs .181; it is hard to read), but I noticed that my winning percentage increased the greater the discrepancy of my predicted score vs. the line. So in the universe of games where my predicted score varied from the line by greater than 5 points, my performace, relative to the line, improved, though both r-squared got worse but my performance improved relative to the line (r-squared: .111 vs. .042).

     

As I dug further, I noticed that my model seems to do much, much better when it favors the home team vs. the visiting team. Part of my "digging" was getting burned on some sizeable bets. Why this is so is a bit of mystery. If anything, I thought that my advantage vs. the line would favor the visiting team with my hypothesis being that more money would be bet on the home team than the visiting team, so the line would adjust to give more points to the visiting team. I actually do not know if more money is bet on the home team or not, but I still don't know why my model works this way, but the results are striking:

          

The charts show the winning percentage (left vertical axis) against the "line gap" (the absolute difference between my predicted score and the line). The right vertical axis (the blue line) is the number of games that fit exist at a various line gap points. So, as the charts show, my model does very well when I predict that the home team will cover the line vs. when my model predicts the visiting team will win. Further, the greater the gap between my predicted score and the line, the better my chances of winning, at least with home teams. These charts are based upon available games (I did not grack all weeks in 2005, starting in week 6), weeks 2-14. The model coefficients are based upon 2004 data.

I looked into adapting this to college, but lopsided nature of many of the games created an obstacle. I am curious if I set some boundaries, to isolate games that similarly matched as professional games and see if I can get similar results.

Next Steps:

1. Try, again, to adapt the model to college.
2. Track and bet upon the 2007 season.