I've been doing some research and regressions and have come up with a few formulas that look promising.
The first is winning percentage which has an r-squared of .77 - the inputs for this are offensive big play % (runs of 20+, receptions of 30+, KR of 40+, PR of 20+), defensive big play %, turnover margin per game, offensive yards per play (including KR, PR), defensive yards per play.
Next is winning percentage to winning margin which has an r-squared of .84 - the only input is winning percentage.
Last is a simple points predictor which has an r-squared of .85. The only inputs for this are offensive yards per play and turnover margin per game.
All of these correlations work using 2010 and 2011 stats. The formulas have been scaled to maximize results using 2011 stats, but the r-squared difference is negligible for each year.
Right now, I'm using a spreadsheet to compare team A's stats to their opponent's and then dividing by overall average. ex: (Ox*Dx)/Ax where Ox is offense, Dx is defense and Ax is the average. This gives me the projections for each game in each category I need (thank god for excel here, there are something like 50 different calculations to come up with those numbers).
Once I've calculated a winning percentage, I calculate a margin which I compare to the actual line. A five point difference is a 1 unit play, 10 points is 2, etc. I used this during the last couple weeks and the bowl season and ended up at just 55%, but up 16.2 units for an ROI of 17%. One unit plays were 7-15 (-9.5 net), two unit plays were 8-7 (.6), three unit plays were 6-1 (14.7), and four unit plays were 2-0 (8). I also experimented with O/U plays based on the points correlation and went 7-2 (2.4 - .5 unit plays) so I'm thinking there may be something there.
My question to you all is, is there a better way to use these numbers? And, considering the correlations, do you think this will prove to be profitable?
Thanks for reading.
I've been doing a lot of things very similar to this for a couple of years - I'm a stats nerd (actuarial science major). In my opinion, there is way too much variance in college football to produce any type of consistency. Combine the severe differences in SOS with inflated HFA and emotional factors and it's difficult to produce anything significantly and consistently different from noise. Just for an example of simply the HFA factor, which is most likely the least significant statistical marker with regard to SOS and certain emotional factors, the HFA for Hawaii has been calculated (for a span of years between 2000-2010) to be 12.62 +/- 2.9, 4.9 +/- 2.7 for Notre Dame, .6 +/- 2.8 for LSU, -5.5 +/- 3 for Navy, just to name a few. To accurately account for a HFA margin between two teams consistently over the course of a single season, let alone individual games, would be statistically impossible. Combine that to the far more significant individual SOS variable (in addition to numerous others including injury, emotional factors, travel, rivalries, etc) and in my opinion, you could combine Einstein's creativity and critical hypothetical reasoning, Gauss's mathematical and statistical genius, a pool of the combined football knowledge of Vince Lombardi, Jon Gruden, Peyton Manning, Bill Bellicheck, and Dick LeBeau - then incorporate all of those ideas into as many Monte Carlo simulations as you could produce or as many applied statistical programs as you wanted, and you still couldn't produce anything near a statistically significant spread beating formula for college football. Now that's just based on statistics alone; one could argue that it may be possible to limit the amount of variables, constrict the samples to only power conference games in certain situations, and introduce a subjective basis or knowledge of a percentage of previous occurrences (somewhat similar to meteorological projections of weather if you're familiar with that) and produce something of interest. But then with that, weather is continuous with a historical basis that's (more or less, at least in recent applicable years) unchanging. When you introduce the fact that players, coaches, teams, facilities, philosophies, and schedules change regularly, the entire idea becomes somewhat frivolous. It could most aptly be compared to predicting the weather, based solely on weather patterns and data that had been collected over a time span much shorter than a year, even half a year. And we all know how accurate weathermen are, and that's with years and years of reliable and "unchanging", objective data and analysis.
However, I think that the NFL provides a much better basis for creating a system or program. The differences in SOS are reduced ten fold; there are four times as many teams in the FBS than there are in the NFL, and in the NFL the lowest ranked team (based on say, a set of power rankings) can beat the highest ranked team on any given day - I don't think the same could be said about a team like UMass beating a team like Alabama. The NFL is comprised of professionals who work and practice every day for a paycheck (not a student with no chance at the NFL who's girlfriend just broke up with him, or who failed an Accounting exam on Thursday, or who just lost to a mid-major last week and now has no shot at a decent bowl or maybe even a bowl period and thus has nothing to play for). HFA is minimized as travel is consistent (private jets and luxury hotel accommodations), facilities and locker rooms are all top of line, and referees are "consistent" professionals who are less prone to be "persuaded" by thousands of home team fans as their impartial performance greatly improves their reappointment probability (possibly less so this year lol). HFA has been statistically calculated in the NFL to be little more than a field goal for the most intense, conference rivalry games, and less than a field goal in almost all other situations - with no team having a much larger significant statistical advantage than any other. There is a minimal difference in skill between the majority of players in the NFL at their respective positions. Certainly the Pro Bowl players at each position are more skilled than the UDFA signed to the practice squad who made the team due to an injury. But the difference between that player and the 10th best is very minimal, with that difference decreasing as the younger player gains experiences and the older player ages. The skill level of players are constantly converging to a central "zone", if you will, which provides a great basis for leveling out SOS and other factors. For example, a quarterback playing the Lions has a similar challenge as a quarterback playing the Bengals. Whereas in college football, a quarterback playing LSU is certainly facing more of a challenge than a quarterback facing Kansas. Trying to compare stats from teams (even teams in the same conference) becomes a huge challenge because the strengths of each respective defense played is so significantly different, a challenge which is minimized in the NFL. I'd dare to say that the difference in strength of defenses/offenses played among teams in the NFL is statistically eliminated (or at the very least relatively close to being statistically insignificant) throughout the course of a single season.
The only hindrance with the NFL is sample size. The total sample size of games for a season is merely 256, and with the variance in stats too large for them to become useful until approximately week 5, you're looking at a total testing window of only around 192 games. I'd love for that number to be more (let's hope for a 2 game preseason and 18 game regular season soon), but by back testing you can expand the sample size and at least make it statistically testable. The only problem there is that by using so many past statistics, you open the door for unwanted biases and regressive formula molding (for lack of a better term). Past occurrences are often not as valuable as you'd wish for future predictions.
With all that said, I think you're initial work and the ideas behind it are good. May I ask what mathematical/statistical and football background you have or what you currently do for a living? I began creating a program for the NFL (I attempted a similar program for NCAAF, but like I said, the variance proved to be insurmountable) last year which produced decent results - I posted my plays on here and I think they were 13-12-1 as of Week 13 when I stopped posting them. I wanted to make a few tweaks to the system, but didn't want to mix the new "tweaked" plays with the previous plays, so I stopped posting the program's plays. The combined program plays finished the year with a very respectable winning percentage near 62%. I will begin posting them here this season in Week 5, when the stats the system uses have enough sample data points to become useful. I wouldn't mind having another knowledgeable, intelligent, innovative person (both in regards to football and statistics/mathematics/Excel) to look over the work and possibly add some new fresh, creative ideas to the program that I may have overlooked or would be useful in adding before the finished program plays are posted. Like I said, I want to post the entire season with no adjustments to the program so its plays and results will be entirely uniform.
If you have any interest, feel free to PM me and we can discuss your background, the program, and any ideas you might have to augment it. If you're not interested, I completely understand and wish you the best of the luck with your college football plays!
