I wanted to take a moment and say thanks to Tibor, one of my 12 readers, for forwarding me some really interesting research papers on modeling NFL and NBA games with neural nets. There are some really good nuggets of information in those papers, especially the discussion on setting the right momentum and learning rates.

I used that information to fiddle around with my neural net model and I’m posting some recent results from my DRAFT NFL neural net point spread model.  As you can see, the model is pretty good at determining if the home or visiting team will win (a negative sign means the home team wins) but the predicted spreads are way off relative to actual spreads.

This leads me back to developing some sort of ranking system to feed the model, which I wrote about in my "Thoughts on Ranking Football Teams" post.  The good news is that the research papers that Tibor sent me allude to a type of football match system where the model learns the results of previous games and then applies its statistical analysis to new match ups.  Despite this good nugget of information, I feel that I have a long way to go to get something solid before the season starts.

In the interest of science, and because I love my 12 readers, I’m uploading my EasyNN Plus data file for this particular model.  However, you’ll have to have the full version of EasyNN Plus to use this file because the model uses 980+ example rows and the test version only allows you 100 rows.  If you follow the link above and buy EasyNN Plus from there, I will get a small commission from Steve.

I’ve made some serious headway last week in analyzing NFL football data to model game point spreads.  I was able to determine with great accuracy (84%), using backfitting, what team would win a matchup. I did this by building a backpropogation model from 2007-2008 data with about 20 matchups as the prediction set.  The only bad thing I discovered was that the point spread prediction was way off.

As I wrote before, backfitting is NOT the way to go for building this type of  model and the fact that the magnitude of the actual point spread was off bore that out.  You might be asking yourself right about now, "if this is not the way to go, why did Tom do this?" 

I did this because I’m still in my data discovery phase looking for relationships and tinkering around.  Now that I understand various data relationships and can detemine the winners relatively well, the next step, and perhaps the hardest now, is determining the ranking system.  I suspect that the ranking system will help get the game spreads closer to what they should be for future games.  If all goes well I should be making spread calls on this blog for the next season as a way to determine if my model is worth a sh*t.

The problem with developing a point spread betting system for football teams is that you can’t initially use a neural net. This is because backfitting raw stat data could lead to poor forecasting, or handicapping. What you have to do is come up with a method where you rank teams first and then feed them into a neural net system to forecast the estimated handicap or point spread.

Why? Teams change over the season; they may lose a key player to injury or evolve their strategies per game. Although neural nets are fantastic, they can’t cope with these paradigm shifts easily. However, if they are fed a ranking system that gets updated weekly based on data from played games, those rankings can be fed into the model and compared to opposing teams in the next game.

Now the trick is to develop this ranking system without backfitting data. How do you do this? It’s not easy but the route to use in my mind is to use neural net clustering to first identify if any pieces of data seem to drive the point spread. Once you know that, then its a matter of devising a mathematical model to help you rank the teams.

FYI, I am using EasyNN Plus for this project. I may or may not post my data files.

I’ve come across some very interesting, albeit very initial, relationships in NFL data. The actual point spread between teams seem to have a strong relationship with the defense’s pass interception ability, the attempted rushing by the defense, and sack yards by the defense.

Way more interesting stuff that I’m not listing at the moment.

If you had to build a neural net model to help you rank NFL teams, what bits of data would you want to include in your input data?

Pass Attempts vs Completed?
Yards Rushed?
Score Differential?
Time of Possession?

What? I’m interested to hear your comments, if any.