Rather than simply posting a table with what might perceived to be made up numbers, I thought it might be helpful to give a high level overview of how the model works.

In the simplest terms possible, the model takes a bunch of “predictors” (called independent variables) and works some mathemagical wizardry (called a support vector classifier) on those predictors in order to arrive at an “outcome” (called a dependent variable).

The independent variables for the model (the things that are going to predict who covers) are the Total DVOA (from footballoutsiders.com) of each team, the preseason DVOA of each team, what week it is, who’s at home, what the spread is, and what the over under is. From those variables the model makes a prediction of “HOME” or AWAY,” indicating who it expects will cover the spread.

To teach the model what to predict, we give it examples of what has happened in the past. The model was fed 3,175 past NFL games with complete data for all the variables above. We held back about 600 of those games (across all weeks and seasons, for consistency) to test its performance. We then took the remaining 2,500 or so games and “trained” the model by turning some mathematical knobs and dials (called “hyperparameters”) and maximizing how well it predicted random blocks of those 2,500 games (this is called “n-fold cross validation”).

Once the model was tuned up how we liked it, we told it to predict those 600 games we held back so that we had a final measure of our performance. The model predicted 52.75% of spread wagers correctly.

There are some additional complications (ie. handling pushes, one or two missing seasons of preseason FO data, stratified sampling of the folds, etc). But in the most condensed way that is how the model works

### Like this:

Like Loading...

## Leave a Reply