This is a metrics-driven prediction system for one full season of college football (FBS). Rather than building a model from scratch, it takes a delta approach, asking: “What has changed since the end of the previous season?”.
Data is gathered at the most granular level possible. For instance, rushing yardage and passing yardage are preferred to total yardage. The data subsequently used in the model depends on relevance and the amount of history available. Relevance is determined by the Pearson Correlation Coefficient. As such, each data category is measured against the season end Jeff Sagarin Predictor power ratings.
Four correlation tests are made:
* Category to Power
* Category to Year-over-Year change in Power
* Year-over-Year change in Category to Power
* Year-over-Year change in Category to Year-over-Year change in Power
Or more popularly: Value-Value, Value-Delta, Delta-Value, Delta-Delta
The highest of the four is used, and the one used also determines how the data is subsequently converted to “field points” (also known as "deltas").
Each category is rescaled into the range of the power ratings. For example, teams will return 0-100% of their rushing yardage. If a team returns 50%, and the power ratings span from 100 to 40, then the 50% is scaled as 70. This allows conversion to field points, and also facilitates comparison of categories.
The main limitation of this system is that the data categories will often intrinsically overlap each other, and thereby introduce “double-counting”. For example, recruiting rankings and unit rankings overlap, as do QB ratings and passing yardage. I try to limit this effect without excluding strongly-correlated data.