Ratings Methodology


One of the things that makes sports so exciting is the debate that inevitably ensues after every week of games about which team or group of teams are better than the others. The human polls tend to weighed down by emotion and bias, whereas the computer polls can’t actually watch the games. My goal was to make a set of ratings that followed common sense by trying to mimic human polls with a solid statistical element driving it. With that, I introduce to you the Donchess Inference Index (DII) version 1.0.

Methods – Layman’s Terms

I have never been a fan of statistical theory. While it is obviously very important, my experience has taught me that theory rarely translates to practicality. The DII is actually a combination of two ratings: a standard rating and an inference rating. The standard rating simply looks at how each team performs in each game (this method is preferred toward the end of the season). The inference rating takes things a step further. It assumes that if Team A beats Team B by seven points or more, that Team A would beat every other team that Team B has beaten (this method is preferred in the beginning of the season where there is a lack of data). I create a rating for both methods each week, and the standard rating is given a higher weight each week in comparison to the inference rating. The underlying basis of my formulas is built around logistic regression. If you would like to be bored with the details of logistic regression then I suggest referencing Wikipedia here. This method is, in fact, not much different than a lot of ranking systems already in place. The DII differs in that it weighs each game differently based on a variety of factors that we humans use either consciously or subconsciously to rank teams.

Factors Included
Game Outcome (by far the most important)
Score Differential (up to 28 points in football)
Home/Away/Neutral Site
Time Since Win/Loss

Factors Not Included
Regional Bias
Future Schedule

Things like Strength of Schedule (SOS) are naturally built in with logistic regression and is thus not “weighted” like the variables described above.