Following the travesty that was the final AP rankings of the 2008(-09) season, I decided that I could do better than the BCS black box at accurately and unbiasedly ranking college football programs to appropriately determine bowl matchups. I spent the next couple years very closely watching the way the polling and BCS systems worked and began identifying metrics that I felt weren't accurately accounted for in those systems. In short, I tried to identify their weaknesses(biases) and determine how I could correct for them mathematically. After much brainstorming on many ideas, countless conversations with fellow fans, and handwriting preliminary versions of the algorithm, I began development of the core code in ~2012.
The original premise behind the code was to rank all the FBS teams against each other based purely on actual performance without the inherent biases of polling-based rankings. The rank of each team is based primarily on whom they have beaten and whom they have lost to, and how well they performed against each of those teams. Each of those teams is also ranked based on their opponents and performance. When this approach is applied to all the teams simultaneously an iterative method is then necessary to get to a converged result. For each team, a Relative Total Performance Score (RTPS) is determined and sorted to rank the teams.
I want to emphasize that my opinions are not in this ranking system, other than indirectly through the fact that I developed the algorithm and chose the metrics it would be based on. That being the case, I am often surprised at how the results come out each week. Because the rankings are based purely on actual performance metrics, the more data I have, the more "correct" the rankings become. Thus, I don't publish any data until about week 5 of each season, and as the season progresses the rankings become more informed and thus more accurately reflect the "correct" order of the best teams.
There is a lot more I am doing with this as well; details in subsequent posts.
The original premise behind the code was to rank all the FBS teams against each other based purely on actual performance without the inherent biases of polling-based rankings. The rank of each team is based primarily on whom they have beaten and whom they have lost to, and how well they performed against each of those teams. Each of those teams is also ranked based on their opponents and performance. When this approach is applied to all the teams simultaneously an iterative method is then necessary to get to a converged result. For each team, a Relative Total Performance Score (RTPS) is determined and sorted to rank the teams.
I want to emphasize that my opinions are not in this ranking system, other than indirectly through the fact that I developed the algorithm and chose the metrics it would be based on. That being the case, I am often surprised at how the results come out each week. Because the rankings are based purely on actual performance metrics, the more data I have, the more "correct" the rankings become. Thus, I don't publish any data until about week 5 of each season, and as the season progresses the rankings become more informed and thus more accurately reflect the "correct" order of the best teams.
There is a lot more I am doing with this as well; details in subsequent posts.
Great stuff, are you starting predictions this week or next?
ReplyDelete