Saturday, September 14, 2013

2013 College Football Modeling: Week 3 Viewing Guide

As we head into another glorious weekend of college football, I'd like to share a Week 3 Viewing Guide I've been working on. The guide includes a new metric I've been working on called Watchability. My purpose in inventing Watchability is to help people identify close, well-contested games they might not have otherwise thought to watch. Going forward I plan on posting one each week on Thursday mornings. I tried to include as much information as I could without being overwhelming. For each game I'm showing:
  • Day and time (Pacific Time)
  • Home team and away team
    • Teams are shaded according to the model's estimation of their chance to win
    • The more green a team is shaded, the better its chance to win; the more red, the worse
    • Ranking (when listed) is this week's AP ranking
    • Teams who have already played are italicized and shaded fully red or green as appropriate
  • Watchability edit: I've since renamed Watchability to Game Quality
    • This is a measure (scaled from 0 to 100) which attempts to capture how likely the game is to be a good game
    • Its purpose is to identify games that might be fun to watch
    • It's based on how good the two teams are, and how likely the game is to be a close game
    • The theoretical max of 100 would be two teams rated 1.000 (team ratings range from 0 to 1) playing each other on a neutral field
    • The highest rating I've seen so far has been around 95 in some simulations of the BCS National Championship Game featuring undefeated teams from the Pac12/Big10/SEC
    • Oregon @ Stanford at 80.9 is currently the highest rated regular season game
    • There is a little more math on Watchability below


I calculate Watchability as follows:

I begin by calculating a Power Rating for each game. This involves taking the model's average rating for the two teams and scaling it from 0-100. A game's Power Rating is a measure of the combined strength of the two teams playing.

Power Rating treats Oregon (.892 rating) vs. Tennessee (.563) the same as Arizona State (.672) vs. Wisconsin (.727). The Power Rating for Oregon vs. Tennessee is 72.7, while the Power Rating for ASU vs. Wisconsin is 70.45.

I then translate Power Rating into Watchability by calculating on how far each team's chance to win the game deviates from 50% and reducing Power Rating accordingly. The reduction is based on the square root of the deviation, so 60/40 games are only slightly penalized relative to 50/50 games, but 90/10 games are penalized heavily because blowouts are so boring. Take the two example games: Oregon is a 90% favorite so that game's Watchability is 27.2 (reduced significantly from its power rating of 72.7). On the other hand, ASU and Wisconsin are nearly 50/50, so that game's Watchability is 70.43 (nearly identical to its Power Rating of 70.45).

A few last comments:
  • Watchability is a context-neutral statistic. It doesn't know which team is your favorite, it doesn't know that this game has BCS implications, it doesn't know that this is a bowl game. It doesn't know that this is a rivalry game or the Rose Bowl or the BCS Championship Game. All it knows is how good the teams are, and how likely the game is to be close.
  • For context, I added average Watchability for some different types of games as an output for the last round of 10,000 season simulations and calculated the following:
  • Regular season games average around 30 Watchability, conference title games & non-BCS bowl games average 50, BCS bowls average around 70, and the BCS National Championship Game averages around 75.
  • Watchability does not predict defied-the-odds kinds of upsets. Those are just surprises to be enjoyed by us all.

No comments:

Post a Comment