Jason Rennie's College Sports Rankings

College Basketball Rankings

2007

The 3/11/07 ranking has (so far) correctly predicted 3 (out of 5 predicted) upsets:

• Winthrop over Notre Dame
• Tennessee over Virginia
• UNLV over Oregon
• UCLA over Kansas
• UCLA over Florida

Bracket points: 120/192

• First round: 28/32 (28/32 correct)
• Second round: 24/32 (12/16 correct)
• Third round: 28/32 (7/8 correct)
• Fourth round: 24/32 (3/4 correct)
• Fifth round: 16/32 (1/2 correct)
• Sixth Round: 0/32 (0/1 correct)

• Games through 3/11/07
• Games through 2/4/07
• Games through 1/28/07
• Games through 1/21/07
• Games through 1/14/07
• Games through 1/7/07
• Games through 12/31/06
• Games through 12/24/06

2006

The 3/12/06 ranking out-performed 62% of all brackets submitted for the Yahoo Tournament Pick'Em. It correctly predicted 2 (out of 4 predicted) upsets.

• Games through 4/3/06 (Post-tournament)
• Games through 3/12/06 (Pre-tournament)

2005

• Games through 4/4/05 (Post-tournament)
• Games through 3/13/05 (Pre-tournament)

College Football Rankings

2006

• Games through 1/8/07
• Bowl Games (using 12/3/06 ratings)
• Games through 12/3/06
• Games through 11/26/06
• Games through 11/19/06
• Games through 11/12/06
• Games through 11/5/06
• Games through 10/29/06

Previous Years

• 2005
• 2004

Links

• Kenneth Massey maintains college basketball and college football ranking comparisons.
• Peter Wolfe provides team and game information for college basketball and college football.
• Ken Pomeroy provides daily-updated college basketball scores with more fine-grained venue and overtime information.

Pensieve

• 1/16/07 - Found a bug in the CF python code---neutral stadiums were not being handled correctly. I'm going to re-do the 1/8/07 run.
• 11/18/06 - It would be interesting to include stadium seating capacity as a variable in the home field advantage variable. Googling college football stadiums provides useful results. CollegeCharlie.com includes a link to 2005 game attendance sorted by average attendance per game and percent of capacity. Looks like it wouldn't be too difficult to make this happen...
• 11/3/06 - Realized that I had been using a weird procedure for initializing parameters for gradient descent when selecting the regularization parameter via cross-validation. I now use the optimized parameter setting from the previous round, and use a small absolute tolernace (1e-10) to terminate the gradient descent.