About
This was designed by Jack Cackler, a doctoral student in Biostatistics at Harvard University.
To make winter travel plans with better information, this site graphically illustrated of all of the bowls a team might play in, how likely they are, and where and when each bowl will be.
With the regular season concluded, you can now view each team's probability of winning the bowl they are in, and purchase tickets for that bowl.
Feel free to contact me at jack.cackler@gmail.com.
Updates
-Previous weeks available to be viewed in links at the top of the page
-Predictions for this week available
-To purchase tickets, click on the bars
-For completed games, winning team indicated in green, losing team in red
With the Conference Championships and regular season concluded, the stage is now set for the bowl games. I've put up links to all previous weeks at the top of the page to view how each team progressed from week to week. Though I feel I've made a lot of progress both in presenting data accurately and in an easily interpretable way, I'm leaving previous weeks up in the format they were originally presented. For the remainder of the season, predictions for each of the bowl games are indicated.
Many of the games are predicted to be fairly lopsided. No teams were given home-field advantage. Close games in which the less favored team wins more than 40% of the time include:
Oregon-Kansas State
Arkansas State-Kent State
Texas A&M Oklahoma
Northwestern-Mississippi
Rice-Air Force
Oregon State-Texas
Rutgers-Virginia Tech
UCLA-Baylor
Washington-Boise State
With two teams out, Northwestern may be the Big Ten's only hope of any bowl victory (unless you count new member Rutgers, which also has a good chance.)
While I'm predicting Alabama to win about 2 in 3 times, I do think Notre Dame may have a shot, primarily because Alabama plays a very heavy rushing game, and Notre Dame has an excellent rush defense. I wouldn't bet on Notre Dame, but I certainly wouldn't count them out. I'm afraid Northern Illinois' luck might run out here, but given the way Florida State played against Georgia Tech, they may still have some hope as well.
This started as a fun experiment to do in my free time, and I've been amazed with its success. To date, the site has garnered 200,000 hits, which far exceeded anything I had originally expected. It's difficult to quantify how successful the algorithm was, as by nature, it never intended to predict absolute bowl destinations for teams, just indicate the probability that each team had. There are a few areas I can point to in which the model was phenomenally successful, even more so than any pundit.
The biggest point of contention with the model was Northern Illinois' presence at the top of the rankings throughout the last month of the season. Before I added Bayesian weighting for conference strength and poll data, Northern Illinois was even indicated as a possible National title contender, which I knew would never happen. Still, starting in Week 11, I predicted that Northern Illinois would make a BCS Bowl game 79.8% of the time, a thought that I only saw started to be suggested before the final week of the season. Even after the conference championships were concluded, particularly with Nebraska, UCLA, and Texas losing, there were still many columnists doubting that Northern Illinois would make a BCS game, but they did. Part of me feels some pride in predicting what few others did a month in advance, but this would be silly, because all I did was interpret the numbers that were publicly available.
One surprise was that I knew two to three teams would be eligible, but not receive a bowl, but I did not predict that they would be Louisiana Tech, who scored more points than any team this season, and a respectable 8-4 Middle Tennessee, who beat Georgia Tech, a team that was about 20 yards away from an Orange Bowl. One of the weak points of this algorithm, as I've mentioned, is all it does is predict conference ranking, and it's hard to determine numerically the multiple machinations behind how bowls select teams based on things like distance, projected attendance, and other intangible elements. I started developing an algorithm to incorporate some of those measures, and may include it next year. Two of these exceptions I eyeballed and predicted were the Cotton Bowl selecting Texas A&M over LSU, and the Sugar Bowl, who could have picked the higher-ranked Oregon to replace the SEC Champion, picking Florida.
As an aside, there's been much handwringing about the presence of Northern Illinois in a BCS bowl, and rule against more than two teams in a conference being in a BCS bowl. I'm actually not convinced this is a bad thing, in a large part due to the brevity of the season and the impossibility to rank teams between conferences with any accuracy. Most teams play 7-9 games within their conference, leaving only 3-5 against other conferences. This leaves a huge sparsity in data available for projecting rank between conferences. For example, undefeated Ohio State beat Cal, who didn't even get a bowl this year, by a very close 7 points. Had Cal run a five lateral play, and won the game, this would have significantly affected all relative rankings between Pac-12 and Big Ten schools. At some point, I'd like to look into just how much confidence there is in ranking schools between conferences. The point is, without a confident way to compare teams in different conferences, the fairest thing is to take the winners of each conference, and play them against each other. The playoff next year should smooth many of these issues out, though I'm sure many will remain.
Methodology
These predictions were generated largely based on the pure points predictor ratings created by Jeff Sagarin, and will be updated weekly.
For each game remaining in the schedule a probability of winning was determined based on the rating of each team and home field advantage.
Based on these predictions, each game was sampled 10,000 times and for each sample, conference championship games and subsequently bowl games were computed.
Overall ratings were given by an Elo rating system used by the computer rankings in the BCS standings. For each team, the number of times it ended up in each bowl was counted and is given as a percentage. The maximum standard deviation of any estimate is .5%. Bowls that occurs .1% of the time or less are grouped as "Other", and are theoretically possible, but highly unlikely
There is also a prior distribution for each conference's poll ranking versus computer ranking, essentially adjusting each team's rating up and down by a number of points depending on two factors. The first was the overall strength of the conference, derived from the Central mean of their Sagarin Pure Points rating. The second was, looking back at the end results since 2005, how teams poll rankings were compared to their computer rankings. These were averaged by conference, and conferences were maximally adjusted up or down half a game against a neutral team.
In other words, an ACC team that went 10-2 would be comparable with an MAC team that went 11-1 against the same teams, as the ACC is an average conference that does very well in polls compared to computer rankings, and the MAC is a mediocre conference that does not do well in polls.
The precision of the end result will be lower, and so you can uncheck the "Use Poll Data" box to view results purely by computer predictions, but the results are probably much closer to what will actually happen.
Each team that plays this week is also displayed with their imputed likelihood of winning this week's game.
The predictions are robust, but have several limitations that could impact accuracy. The prior I used has been trained on 8 years of data, but may have limited accuracy..
As such, these predictions will over-rank teams with good numerical rankings but poor poll standings.
While most bowl seats have contracts with conferences, in some cases a bowl may choose any available team if no team they are contracted to is eligible.
These predictions simply assign the highest rated team not yet in a bowl, but there may be instances in which a bowl might pick a lower-ranked team that they believe has a larger fan base.
Analysis coded in R, graphs constructed with the D3 Library.