Saturday, December 29, 2012

Bowl Game Line Predictions

The power rankings I create for this site are calibrated to a single purpose: to predict, as closely as possible, point spreads for future matchups.  I am trying to understand what (and how) the betting market thinks.  If my rankings worked perfectly, you would take the difference of the opposing teams' GPF (Generic Points Favored), add a home field adjustment, and the resulting number would equal the point spread.

Of course, my rankings do not work perfectly, and the predicted point spreads based on my rankings tend to miss, on average, by about 2.5 points (for college football). In a previous post, I compared my ranking's accuracy to the rankings provided by the Sporting News Linemakers section (the only other publicly available rankings system I'm aware of that has the same purpose as mine).  For the week I chose, I found that the predictive accuracy of the two ranking systems was about the same (my rankings had a mean prediction error of 2.1 points, compared to Linemakers' error of 2.2 points).


In this post, I made a prediction on the point spread for the Alabama-Notre Dame BCS Title Game.  My rankings predicted an 11 point spread while Linemakers predicted a six point spread (both in favor of Alabama).  The opening line was around 8 points, closer to the Linemakers, but the line has since stabilized at 10 points, closer to my prediction.

Bowl Games

The purpose of this post is to test the predictive accuracy of my rankings system using all of the 2012 Bowl Games (a more expansive version of my Alabama-Notre Dame test).  Bowl games can be tricky for my ranking system since, by design, they pit teams from different conferences against each other.  Getting a proper estimate of the relative strength of the various conferences can be challenging since most inter-conference games take place towards the beginning of the season, where the market's evaluation of team strength may still be in flux.

Methodology

As with my prior post, I will compare my rankings' accuracy against the Linemakers rankings.  I am using the Linemakers rankings from November 28, 2012 as these were the last rankings released prior to announcement of the Bowl Games and the opening lines.  My rankings are archived daily, so I pulled my rankings from the same date.

I was also curious how closely a pure stat-based model would predict point spreads. So, I calculated predicted point spreads from Football Perspective's Simple Ranking System, using their Week 13 rankings, which were published on November 25.

The table at the bottom of this post summarizes the results.  The three columns on the right show the predicted point spread and how big the miss was (BMA=this site, LNM=Linemakers, SRS=Simple Ranking System).

Results

Here are the results in total:
  • Average Prediction Error
    • Betting Market Analytics: 2.1 points
    • Linemakers: 2.6 points
    • Simple Ranking System: 2.3 points
My rankings performed better than the Linemakers, with a 20% lower average prediction error.  Oddly enough, the Simple Ranking System, which looks at margin only and ignores point spread, also outperformed the Linemakers (by 0.3 points).

Once again, a straight ensemble average of my rankings and Linemakers outperformed either model on its own, achieving a 1.9 point mean average prediction error.

Here is the table.  It is sorted in descending order by the combined GPF ranking of the two teams (by that metric, Kansas State vs. Oregon is a bigger matchup than the Alabama-Notre Dame title game).

No comments:

Post a Comment