Friday, January 11, 2013
2013-01-10: NFL Divisional Playoff Predictions
For the NFL Divisonal playoff week the predictions for all of our algorithms are in agreement. For our predictions we run a number of different types of algorithms in our research and compare the outputs. The three main algorithms that have consistently had the best performance are a Support Vector Model (SVM), a Multilayer Perceptron Neural Network, and a ranking algorithm. All three algorithms sided with the favorites except for Atlanta. All three picked Seattle for the upset.
All season long our algorithms have consistently shown that Atlanta is over-rated. Yes they have won most of their games this season but they have had the easiest strength of schedule this year out of the all of the NFL teams. ESPN's adjusted strength of schedule shows that Atlanta has had an easy season. Big Lead Sports states that not only was Atlanta's season easy but it was respectively the easiest season for any NFL team in a number of years. Most of our algorithms take the strength of schedule into account and Atlanta has been ranked accordingly. That being said we did not take Seattle's loss of Chris Clemmons into account.
The SVM gives us a binary output, winner or loser. There is nothing in between. The SVM chose Denver, San Francisco, Seattle and New England.
The Neural Network output is a continuous variable that is supposed to be the margin of victory. A positive score favors the home team and a negative score favors the visitor. The Atlanta - Seattle game had an output of -1.49 so Seattle is essentially a one point favorite albeit a very close favorite. San Francisco was not as close with an output of 3.1 making the 49ers a 3 point favorite.
The ranking algorithm is based on Google's PageRank algorithm which we have outlined in previous posts. Seattle is ranked higher than Atlanta so Seattle is the predicted winner. The rankings for the entire 2012 season are visually represented the the figure below.
San Francisco is a bit of a concern as they changed quarterbacks in week 10 after their starter suffered a concussion. We did not make any adjustments to account for this as their was no significant delta in the slope of the performance rating. Our picks for this week.
Superbowl: using preliminary data all of our algorithms are currently predicting San Francisco and Denver in the Superbowl. What they disagree on is who is going to win.
-- Greg Szalkowski