I've been leery of neural networks for some time because of their perhaps undeserved reputation for overfitting data and because of the practical difficulties that have existed in using them for prediction. Phil Brierly's Tiberius software includes an implementation of neural networks that has, at least for now, converted me. As a consequence, I'm adding one final margin predictor to the mix for 2011.
Within Tiberius it's easy to split an available sample into "train" and "test" components, which somewhat reduces the risk and extent of potential overfitting, since any model constructed on the training data set needs to also perform well on the test set to be considered adequate. Probably more importantly though, Tiberius also produces an Excel spreadsheet via which it is possible to predict outcomes for future games.
So, over the past couple of days I've been training a neural network to predict game margins using the probabilities predictions of the TAB Bookmaker, ProPred, WinPred and the Adjusted and Unadjusted versions of Head-to-Head for the games from seasons 2007 to 2010. As for the other margin predictors, I've used just 50% of the available data as the training set, leaving the other 50% as the test set.
The final model performs about as well as Combo_7, with a Mean APE of 28.71 points per game and a Median APE of 24.23 points. Since it's also a Combo model - actually more of a Combo than Combo_7 as it uses all of the probability predictors and not just a couple - I'll be labelling its predictions with Combo_NN.
I'll be fascinated to see if it can permanently dispel my mistrust of neural network models.