How Different is C_Marg, Really?

Over the first seven rounds of season 2014 it's not been an unusual occurrence for C_Marg to be the Margin Predictor most different from the all-Predictor average for a round.

One way of quantifying how different C_Marg has been from the other Margin Predictors across the 63 predictions they've all made so far is to calculate the average absolute difference in C_Marg's margin predictions from those of each of the other Margin Predictors. While we're at it, we might as well calculate these "manhattan distances" for all pairs of Predictors.

View fullsize

Looking across the row for C_Marg we can see that it's managed to be distant from every other Predictor. Its minimum average manhattan distance to any other Predictor is 8.1 points per game, which is the space it's put between itself and the two RSMP Predictors. It's also about equally proximate to Bookie_9 and Combo_7. C_Marg is most distant from the two Predictors based on ProPred and the two based on WinPred.

An inspection of the rest of this distance matrix shows that all of the other Predictors are at least reasonably proximate to one or more other Predictors.

One way of visualising this distance matrix in two dimensions is to undertake some classical multidimensional scaling and then plotting the Predictors in the space of the first two principal co-ordinates.

Using R's cmdscale function from the stats package yields the following diagram, in which similarity of predictions is encoded by proximity.

View fullsize

C_Marg's uniqueness is starkly illuminated by this chart, as is to a lesser extent that of Bookie_3 and Combo_NN_2. Also depicted are the very high levels of agreement amongst the four H2H Predictors, and amongst the two ProPred and two WinPred based Predictors. The two RSMP Predictors' Margin Predictions are also shown as being, in general, very similar as are, somewhat surprisingly to me, those of Combo_NN_1 and those derived from Bookie_LPSO. The other result that surprises me a bit is the distance between Bookie_3 and Bookie_9.

SUMMARY AND CONCLUSION

This year's major new Margin Predictor, C_Marg, has been shown to provide a quantitatively distinct set of margin predictions to MatterOfStats. Though this has not, so far, led to high levels of accuracy in margin prediction, it has provided us with a Predictor showing superior line betting proficiency.

It'll be fascinating to see whether this ability persists over the remainder of the season.