Super Smart is Taking Heed of Bookies

Across a series of blogs now we've explored the Super Smart Model (SSM) and investigated its ability to predict victory margins. In this blog we'll look more closely at which variables most influence SSM's forecasts.

SSM uses just six pieces of independent data to come up with its predicted margin for a game:

  • The bookie's assessment of a team's probability of victory (implicit in the bookie's head-to-head prices)
  • The average result for the last two games for each of the teams playing in the game
  • The MARS Ratings for each of the two teams
  • An indicator that flags if the contest can be designated as an Interstate Clash or not

The six pieces of information don't equally effect SSM's predictions. One way to demonstrate this is to calculate the historical correlations between SSM's predictions and each piece of data.

2010 - SSM Correlations.png

For now, focus on the top row of the table. There you can see the numerical evidence for the overwhelming effect of Bookie Probability of SSM's predictions: the correlation is +0.98.

If you glance across to the far right of that first row you can see that the difference in the MARS Ratings of the two teams also has a significant effect. The correlation of this difference with the SSM prediction is +0.88. Other variables have smaller effects. The correlation of SSM's predictions with the difference in the average results for the last two games is +0.66 and with the interstate clash variable is +0.36.

Since SSM's predictions are highly positively correlated with both the bookie's probability and with the difference in the team's MARS Ratings, it seems likely that the bookie's probability and the difference in MARS Ratings will themselves be highly correlated. They are: the last column of the second row of the table above records that this correlation is +0.89.

One way of describing this is that there's quite a bit of overlap in the information content of MARS Ratings and bookie probabilities.

Some other of the correlations in the table above are also interesting, not because they provide any further insight into the workings of SSM but because they shed light on bookie price-setting behaviour and the information content of MARS Ratings.

For example, the second row shows that the TAB Sportsbet bookie is more influenced by the fundamental abilities of teams (as encapsulated in their MARS Ratings) than he is by the recent performances of teams. This we can deduce from the correlations of Bookie Probability with the difference in MARS Ratings (+0.89) and with the difference in the average of the teams' last two results (+0.62).

Further, since the correlation between Bookie Probability and the Interstate Clash indicator is only +0.22, we know that the fact that the home team does or does not enjoy the benefit that accrues from playing in its home state while its opponent has been forced to travel is of less significance to the bookie than either of these two measures of team form.

The very last number in the table shows that, while MARS Ratings are somewhat influenced by recent form - the correlation of MARS difference with the difference in recent results being +0.57 - they are not completely determined by them.

Another way of demonstrating the high correlation between Bookie Probability and Predicted Margin is via a scatter plot of these two variables.

2010 - SSM Pred Margin v Bookie Prob.png

It's interesting to note that, historically, SSM has had a slight bias towards predicting victories for home teams that were equal favourites with the bookie, as evidenced by the fact that the trend-fitting line shown in the chart above is above zero for a home team probability of 50%.

That aside, as this chart suggests, most of the time SSM predicts the same team to win as does the TAB Sportsbet bookie. In fact, the two have disagreed on only about 6.5% of games since Round 1 of 2006 - and this figure treats SSM as agreeing with the bookie whenever there is equal favouritism. SSM finds, as have we all, that it's more often wrong than right when it disagrees with the bookie. Its tipping accuracy in those games where it predicts the underdog to win is just over 40%. But, so far at least, it compensates for this shortcoming by being a superior tipper of margins than is the bookie.

One final chart to finish. It shows the tight relationship between Bookie Probability and MARS difference, though here I've chosen to use as the probability measure the difference between the Home Team and the Away Team probabilities. You can see that it's quite rare for the TAB Sportsbet bookie's opinion of the difference in each team's chances to be unreflective of the difference in the teams' abilities as measured by MARS Ratings. I'd like to pretend that this was a compliment to the skills of the bookie but, in truth, it's more a nod to the validity of MARS Ratings.

2010 - SSM Prob Diff v MARS Diff.png