I have recently been working on my AE horse ratings, looking at a new and perhaps more productive way of combining the AE values assigned to each attribute, than simply multiplying them together. This blog entry will examine my work with a neural network approach, the Weka freeware software and the results I have achieved.
AE values are calculated by taking a given characteristic, for example if we examined all horses running today that have won previously going left handed if todays race is on a left handed track and involves a bend. We would then calculate the expectancy of this horse winning based on the market odds and allowing for over round. We would add this to a running total called expected wins and then if the horse won we would add 1 to a second running total called actual wins. If we do this over a reasonable length of time we arrive at a total for each count. These totals are called actual wins and expected wins. If we divide the actual by the expected we arrive at an AE value. Clearly if the AE value is greater than 1 then more horses than the market predicted have actually won and if it is less than one then fewer have won than the market predicted. If we compile AE values for a number of different attributes we can then use these for todays runners. The typical approach would be to check the characteristics of each horse today for the chosen attributes eg has it won left handed or not. We would have as our AE values figures for won left handed and not won left handed along with any other attributes we had deemed predictive. We would then multiply these values together to get a master AE value for a horse.
The purpose of this article is to look at whether simply multiplying is the best approach or whether a NN could find a more predictive way of combining the individual AE values for each attribute.
Firstly I took a broad brush approach and calculated the average AE rating across all handicap flat races for 2013. I then looked at how I would have performed backing all horses to BFSP that had a rating above the average. This produced 16320 bets and a profit of +93.7 pts ROI = 1.47%.
Backing all horses that had a rating below the average produced 17140 bets and a loss of -412 pts ROI -3.1%. The ratings in there basic form seem to have done a reasonable job of sorting out profit from loss.
Now if I feed the individual AE values for each horse component into a WEKA neural network, rather than simply multiply them together, and ask the NN to create a master rating for each horse, we get the following results, once again using the average rating as a dividing point and having trained the network on 2009 to 2012 data.
Those above the average produced 17228 bets with a profit to BFSP of +455.3 pts ROI +2.6%
Those below the average produced 14641 bets with a loss of -812 pts ROI -5.5%.
The suggestion here seems to be that the NN made a better job of combining the AE values for each rating component when compared to simply multiplying the AE values together.
WEKA is freeware software and is supported by the University of Waikato. There is a rich source of online material and even free online courses run by the University which you can tap into.