Making a profit from rating horses 




IntroductionFor many years I have been analysing horse racing data looking for profitable systems. I am a Chartered Statistician by trade but horse racing is a hobby of mine and the opportunity to make some extra cash has always been a big incentive. Actually, it’s quite easy to find a reasonably simple system that makes a profit … at least for a while! However, the nature of horse racing, with many unknowns and trends makes it very difficult to find something that stays in profit over a number of years. Once I had collected enough data and experience, it was always my intention to progress onto horse ratings. In theory this approach can take into account many more factors than a system and it also rates every horse in the race giving a better view of the opposition. At the start of 2003 I started to compile ratings for sprint handicaps. I worked closely with Dave Renham from Drawn2Win (who specialises in draw bias in sprint handicaps) and using his knowledge and my technical expertise, rated every sprint handicap over the flat season in 2003. We refined the models over the winter AW season and then again for this year’s flat season. We also embarked on a detailed analysis of the ratings looking for angles within the ratings that should make a reasonable profit over the summer. We now have a number of approaches – covering both backing and laying horses – which we are trialling with Drawn2Win members and any other interested parties.
What makes Good Ratings?
There are many ways to compile ratings. There are some very basic approaches which can be carried out by hand (e.g. “10pts if 1^{st} last time out; 5pts if 2^{nd}“ etc.) or there are the more complex kind based on statistical modelling and which need a computer to run. Whatever the approach, good ratings should exhibit 2 key features:
If these 2 “rules” are obeyed, it makes the ratings much easier to use and therefore profit from. The ratings I produce follow these rules and are designed so that the horse with an average chance of winning has a score of 100. The highest a horse can get is around 140; the lowest about 60. Here’s a breakdown of how the ratings have worked on 5f sprints over the last 4 years:
The table shows that horses with a higher rating win more. If every horse with a rating of 120 or higher had been backed over the last 4 years the profit would have been just over 50pts.
Here’s how the top 3 rated horses performed over 6f over the same period (20002003):
Top rated horses won in just over 20% of races and made a profit of 95pts. 2^{nd} rated also made a profit. Finally, here’s the strike rate of top rated horses with different gaps to 2^{nd} rated over 6f:
Top rated horses 10 or more points ahead of 2^{nd} rated had over twice the strike rate of those within 5 pts of 2^{nd} rated.
How ratings are compiled
The method by which ratings are compiled is a 3stage process. First we look at different factors that we feel might affect a horse’s chance of winning such as position in last race, age, official rating etc. The following chart shows the strike rate of horses depending on their position last time out (again for 5f sprints over the last 4 years). You can see that horses that won last time are more likely to win next time; and the further back they came in their last race the less likely they are to win next time.
However the dots are not always a smooth line. So the second stage is to fit a line to this data which can then be used to estimate strike rate from position last time out (see graph below). We repeat this process for all other factors (we currently look at over 20 including age, weight carried, trainer strike rate etc.).
The third stage is then to combine together all those factors we find important to produce an overall rating. This requires complex statistical modelling and is the stage that ensures our ratings have the two key properties outlined above.
How the ratings perform
One of the drawbacks of statistical modelling is that the model is designed to produce the best possible fit to the data it is given. What this means is that when the model is used to forecast forward, there is some reduction in accuracy. So, although the top rated strike rate for 6f was 20% for the model, the actual strike rate when the model is used “live” will probably be around 18%. This makes a strategy of backing all toprated horses less appealing since it reduces the likely profit.
The way around this is to analyse the ratings and find areas where the strike rate is high enough such that any reduction will still make a comfortable profit. For instance, backing top rated horses only when they are 10 or more points ahead of 2^{nd} rated. This has a modelled strike rate of nearly 35% which, even with a 10% drop in accuracy would still produce a healthy profit.
Our first attempt at ratings in 2003 produced an actual strike rate for toprated of 17% against a modelled level of 20% and actually made a slight loss. However backing top rated when 10 pts ahead produced a strike rate of 29% and 20pts profit from 69 bets (29% profit on total staked).
For 2004 the models have been improved further and the analysis more detailed so we are hoping for even better profits.
Paul Dyson 



