Can a Machine Learning algorithm outperform Hugh Taylor, lead tipster at ATR. Hugh regularly ends a year up to the tune of +30% ROI. An impressive haul but can a Machine Learning algorithm do as well if not better ?.

If you are unsure what I mean by ML then check out my earlier blog on this topic

https://markatsmartersig.wordpress.com/2018/10/07/form-reading-systems-and-machine-learning/

For the last couple of weeks I have been tweeting the predictions of an ML algorithm I have devised based on the ideas of big race trends. Trend analysis involves taking a few parameter and from past runnings of the race in question and applying them to the up coming race. For example we might find that no horse has won the Grand National in the last 10 years who is aged over 10 (please dont check this I have just made it up as an example) so we then throw out all horses aged over 10 and then move onto the next parameter. This approach then whittles the runners down to a shortlist or perhaps even one runner. It is a bit brutal in that you throw horses out rather than update their likely hood of winning based on a metric. ML might be an interesting approach to evaluating past big race data mainly because the data although sparse is at least specific to the race and ML takes a more nuanced approach to evaluating the parameters. An investigation would also answer the question, does this approach actually help find profit.

My intention was to keep the tweets going until the end of the year and see what is produced. One or two big races per week can be evaluated and by the end of the year I should have around 40 or 50 races analysed to Betfair SP.

My curiosity and impatience got the better of me especially after it had just found the winner of the Bet365 Gold Cup at Sandown. I decided to take a look at how the approach had faired in 2018 using data from 2009 to 2017

The code took about 3 hours to write and debug and the selection of races means trawling through the results to find the pertinent race characteristics to grab the past data. For example 14f, greater than 100k, handicap at York in August. So far I have run the code on 13 races from 2018 and the results are as follows.

Note the code produces a short list of 5 with ratings for each

Top rated bets 13 PL before comm’ +49.87
Top two rated bets 26 PL +43.3
Top three rated bets 39 PL +82.5
Top 4 rated bets 52 PL +69.5
Top 5 rated bets 65 PL +61

Among the winners top rated was Euchlen Glen at 31.0 in the John Smiths Magnet cup

I will update this post when I have entered a few more races for 2018 but at the moment there is some promise that the key to today’s big race lies in previous years races and Machine Learning.

FOOTNOTE – Hayadh 3rd top in the Thirsk Hunt Cup wins today at BFSP 21.0 🙂