Machine learning the hard way

[DeltaLover]

An interesting video about machine learning and horse racing, created by a novice bettor who still is touching a lot of the interesting problems we have to face when implementing such an approach...

xeAB10QgDW8

[raybo]

An interesting video about machine learning and horse racing, created by a novice bettor who still is touching a lot of the interesting problems we have to face when implementing such an approach...



Enjoyed the video, very amusing, and it definitely spoke to the difficulties we encounter while trying to profit from racing. I don't know when the video or the speaker's testing took place, but perhaps it was too far removed from the Hong Kong phenomenon, because he said that the real limiting factor was the size of the pools and the affect larger bets had on odds/payouts and degree of profitability. I suspect that he also did not understand the game even remotely as well as most on this forum do, like if you're wanting to make large bets then you only play the tracks with the largest pools. Of course, then you run into those Hong Kong guys more often, and we know that they are trying to do the same thing, bet more money.

However, he touched on exotics, zero times, and only mentioned place and show wagering as an alternative to win betting, saying of course, what most of us already know, that being that the payouts are too small to exploit those 2 pools, regarding market efficiency.

One thing that he did mention that I thought was very important, to almost anyone playing the horses, was that the historical data is "ordered", not random. Meaning that the data is grouped, by track, by date, by surface, by race number, etc.. So, when you run simulations, or you otherwise back-test historical data, even though you train with a certain set and test a different set of races, the data is still ordered", and then when you go to the real world, and you don't play races that are ordered similarly, your real world results are dramatically different than what your testing led you to believe. In other words, if you train/test on multiple tracks, at multiple times of the year, on multiple surfaces, multiple track conditions, etc., and then you go play for real at a single track, during a single time of year, on a single dirt and turf surface, your real world results suffer dramatically. This of course, leads some, me included, to believe that if you're going to use historical data for testing methods, you should at least do that testing on data that is similarly "ordered" as the races you will be playing: same time of year, same track, same track surfaces, same overall surface conditions, etc.. Which means that you necessarily must find a way to deal with small sample sizes of historical data. I know that viewpoint is not embraced by many, sorry.

Overall, I wish the speaker had had longer to take questions from the audience, and I wish there had been some serious horse players in the audience, as I thought that some of the questions were actually more interesting than what was presented in the presentation. I also wish it had not been so focused on programming and more easily understood by non-programmers. "Machine learning" need not be so formal, one does not need to be a programming guru to use computers to accomplish what the speaker finally accomplished, it just takes longer to get there.

:ThmbUp: :ThmbUp:

[Capper Al]

An interesting video about machine learning and horse racing, created by a novice bettor who still is touching a lot of the interesting problems we have to face when implementing such an approach...



Enjoyed the video. The high tech guys are tough to compete against especially if they get a backing of a whale.