InControlX
03-08-2011, 07:19 PM
In the past three years an unusual anomaly has surfaced in the results data of one of my three-race progression spot plays. The spot play defines a specific set of performance (horse pace profile in the two "prep" races) and artifact (trainer/owner selected race purse, distance, type) patterns that have traditionally been profitable but by no means a windfall generator. The anomaly arises when I look at individual trainer results for the spot. While expecting a typical bell-curve of trainer win percentages, peaking at the average, the data instead exhibits two distinct bells of roughly equal population, one centered at dartboard (15) percentage and another at a very high (45) win percentage using only trainers having at least ten tries at the spot with a total used sampling count of 252 trainers and 3750 races over the three years. Within the data set are an unusual, at least to me, number of trainers with 0-for-10, 1-for-15, and even up to 0-for-20 individual spot results while others are 8-for-10, 12-for-15, and one even is 16-for-16. Very few come in the middle. Adding the lower sample count (<10) trainers skews the percentages as I would expect, but not the overall bell averages.
Because it is a statistical sin to place arbitrary winning feedback in a prediction loop (and thus "invent" a wildly successful filter for past data), I've never filtered the play to specific trainers. But I'm going to filter them from now on for wagering and see how it goes. It's like there are two sets of trainers, those whom know how to prep the horse in this form cycle (and win) and those whom don't (and lose).
Anybody run across this before? Any ideas if something else is at work here? I’ve tested for days off, purse differences, month of year, tracks, state bred restrictions, age of entrant, final odds, race type and lots of other weird stuff to try to separate the trainer groups all with no correlation. I have a hunch it might be workouts (my database is poor in workout data) or unreported medications, but I have no proof.
ICX
Because it is a statistical sin to place arbitrary winning feedback in a prediction loop (and thus "invent" a wildly successful filter for past data), I've never filtered the play to specific trainers. But I'm going to filter them from now on for wagering and see how it goes. It's like there are two sets of trainers, those whom know how to prep the horse in this form cycle (and win) and those whom don't (and lose).
Anybody run across this before? Any ideas if something else is at work here? I’ve tested for days off, purse differences, month of year, tracks, state bred restrictions, age of entrant, final odds, race type and lots of other weird stuff to try to separate the trainer groups all with no correlation. I have a hunch it might be workouts (my database is poor in workout data) or unreported medications, but I have no proof.
ICX