Twenty years ago, before he had access to computer technology, Rich Janava, only imagined being able to predict the order of finish in a six furlong claiming race. As a teenager, his father dragged him to the races to perform the legwork of running money to betting windows. Over the years, Rich watched his father lose a small fortune. Armed with the belief that predicting the most probable horse in a race should be easier than putting a man on the moon, and having seen too much to be a gambler, he set to work on a solution. An engineer by profession, Mr. Janava worked on the problem for seven years using calculators and some early IBM 360 machines, but a major obstacle was the need for better technology enabling the required non-linear optimized solutions.
Four years ago after being exposed to neural net analysis, Rich discovered the tool required to accomplish his goal. Using BrainMaker Competitor and Lotus macros, he has developed and automated a method for predicting the order of finish of six furlong claiming races at a Philadelphia race track.
So far in the first 300 races, 39% of the winners have been predicted at odds which average better than 4.5 to 1. The real power is in three horse box exactas and four horse box triple wagering which are both hitting better than 35% of the time. Rich says the BrainMaker Competitor formulation is the only software he is aware of which sets up the fact files in the manner necessary to formulate and solve this type of problem. It also provides the convenient ranking of results for the many hundreds of races necessary to train and solve this problem.
According to Mr. Janava, the key to his success lies in the fact that he has limited his focus to one type of race and one race track. As any horse player knows, every type of race is affected by different racing variables. Every race track is also different. So far Janava has thoroughly explored and compared nine major race tracks. He explains, "If you try to generalize too many types of races and race tracks, you will lose the fine edge necessary to be successful at any race, distance, or track." In other words, the key to success is to become a specialist.
To make a generalized prediction, Mr. Janava analyzed hundreds of quality six-furlong races using the standard Racing Form paper. He started with an initial set of over 50 variables and determined that 24 variables were statistically significant. These include variables related to speed, horse position during previous races, class, earnings, recent activity, in-the-money percentages, and post position in today's and previous races. Several variables in each category are also combined to yield composite variables which extend the total number of variables considered for each horse to 31.
After much thought, Rich derived an approach to calculating probability functions for each significant variable. For fact file generation and training, non-linear-regression-fitted probability values are used as inputs to Competitor. While using raw data values instead of probabilities for training has not been explored, Rich believes that raw data should work. In order to produce a sufficiently generalized network, Rich trains using at least 200 races.
Rich says that, had it not been for the BrainMaker software, his efforts would have been nearly impossible. He intends to constantly refine his analyses with additional races during his yearly winter recess, and looks forward to a BrainMaker Competitor version that runs race fact files in batch mode.
INPUTS:
speed
horse position during previous races
class
earnings
recent activity in the Monty percentages
post position in today's race
post position in previous races
OUTPUT:
order of finish