Black Box Output(Ratings)

[bucktron]

Much attention is placed on input variables in a Black Box System(neural nets, genetic algorithms, multiple regression , grnn, etc.). Little is mentioned of what developers are trying to predict. A quality Rating reflects a balance of all critial output factors and can greatly influences the accuracy of your projections given the same data set. Projections at different Distances and Surfaces are more accurate using certain combinations of output factors versus others. What do you use as output?

Example: I have found that projecting the horses Speed Rating is not nearly as accurate as a projection based on Finishing Position and Lengths Beaten for Routes run on the Turf.

Inputs ............ Output

Variable(1) ---> Rating = (300-((Lengths Beaten * 2) + Finishing Position))/3
….
Variable(n)


I would like to hear thoughts or ideas the subject of Output/Computer Generated Ratings.

[Tadek]

bucktron

I am using a rating system, which computes comparative speed values and comparative workout values. Then I conglomerate both with a Factor Influence parameter defining relationship between both elements.

I apply this combination to a sufficient number of races on a track I want to play. I repeat the same evaluation with different values for Factor Influence parameter. I use the Speed factor as an anchor – factor, which has its Factor Influence, set to 1.0; the Factor Influence of Workouts is varied from 0.05 to 1.0 in steps of 0.05.

The results of those computations showed that the number of winners as a function of the Factor Influence parameter varies up to 23% - one example at Woodbine was 97 to 74. Profits vary much more substantially, in the above example:
Speed, Workout(1.0-0.2) – 97 winners $63.40 profit 364 races
Speed, Workout(1.0–0.6) – 78 winners -$124.00 loss 364 races

I’ve repeated those computations many times for different tracks and different profiles of races. I agree with you bucktron that the proper balance of factors is absolutely essential for a successful rating system.

Unfortunately, a properly balanced method for Woodbine (last season) proved to be completely useless for Belmont (this year). Before I start playing a meet I gather about 100 races and perform factor influence analysis and then, while the season progresses, I recalculate factors every week.

Regards

Tadek

[osophy_junkie]

Originally posted by bucktron
Much attention is placed on input variables in a Black Box System(neural nets, genetic algorithms, multiple regression , grnn, etc.).
These aren't black box systems, it is well know how they work.

Little is mentioned of what developers are trying to predict. A quality Rating reflects a balance of all critial output factors and can greatly influences the accuracy of your projections given the same data set. Projections at different Distances and Surfaces are more accurate using certain combinations of output factors versus others. What do you use as output?
That is the fun of it :) Feature selection is an incredible difficult problem, with the easist solution being adding human knowledge to the algorithem being used.

Example: I have found that projecting the horses Speed Rating is not nearly as accurate as a projection based on Finishing Position and Lengths Beaten for Routes run on the Turf.

Inputs ............ Output

Variable(1) ---> Rating = (300-((Lengths Beaten * 2) + Finishing Position))/3
….
Variable(n)


I would like to hear thoughts or ideas the subject of Output/Computer Generated Ratings.

This is highly dependent on the system being used. I've used a NN with much ignorance in the past with little success. I'm currently useing a statistical learning system for regression and classification with much success. I use it for extrapolating pars, and finish position prediction.

Ed