Commercial Software Systems

[bucktron]

System Sellers standard buzz sentences:

[list=1]
System generates one or many figures to base your betting decisions on.
Users are given and also may develop strategies that should generate profits.
The System will provide varying degrees of success depending on the users experience and expertise.
[/list=1]

System Sellers, I have several questions:
[list=1]
Have any of you Tested your systems output on a viable Sample, lets say 10,000 races, that the System had not seen before?
What were your results?
Did you find that by increasing your Training Set Size that your Backfitted Equations/Model more closely approached your Test Set Results?
If anybody has done this, even if you are not profitable, I would like to hear about your methodologies.
[/list=1]



Thanks,

bucktron

[Derek2U]

I really dont think MOST try to replicate their data; instead, they
keep on tweaking their data dry (backFit) and then proclaim
good results. To me, the best thing is this: Start with ideas;
measure those ideas; throw-away the sample on which you
tested-measured your ideas; get a fresh sample; validate those
ideas using the fresh sample; use this method: Suppose your
fresh sample is only 5,000 races. Don't use all 5k, but instead
take 200 RANDOM races & plot the results; take another 200
RANDOM sample & plot; keep doing this maybe 100 times & then
report on the DISTRIBUTION on the 100 x 200 races.

[Handle]

EquiSim has the ability to track how well its various strategies perform built into it -- its part of the program and its one of its driving concepts.

The S.P.A.T (simulation performance analysis tool) lets you query how well the simulation winner does in many different cases.

The Profiler allows you to track how well various attributes do based on past races that are similar to the upcoming race (similarity is configurable through filters). Attributes (called abilities in ES) are assigned by the software to each horse based primarily on the selected pace line (mechanically selected or user selected).

I've done well by looking for race types where factors are strong, since I know that no "backfitting" or whatever is used to produce the stats. That's the primary concept -- the software does X. It can't do X well everywhere. So you have tool Y (the statistical analysis stuff - Profiler/S.P.A.T) to show you were its performing well.

-Nathan