Quote:
Originally Posted by Nitro
So, apparently you support the absurd concept that viewing the running lines of a previous race without consideration for any other race related factors will lead to drawing better conclusions when contemplating a future race? The only basic validity this has is that you know that the race has already been run, and where each of the entries were located at various points on the track. Now trying to assume that these running lines will in some incredible way provide “valuable” information pertinent to a future racing event for any of the entries previously involved is bizarre at best...
|
He never said without consideration for any other race related factors.
Quote:
Originally Posted by steveb
If you model the rankings at various points of the race, then you can glean valuable information.
That other ways may yield better results, does not make it any less valid.
|
Having done it I'm pretty sure it's worthwhile to create algorithms that score running line data (positional calls, beaten lengths, gain or loss between calls, etc.), normalize the scores - and create running line factors from the normalized scores.
Then include some of your running line factors in a fundamental model - with many other factors.
Benter wrote about doing this (with time decayed positional calls) in one of his presentations.
Quote:
Originally Posted by Dave Schwartz
It's not enough to win, so I'd say not.
And the calcs are monstrous for someone using paper & pencil.
But your point is not too far off in the sense that doing what everyone else is doing isn't enough.
I used to think that meant "do something different."
Now I believe it means "do more." That is, add another step of analysis.
|
The way it often works (but not always you have to test it) if you add something new to an already decent fundamental model --
For example factors based on positional calls data --
And provided the existing model doesn't have anything in it with a strong correlation to what you just added:
You tend to get a small incremental improvement in the performance of the resulting new model vs. the old model.
-jp
.