That site has good examples of how to calculate impact values, but they fall way short in drawing conclusions based on too small a sample size.
5,000 races is a drop in the ocean, I wouldn't be happy with 50,000.
The guys such as the Reads and Bartholemews have sample sizes in the millions.

The reason for such sample sizes is because not all horses handle factors the same way.
For example, some horses carry weight better than others, so if the impact value of topweights were 1.10, you are over crediting or under calculating for some horses than others, so the prices or ratings will be quite out of whack.

Another example is last start winners.
if the impact value is 1.10 for won last time out, and .90 for didn't win last time out, then a last start group1 winner is going to be under credited and a last start hanging rock winner over credited.

There are so many variables, that drilling down within the data is the only way to go, which is why you need huge sample sizes.

Perhaps this is something I might have a crack at in the future, but it's a mammoth undertaking.

__________________
RaceCensus - powerful system testing software.
Now with over 426,000 Metropolitan, Provincial and Country races!
http://www.propun.com.au/horse_raci...ng_systems.html
*RaceCensus now updated to 31/10/2025
Video overview of RaceCensus here:
http://www.youtube.com/watch?v=W821YP_b0Pg