So often we hear that we must test a large number of selections. Well, shaun pointed to a link with a post from hermes that put forward an argument that we don't need to have a large sample at all.

I've been of the belief that we only need to have a decent sized sample on our base filter. If for instance you base filter is a Win% S/R then your sample size should be fairly large, but additional filters, provided they fit hand in glove with the base filter and each other, do not need to end up as a large number of selections.

It's logical really when you think about it, so your total number of selections could be as low as say 200, but your initial sample size may have been well over 1,000. The key I believe is compatibility of the filters with the base filter, and compatibility with each filters to the other filters.

Also, the notion of having as few filters as possible is ludicrous. You need 4 filters just to select an on-pacer over it's previous two runs, add several other filters and you're up to 8 without batting an eyelid, and none of it backfitted.

http://www.stat.wvu.edu/srs/modules...l/binomial.html

Have a read barny