Fixing HHFs

[CodeHowlerMonkey]

So how do we fix current HHFs we have on file? We have both classifiers that are too easy and are GM factories and ones that are literally impossible to class on.

Here's an idea...

Pretty sure the main principle I stole from reading your competition.shooting.analytics on IG. TLDR: we look at real A-class shooters percentiles and use that as a calibration point for 75% scores for each classifier.

Quick algo:

1. We take table of shooters in division and calculate their percentiles based on curPercent (not classification or high percent). This gives us approx 1/5/15percentile for G/M/A (1.1, 6, 14.7 for CO currently)
2. This gives us points for calibration. So far A-class looks like best calibration point — it’s high, yet has a lot of people/scores compared to M/GM
3. We take that calibration point as percentile and say that this is where 75% score should land. Meaning that hundo should be set as thatHF / 0.75 of that score closest to our target percentile.

Result we have easy classifiers made harder and hard classifiers made easier. I don’t really know statistics but basically we’re smoothing HHFs out for existing classifications system has, with some bias towards realism, since curPercent also accounts for majors in some shooters.

Here are two examples with hard and easy classifiers in CO. r15 prefix is that A class targeting: