The initial results are promising with concentration ratios at 200% to 300% before running the optimizer. This is running against what is supposed to be randomly moving price data so classification at this level should be pretty difficult.
Since this is before running an optimizer it cannot be over fitting the curve since it is doing no fitting at all.
I pushed the KNN in some different directions and have it produce both a probability and magnitude score. Not sure yet how accurate the magnitude will be.
To ensure we didn’t cheat we filter all KNN data from earlier less bardate – (goal + 1 days) prior to day being ranked. This absolutely eliminates risk of cheating by looking into the future.