Due to the very small sample I ran several random combination of crashes and optimized the standard deviation (so sample 1- was 2010,2019,2015, sample 2–2010,2014,2016). after running about 10 combinations the optimal distance came as 5 standard deviations across the different assets