It seems that the time complexity is O(2^S), where S is num_to_sample

[hzhangxyz]

so it is hard to do some parallel computing, is there some solution?

besides, could you mention it in the document? not only "O(N^3), where N is the number of historical points"