@freemo i should say, that's for a hash set. so the cost of the 'not' should be the only diff

@2ck Yea, normally I can get pretty close to optimal before i actually do performance profiling. But with python I was way off, things that normally would have virtually no cost to it seemed to be quite expensive for me. In the end I relied on the profiler more than an algorithmic understanding to fix it, which I dont like to do but I kinda had to in this case. In the end it worked though, its now fast enough that I can backtest a year worth of day trading in just a few minutes.

Your description of what you find yourself doing when you optimized seemed to match my own expiernce. which makes sense since in python an attribute/property may actually be a method so it really doesnt tell you much about if reading it is atomic or not. But the collections were also a damn mess and really hard to tell what algo backed the various collections I was using since they mostly came from third-party libs. a good example of this being numpy series which seemed to be any of several collection types backing it and the constructor was rather convoluted.