Robustness of move_std

[jaimefrio]

The algorithm used by move_std is prone to numerical errors. A while back I rewrote the implementation of the equivalent function in Pandas to use what some call Welford's method, a one pass stable algorithm, see here for the relevant PR. I took some timings on the Pandas code, and it resulted in a 25% slowdown.

Would you like to see something like that implemented for bottleneck?

[kwgoodman]

Sounds like a good idea to me. But I'd like to hear what others think. Anyone?

bn.nanstd uses a two-pass algorithm. If the bottleneck is memory access, which I believe it is, then perhaps the one-pass algorithm you are proposing is faster. If so, assuming comparable stability, then maybe the place to start is with bn.nanstd?

[jaimefrio]

If you are not frontally opposed to it I'll put a PR together, and we can discuss its merits after it is up and running.

[kwgoodman]

That's great. Is this for move_std or nanstd?

[jaimefrio]

I think it is only clearly advantageous in move_std, the two pass algorithm is typically faster, and probably even more accurate, for a non-moving window.

[shoyer]

Indeed, this seems like a nice improvement 👍.

[jaimefrio]

What should go into the PR? Just the move,pyx file in templates, or the generated C files as well?