Greg Brandt
Greg Brandt
@hntd187 Nice I will check those libraries out, and thanks for the suggestions. I agree that we probably won't hit Fortran speeds, but my intuition is that we could probably...
@hntd187 Sounds good to me. It might also be interesting to run something like [f2j](http://icl.cs.utk.edu/f2j/index.html) on the original STL fortran implementation, and see if we can wrap the result in...
@hntd187 Yeah, good point - I'm not really sure what the eventual goals are here. The original goal was just to have something comparable to R to run in MR...
Also FWIW running on smaller data sets, e.g. couple months of hourly data with weekly trend, does not have this performance problem, so it may not be worth it to...
@hntd187 Nice, we should give this a shot and see if we can eliminate the order of magnitude performance discrepancy. Maybe make another config in `StlConfig` that can switch between...
I was able to reduce the number of inner loop passes from 10 to 2 (R's choice) after commit 8bbc947a621eb51fcdea8ce124a6591934cbebf9. The cycle series were not being properly padded, so we...
Here's the data in CSV [hourly.txt](https://github.com/brandtg/stl-java/files/187090/hourly.txt)
@jcrotinger I see, the skip points thing may indeed be the reason. Do you think that it would be worthwhile to submit a patch to the commons-math LoessInterpolator? Or would...
@jcrotinger Got it, a port of the original version makes sense
@jcrotinger Sorry I was out sick for a couple days, so didn't have a chance to get to this. But thank you for the awesome analysis! We have been wondering...