polydiff fits a consistent polynomial order to a window of some number of samples, which is fine if there is just slight jitter in the data but not fine if there is big density somewhere, because then the model wants to overfit noise in that dense region compared to elsewhere in the domain.
Discovered while trying to get Claude to extend lineardiff to handle variable step, which also depends on slide_function. Other models that support variable step--splinediff, rbfdiff, rtsdiff, and robustdiff--do not suffer this problem, because they work globally or by evolving over t, not working merely with a set number of samples in windows. I'm going to push the changes that make lineardiff more like polydiff and then try to get a fix to both of these at once.
This one continues #173, revisiting a concern I swept under the rug at the time. One of the more annoying curveballs is you could end up with too few samples in a fixed window.
polydifffits a consistent polynomial order to a window of some number of samples, which is fine if there is just slight jitter in the data but not fine if there is big density somewhere, because then the model wants to overfit noise in that dense region compared to elsewhere in the domain.Discovered while trying to get Claude to extend
lineardiffto handle variable step, which also depends onslide_function. Other models that support variable step--splinediff,rbfdiff,rtsdiff, androbustdiff--do not suffer this problem, because they work globally or by evolving overt, not working merely with a set number of samples in windows. I'm going to push the changes that makelineardiffmore likepolydiffand then try to get a fix to both of these at once.This one continues #173, revisiting a concern I swept under the rug at the time. One of the more annoying curveballs is you could end up with too few samples in a fixed window.