KernelReg is slow, and with bandwidth selection it is even slower. My guess is that you don't need the extra features of KernelReg and a faster substitute would work better in this case.
KernelReg is doing the estimation loop in python and is not optimized for speed.
Also, if you do bandwidth selection with one of the cross-validation methods, then doing the bandwidth selection only once and reuse for other series would be faster, if all profiles require approximately the same amount of smoothing.

First, can you add or post somewhere one of your datasets/profiles, mainly because this one of related type of usecases that I want to look into (eventually).

For related smoothing:

• lowess is faster and should work pretty well if only the smoothed series is required.
• if the x points are evenly spaced on a grid, then scipy savitsky_golay should work and be very fast.
  (I have an open PR for local smoothing that works similar to savitsky_golay but does binning if the x-values are not evenly spaced.

without smoothing
even without smoothing a nonlinear fit should work quite well, if not with a standard least squares fit, then a robust fit might work. scipy now also allows for Huber robust regression in the extension of leastsq fit, maybe lmfit includes this. I haven't tried those and so I have no idea how well they work for outlier-robust nonlinear fitting.

(There might be also other ways to estimate this in a relatively robust way, e.g. assuming Poisson noise, but I don't have any experience in what works well in cases like these.)

I just saw https://stackoverflow.com/questions/48069998/kernelreg-performance-in-a-for-loop
which includes one example profile

Yes.
I posted the example to https://stackoverflow.com/questions/48069998/kernelreg-performance-in-a-for-loop.

You can also post your answer to stackoverflow.

I have a solution to speed up bw estimation for Nadaraya–Watson kernel regression (local constant reg_type). The idea is to vectorize the cv_loo method, and I can achieve in general 2x~2.5x speed up for hundreds to a thousand rows of 3-dimensional data (can't do more rows/dimension because my current machine is basically a tablet). For Gaussian kernel and continuous var_type, I can achieve 50x speed up (a few hundreds row) and ~7x speed up for 1000 rows. This obviously do not solve all the performance issues with kernel regression as there is no change to the fit function, and it only works for local constant reg_type. Can I still raise a pull request for this partial solution?

Please open a PR.

Should I tag someone to review this? Any comment would be appreciated.