I had to compare my noisy data against a product that was smoothed by the LOESS implementation in R. My take-home messages are:
- The speed of Savitzky-Golay is beautiful!
- Multiple passes with a smaller window preserve the shape of the original data (better than e.g. a single pass with a big window)
- The position of peaks in a derivative may shift with different choices of window length!
- I have the feeling that it does not make a difference if you first take the derivative and then do multiple smoothing iterations or first smooth and then take the derivative. But I did not check the math (or extensive tests)
So if in doubt, I will use 2-5 passes of an Savitzky-Golay filter with quadratic polynomials.