I'm analyzing a time series (terms of trades) on which I want to perform a trend estimation by nonparametric methods like the above mentioned. By the way, I'm a total beginner with R and using the help files is already cryptic enough. Using the default settings of R gives me a smoother which simply follows the curve of the original series, although I thought that by using the cross validation method (which is set to default in R) optimal parameters would be used, i.e. which are a perfect trade-off between smoothness and best-fit. By comparison, I set df=8 which was a much nicer curve, however randomly choosing values isn't really scientific. Can you tell me what's the correct procedure here?
mgcvpackage that comes with base R. There is a learning curve, but it generates smoothers that are optimal under a few choices of loss definitions. If you just want a picture that looks good, try a few different bandwidths forloessand call it good. – Shea Parkes Oct 28 '12 at 16:47mgcvis the only one that comes to mind; it offers Generalized Cross Validation (GCV) and REML to optimize the curvature. You could define your own loss function and use cross-validation to test different smoothers at different tuning values to see what is "optimal." Be warned that using cross-validation to both tune and choose between multiple methods will overfit your training data without complicated nesting. – Shea Parkes Oct 30 '12 at 0:56