Selecting smoothing parameters

[#1]

I have a large number of datasets where the abscissa is date/time and the ordinate is some response variable. I want to show variation over time as a smooth curve. I generally adjust the choice of fit type and the fit parameters (e.g. lambda) interactively until I get a smoothed curve that subjectively seems informative or a good fit to me.

But it's wholly subjective. Have statisticians thought through the question of whether there are objective ways of determining an optimum degree of smoothing?

I imagine optimum could be defined in a myriad of different ways, so maybe this isnt so easy to answer. If there is literature on this, I'd be grateful for some pointers.

[#2 · Accepted solution]

Hi, john_madden!

As you imagine, this isn't such an easy question to answer.

This might not be the best paper with which for you to start, but Grace Wahba has serious spline chops!  She wrote the book on it!!

Good luck.

[#3]

Thanks much Kevin. That paper is obviously pretty technical, but I'm going to see if I can pick some of the more intuitive ideas from it and move on from there as a starting point