Hi all,
I am wondering about estimating predictive variances for regression splines. In particular, I know that I could fit a cubic spline to predict the mean given some input, but what if I am interested in the output likelihood. I know that gaussian processes do this, but for reasonable size datasets with ~5000 points, gaussian processes take a long time.
Why not simply fit a local variance using another cubic spline, such that the predictive gaussian distribution is then parameterized by two local estimates.
This seems wrong, but I'm not sure why. Does anyone have any suggestions for how to think about this?
[link][2 comments]