Let's say I have a set of 1D time series which values have been samples in equip-distant time steps with timestamps $1,2,3,...$, they have all the same lengths and are somewhat similar in shape. I want to apply non-parametric regression (e.g. with Gaussian Processes or Kernel Regression) on the time series in order to infer values for timestamps that are between sample timestamps (e.g. $5.3$).

The obvious way of doing this would be to simply build a regression model for each time series. Overfitting, if a single regression model is build for each time series, can be avoided by using established techniques in using the right smoothing parameter in the respektive kernels.

Is there some game-changing merit in taking the entire time series data and applying the regression model to the entire time series dataset? So far I only see the disadvantage of over-smoothing (aka underfitting).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.