I am looking for some reassurance that I understand the sliding windows approach to time series modelling / training correctly.

From what I gather, when training a model, the dataset is divided into overlapping training sets of size k that are tested on test sets comprised of the next m observations proceeding each training set. I believe that that the test sets do not overlap? Hopefully the attached image from "I. Kaastra, M. Boyd /Neurocomputing" illustrates what I'm describing.

As far as I can see, this approach does not gauge the reliability of a single model, but rather tests the reliability that a model trained on k observations will be effective in forecasting the next m observations. This is unusual to me, as I am familiar with the training process being used to derive a single set of model parameters.

Any help in clarifying this is much appreciated.

enter image description here




1 Answer 1


You're correct, you are not getting only one model, but the performance of many. This is what happens every time we do cross-validation in usual (non time-series related) problems. In short, this is cross-validation.

The figure you posted depicts the usual way to do cross-validation in time series. The principle is that you should always evaluate your model with future data. You could also do that without superposition of intervals, but then you'd have weaker estimates.

  • $\begingroup$ Brilliant! Thank you Franco. :) $\endgroup$ Commented Jul 17, 2022 at 14:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.