0
$\begingroup$

If you are about to answer "because k-fold incorporates future information", I'm going to challenge you on that answer :)

If my time series exhibits some pattern (e.g. annual seasonality), that pattern must obviously occur every year, in the past and in the future.

For example, if some summer seasonality appears in the 2017 data, it should also appear in 2010, or in 2007, etc. Otherwise it's not really a pattern!

But if seasonality appears every year, then training the Model on chunks of future data and validating the Model on chunks of past data, should be fine as well.

So cross-validation rocks on time series too, right?

Improve this question
$\endgroup$
5
  • $\begingroup$ CV works on random splitting of the data, so among (many other things) it's not clear that you would be able to split the data so that seasonality is preserved. $\endgroup$
    – meh
    Commented Sep 11, 2017 at 17:18
  • $\begingroup$ ok but that would make CV more pessimistic, but that's not too bad. The problem is usually the opposite ... being too optimistic. besides K-Fold is no based on splitting the data random, rather in K chunks :) $\endgroup$
    – elemolotiv
    Commented Sep 11, 2017 at 17:23
  • 1
    $\begingroup$ See this answer. Let me know if your question should be marked duplicate of the other one. $\endgroup$
    – Richard Hardy
    Commented Sep 11, 2017 at 17:41
  • $\begingroup$ @RichardHardy great link! let's mark it as duplicate, thanks :) $\endgroup$
    – elemolotiv
    Commented Sep 11, 2017 at 18:13
  • 1
    $\begingroup$ See robjhyndman.com/publications/cv-time-series $\endgroup$
    – Rob Hyndman
    Commented Sep 11, 2017 at 21:15

0

Reset to default

Browse other questions tagged or ask your own question.