Skip to main content
5 events
when toggle format what by license comment
May 19, 2020 at 17:37 vote accept Santi Peñate-Vera
Dec 16, 2016 at 13:20 comment added Santi Peñate-Vera Imagine you have 4 weeks data in hourly steps. To test the method you pick 3 weeks to train and the last week to forecast. If you shuffle the 4 weeks data into train and test sets, you'll have data from the fourth week in the train set, hence hours from the 4th week are used to predict other hours from the fourth week having those hours a great similarity. By not shuffling we won't have any info about how the hours in the fourth week look like.
Dec 16, 2016 at 12:05 comment added einar Plus also, you could look into the block bootstrap and its variants if you want to do resampling/data perturbation/"shuffling" in a time series setting
Dec 16, 2016 at 12:03 comment added einar I'm not sure I 100% understand your answer, is it that in the "non-shuffling" case you're trying to predict the future, but when doing shuffling you're just trying to fill in missing points in the past?
Dec 16, 2016 at 11:34 history answered Santi Peñate-Vera CC BY-SA 3.0