Improve cross validation to handle time structure

[dirkschumacher]

See here: https://stats.stackexchange.com/a/14109

[dirkschumacher]

Also see mase https://yardstick.tidymodels.org/reference/mase.html

[dirkschumacher]

I think the n-fold CV is a good default without good source

[stephaneghozzi]

I think it is necessary if, as is the case here, the predictive power of the model is to be evaluated. (A retrospective fit, looking in the whole data set which points are outliers compared to all others, would work fine with usual cross-validation.) One further reference: https://scikit-learn.org/stable/modules/cross_validation.html#time-series-split