What % is the best train/test split for Time Series Data?

Question

What % is the best train/test split for Time Series Data? Do you think it is still 70/30 ?

I am talking about leaving a whole continuous period (in the beginning or in the end) for TEST.

I have two intuitive assumptions:

• For Time Series Data it is better that the split is more weighted towards the train data because of interdependence. Maybe 80/20 or 85/15.
• If you are to choose from leaving a TEST period in the beginning or in the end probably in the beginning is better, because the latest data brings more 'information'.

Please give your opinion on the validity of the assumptions. Both a theoretical and 'from experience' point of view.

Answer 1

Regarding your first assumption, maybe you could check the minimum value where the prediction error stabilizes. What I used to do with classification algorithms was to determine the optimal percentage of training/testing by doing the training with 5%, 10%, 15%, 20%, etc., calculating the misclassification error on each case, and plotting them in order to find that minimum value.

About your second assumption, your statement is probably true, but usually, when you fit a time series model, you are interested in predicting the future, what is yet to happen. If you test your algorithm with the latest data you will be more confident that the resulting predictions are more accurate.