Is it correct or incorrect to split data into train and test portions before using them on tumbling and sliding windows?

A user asks whether time series data should be split into training and test portions before or after applying tumbling and sliding window transformations for an LSTM predictor, noting they could not find authoritative guidance on the topic.

Written by AI from AI Stack Exchange's published text. Read the original for full details.