I am attempting to derive a single multivariate/vector autoregressive (VAR) model from a large dataset (6-minutes sampled at 250Hz in total w/ 50 vars) using cross-validation (CV) to optimize model-order. To run CV, I first split up the entire dataset into non-overlapping epochs containing 350 samples each. I then randomly split the epochs into train/test sets for each fold of CV, and train one model per CV fold. Just to be clear, the training process for a given fold involves training one epoch at a time rather than concatenating epochs (which would likely introduce non-stationarity given that epochs are randomly selected for each fold).
Regarding my actual question, I am wondering if I should evaluate whether or not my data is stationary when it is segmented into epochs (350 samples). I know it is essential for all variables in an VAR process to be stationary (see: https://www.researchgate.net/post/Is_it_necessary_to_ensure_stationarity_of_all_time_series_variables_when_you_run_a_Vector_Autoregressive_VAR_Model), but I'm not sure if, in practice, I want to ensure that the variables are stationary within each short epoch of data. Since the end goal is to derive a single model trained on all 6 minutes of data using the most optimal model, my intuition says that I should only worry about the overall process being stationarity (i.e. all 6 minutes). Is this intuition correct or not?