Why is sliding window evaluation important in time series analysis?

I have been working dynamic graph neural newtork survey, and what I realized is that all of the well known paper (from pretegious university) do not use sliding window evaluation on dynamic graph model. They only use simple train-test splits.

I find this very confusing. Then I start asking question why sliding window is important in time analysis in the first place.

From my own experience, I know for a fact that dynamic graph models are VERY VERY sensitive to window size, batch per window, and epoch for window.

I don't know too much about time series analysis, but I suspect that sliding window evaluation must have sometimes to do with time series property.

Any thought?

Thanks in advance