When applying windowing techniques in time series or sequential data moving windows takes windows of a fixed length starting at every position, for example [1,2,3,4], [2.3.4.5], [3.4.5.6], etc. This is in contrast to non-overlapping windows.
Moving windows are often used when wanting to make a prediction of the item following the window, for example, predicting data point 5 based on points 1 to 4. In time-series analysis moving averages do this with simple linear predictors, often over a small number of past observations. Large-language modles use the same technique but with a deep neural network to perform the prediction and a window size that is many thousands, if not millions of tokens long.
Used in glossary entries: deep neural network, non-overlapping windows, time series, windowing