The span corresponds to the number of data points the algorithm gets to see before making a prediction. The donut model uses a fixed window that slides over historical data to learn a meaningful representation.