Regarding the issue of step_size

[zhizhuaa]

1. If step_size is set when defining the model, will the training proceed with movement based on step_size when calling the fit method for training?
2. Does the predict method have a step_size parameter? Why can't we input step_size in the predict method and let it automatically output multiple results, just like how hist_exog_list and futr_exog_list are automatically retrieved in the fit method.

[marcopeix]

Hello! I'm not sure I completely understand your first question. step_size in the model's initialization controls the distance between consecutive training windows. The default is 1, meaning that we maximize the number of training windows. So, if the first window starts at index 0, the next training window will start at index 1. If step_size=2, then the first window would start at index 0, and the next would start at index 2. So you will end up with less training windows.

For your second question, no, there is no step_size for the predict method. I think you're talking about specifying an arbitrary horizon and let the model do recursive forecasting if the horizon exceeds the training horizon. We're working on that at the moment.

[zhizhuaa]

The second question is about the following scenario: for example, if the input_size of the model is 2, and when the predict method is called, the length of the input df is 4. Why can't it automatically output the prediction results twice, but only the last one?

[marcopeix]

Because the input df serves as a single input to the model, so it predicts only the next h steps. If you want to take the first 2 steps as input, make forecasts, and take the next 2 as input, and make forecasts, then that's cross-validation.