Documentation on cross validation parameters (h, step_size, n_windows).

[ABJ66]

Hi all,

While using statsforecast, I was not able to understand how the cross validation parameters of h, step_size and n_windows (aka no. of cutoffs) work together. Hence, tried running some experiments with various combinations of the above and came up with observations and a reference table.

It would be of help if the authors/maintainers/user community can verify the same. Note: I'll be using integer representations of date-stamps/'ds' for all relevant columns as seen in the below reference table:

1. History end (ds) is & horizon 'h' are the anchor points for computing various other columns in the table. I had 57 time points per series in the data, hence History start = 1 and end = 57 for all trials.
2. "Cut-off max": Is the latest/max cut-off 'ds' integer representation and is calculated as (History end - horizon 'h')
3. "Cut-off min": Is the earliest/min cut-off 'ds' integer representation and is calculated as (Cut-off max - ((n_windows -1)*step_size). We reduce 1 from n_windows as we have already accounted for cut-off max.
4. "Consecutive Overlap": Horizon 'h' - step_size (this denotes how many 'ds' points are common between a certain cut-off value and its previous value).
5. "Cut-off min ds min": Is the minimum 'ds' for the minimum cut-off and is calculated as Cut-off min + 1 (as its supposed to start immediately after the cut-off point).
6. "Cut-off min ds max": Is the maximum 'ds' for the minimum cut-off and is calculated as Cut-off min + Horizon 'h' (as its supposed to generate predictions till this 'ds')
7. "Cut-off max ds min": Is the minimum 'ds' for the maximum cut-off and is calculated as Cut-off max + 1
8. "Cut-off max ds max": Is the maximum 'ds' for the maximum cut-off and is calculated as Cut-off max + Horizon 'h' (these 2 calculations are applicable for all cut-off points irrespective of min or max and in-between).
9. Additionally, the "Cut-off max ds max": Also happens to be the same as History end.

Hope this helps. Look forward to your comments!