I do not understand this API with initial=21, what is the initial=21 doing here?

Also the default test size is 7 and there is only one sample in the test set.

Thanks for the response. I should explain this better.

Assume the blue dots represents dates to train on and the red dots are for testing.

In this picture the initial size of the blue dots are much larger than the size of the testing dots. (6 blue dots vs 1 red). With each iteration train size increases by 1 blue dot.

My PR allows this kind of flexibility. A user may set a larger initial training period, which is not bound by the size that the training set is increases by on each iteration. It decouples the initial size from the increment size.

Oh and 21 represents 7 days a week times 3 (a three week training period by default). I am assuming people could use this function in a forecast setting for daily forecasts.

I am assuming people could use this function in a forecast setting for daily forecasts.

This assumption is not always valid. I've seen the time series splitter used for hourly data, etc.

In that case with initial=21, should the first split have training indices [0, 1, 2, 3, 4, ..., 20], and the test set be [21, 22, .., 27] because test_size=7?

It looks like you want a walk forward as in #23780 but with an additional min_train_size.

That is fair enough. I am not married to initial=21. You let me know what is preferred and I will make the change. The functionality is much more important to me than the initial args.

It looks like you want a walk forward as in #23780 but with an additional min_train_size.

I am afraid that min_train_size means that I want a consistent train size. That is what I do not want. I want one train size for the initial training then a second train size for each iteration.

If you sign off on this then I can keep working. I remember seeing a contributors document. I would have to see what the expectation is for writing tests and documentation.

I am not signing off on this because the API is not finalized and unclear. Let's move the discussion to the issue.

After the API is clear, we can evaluate if we want this in scikit-learn.