time series indexer about mlens

[flennerhag]

Agree. A make_indexer function would be a great utility to convert any list of train-test splits of fold generator into a mlens compatible indexer. Actually think the mlens.index code base (in 0.2.0) already has the base code needed for that, worth looking into!

Come think of it, the time series indexer is quite straightforward to implement as the folds are contiguous:

# Params
require: burn_in
require: window_size
require: test_size

start_index = 0
stop_index = burn_in
stop = False
while not stop:
    start_index += window_size  # if moving window
    stop_index += window_size

    if stop_index + test_size >= X.shape[0]:
        stop = True
        # Something here to grab the last obs in the final fold

    train_index = (start_index, stop_index)
    test_index = (stop_index, stop_index + test_size)  # test_size might not always be 1?

    yield train_index, test_index

or something like it maybe.

[JoshuaC3]

This would be great.

In addition, it would be useful to add in a base training size e.g. 3 (of the 6 "folds"). This better mimics many real life scenarios where you have a good amount of training data with an expanding window.

fold	train obs	test obs
0	0, 1, 2, 3	4
1	0, 1, 2, 3, 4	5
2	0, 1, 2, 3, 4, 5	6

It would also be useful to include a moving window size, e.g. 4 (on 6 folds)

fold	train obs	test obs
0	0, 1, 2, 3	4
1	1, 2, 3, 4	5
2	2, 3, 4, 5	6

Ultimately, this could all be achieved by allowing a user to define their own indices for the train-test-split.

[flennerhag]

Resolved in #95.