Time Split#

Time-based k-fold validation splits for heterogeneous data.

Experimenting with parameters#

The Time Split application (available here) is designed to help evaluate the effects of different parameters. To start it locally (requires the time-split[app] extra), run

python -m time_split app start

in the terminal. Alternatively, you may run

docker run -p 8501:8501 rsundqvist/time-split

to start the application using Docker instead.

Parameter overview#

Overview of parameters used by time_split.split() and plot(). Integrations such as split_pandas() may add or remove parameters, but the base function remains the same unless otherwise stated.

Name	Default	Type	Description
`schedule`	N/A	Valid `Schedule` types: `DatetimeIterable` pandas offset alias cron expression	Generates training dates (`DatetimeSplitBounds.mid`). Examples: `['2019-05-04', '2019-05-11']` \| Hand-picked dates. `'7d'` \| every 7 days, aligned to the end of the available data. `'0 0 * * MON,FRI'` \| every Monday and Friday at midnight.
`before` `after`	= ‘7d’ = 1	Valid `Span` types: pandas offset alias Literal ‘all’ [1] Literal ‘empty’ `int >= 1`	Range before/after schedule timestamps. The default `after=1` stretches the `Future data` until the next schedule timestamp, simulating models staying in production until a new model takes its place. That is, `fold[i].end = fold[i + 1].mid` for `after=1`.
`step`	= 1	`int >= 1`	Keep every step:th fold in the schedule. Default (1) =keep all.
`n_splits`	= 0	`int >= 0`	Maximum number of folds. Default (0) =keep all.
`available`	= `None`	`DatetimeIterable`	Limits `(min, max)`, or an iterable of datetime-like types that support the built-in `min()` and `max()` functions. Binds schedule to a range.
`expand_limits`	= ‘auto’	Valid `ExpandLimits` types: Literal ‘auto’ [2] `bool` ‘round_to[<tolerance]’	Expand available data outward to its likely “true” limits. Disabled if `False`, `True == 'auto'`. Use `expand_limits()` to experiment. Example: Passing `expand_limits='d<3h'` expands `(min, max)` -limits (derived from available) to the nearest day, at most 3 hours from the original limit. The tolerance argument is optional; `expand_limits='d'` performs regular `floor(min)` / `ceil(max)` expansion of the limits.
`filter`	= `None`	`Filter` or `str`.	A callable `(start, mid, end) -> bool`. Strings are converted using `get_by_full_name()`.

Later folds are always [3] preferred. For more information about the schedule, before/after and expand_limits-arguments, see the User guide. See the Examples page for plots using the various parameter options.

Footnotes