splits.rolling_forecast_origin()

Create rolling forecast origin train/test samples.

Usage

Source

splits.rolling_forecast_origin(
    data,
    horizon,
    step,
    min_train=365 * 2,
)

Samples are generated from the most recent data backwards. The test period moves back by step days for each fold.

Parameters

data: pd.DataFrame

Data containing a ds column.

horizon: int

Number of daily observations in the test set.

step: int

How many days to move by before creating a new sample. Warning: using a step of 365 will produce test samples all at approximately the same time of year.

min_train: int = 365 * 2

Minimum number of days to include in training sample. By default, set to 2 years as that allows detection of yearly seasonality.

Returns

train, test : tuple[list[pd.DataFrame],list[pd.DataFrame]]

Training and test dataframes for each fold, ordered from most recent to oldest.