cv_split_temporal: Special resampling strategy for K-fold cross-validation on...
In statist-bhfz/resampleR: Easy resampling with data.table

Description Usage Arguments Details Value Examples

Special resampling strategy for K-fold cross-validation on time series data with stratification by target variable.

1 2	cv_split_temporal(data, y, id, time, nfolds = 5L, probs = seq(0, 1, length.out = 11))

`data`	data.table with `y`, `id` and `time`.
`y`	Target variable name (character).
`id`	Identifier of each time series (character).
`time`	Time variable name (character).
`nfolds`	Number of folds (min 2, max 20).
`probs`	Numeric vector of probabilities for quantile binning with values in [0, 1] range.

Numeric target: quantile binning is used for stratification.

Character/categorical target: resampling performs within categories.

probs can be a vector like c(0, seq(0.99, 1, length.out = 10)) for target with very skewed distribution, e.g. for financial data with 99% of 0's.

When some observations from one time series fall into validation fold, train/validation indices for this time series will be reassigned: only last observation will be in validation fold. This ensures that training performs on past data and predictions are made for future observations.

TODO: allow to specify arbitrary number of observations for validation set.

data.table with nfolds columns. Each column is an indicator variable with 1 corresponds to observations in validation dataset (stratified by target).

dt <- data.table(
    user = rep(1:100, each = 5),
    date = as.POSIXct(rep(seq(1.8*10e8, 1.8*10e8 + 388800, by = 86400), 100),
                      origin = "1960-01-01"),
    target = rnorm(5e2)
)
cv_split_temporal(dt, "target", "user", "date")