cv_decomposition.Rdcv_decomposition() applies rolling cross-validation to the portion of an
STL decomposition conditional on future data. It provides
the predicted values and associated "errors" for each "forecast" horizon and
decomposed component across all stably-estimable time points.
cv_decomposition(
.data,
.col = "observed_cleaned",
trend = "31 days",
period = "7 days"
)A data frame containing the time series data to resample
The column containing the data to resample
The length of time to use in trend decomposition; can be a
time-based definition (e.g. "1 month") or an integer number of days. If
NULL or "auto", trend is set automatically using the tunable
heuristics in the timetk package.
The length of time to use in seasonal decomposition; can be a
time-based definition (e.g. "1 week") or an integer number of days. If
NULL or "auto", period is set automatically using the tunable
heuristics in the timetk package.
A list of tibble objects, each containing the results of one
sampling step for the dates in start_date + trend to
end_date - trend/2, where end_date is the last completely observed
date. See the Value section of
validate_decomposition()for information on the components of each sample.
When smoothing time series using LOESS, a portion of the smooth will change
as future data comes in (approximately the last trend/2 data points). These
points are said to be conditional on future data, and the difference between
the smooth with and without future data can be computed for all time points
that have already reached the "stable" portion of the series. This is done by
rolling the decomposition step across the time series and comparing the
smoothed point at each "forecast" horizon to the stable estimates. In effect,
this is rolling cross-validation of the STL decomposition using the fully
smoothed data as the reference values.
The workhorse function
validate_decomposition()
and the higher-level function
cv_linelist_decomposition()