crossv_ts: Generate test/train sets for time-series

In jrnold/ramsleep: Resampling, Cross-Validation, and Permutation Methods Infrastucture

Description

Generate test/train sets for time series. Each training set consists of observations before the test set. This method also called "evaluation on a rolling forecasting origin".

Usage

crossv_ts(n, horizon = 1L, test_size = 1L, test_partial = FALSE,
  train_partial = TRUE, train_size = n, test_start = NULL, from = 1L,
  to = n, by = 1L)

Arguments

n	A positive, scalar integer representing the number of observations (items to choose from).
horizon	Difference between the first test set observation and the last training set observation
test_size	A scalar integer indicating the size of the test set.
test_partial, train_partial	A logical scalar or positive scalar integer which handles how partial test (train) sets are handled. If TRUE, then partial test (train) sets are allowed. If FALSE, then partial test (train) sets are not allowed. If it is an integer, then it is the minimum allowable size of a test (train) set.
train_size	The maximum size of the training set. This allows for a rolling window training set, instead of using all observations from the start of the time series.
test_start	An integer vector of the starting index values of the test set.
from, to, by	Integer scalar values used to generate values of test_start if test_start = NULL. In that case, test_start = seq(from, to, by).

Details

In time-series cross-validation the training set only uses observations that are prior to the test set. Suppose the time series has n observations, the training set has a maximum size of r <= n and minimum size of s >= r. and the test set has a maximum size of p <= n and minimum size of q >= p.

For indices i in {1, …, N}:

• Select observations i, ..., max(p, n) for the test set.

• Select observations max(i - h - p), ..., i - h for the training set.

• If the test set has a size of at least q and the training set has a size of at least r.

References

• Hyndman RJ (2017). forecast: Forecasting functions for time series and linear models. R package version 8.0, URL.

• Hyndman RJ and Khandakar Y (2008). "Automatic time series forecasting: the forecast package for R." Journal of Statistical Software. URL.

• Rob J. Hyndman. "Cross-validation for time series". December 5, 2016.

• Rob J. Hyndman. "Time series cross-validation: an R example". August 26, 2011.

• Rob J. Hyndman and George Athanasopoulos. "Evaluating Forecast Accuracy." URL.

• Max Kuhn. "Data splitting for Time Series." The caret Package. 2016-11-29. URL.

jrnold/ramsleep documentation built on May 29, 2019, 11:43 a.m.