Create time series cross validation slices by key in the tidyverts package

Question

Is there a way to create time series cross validation sets by key using the tidyverts package? I can't seem to get it right. Below is a reprex of my attempt.

The example involves creating time series cross-validation (slices with 1 step ahead) for forecasting. The key variable has 2 distinct values and I will like to have one tsibble containing the time series slices for both keys. When I try to row-bind both tsibbles, I get an error.

library(dplyr)
library(tibble)
library(tsibble)

# helper function
create_cv_slices <- function(data, forecast_horizon) {
  data %>%
    dplyr::slice(1:(nrow(data) - forecast_horizon)) %>%
    tsibble::stretch_tsibble(.init = nrow(data) - 2 * forecast_horizon, .step = 1)
}

# get data
raw_tsbl <- tibble::tribble(
  ~index,      ~key,    ~Revenue,     ~Claims,
  20160101, "series1",  11011836.1, 5386836.696,
  20160201, "series1", 11042641.16, 9967325.715,
  20160301, "series1", 11445687.52, 10947197.89,
  20160401, "series1", 11252943.11, 6980431.415,
  20160101, "series2",    12236155,    12526224,
  20160201, "series2",     8675364,     9812904,
  20160301, "series2",    10081130,     8423497,
  20160401, "series2",    14840111,     8079813
) %>%
  dplyr::mutate(index = tsibble::yearmonth(as.character(index))) %>%
  tsibble::as_tsibble(index = index, key = key)

keys <- unique(raw_tsbl$key)

# split & combine
tbl1 = raw_tsbl %>%
  dplyr::filter(key == keys[1]) %>%
  create_cv_slices(., forecast_horizon = 1) %>%
  tibble::as_tibble()

tbl2 = raw_tsbl %>%
  dplyr::filter(key == keys[2]) %>%
  create_cv_slices(., forecast_horizon = 1) %>%
  tibble::as_tibble()

dplyr::bind_rows(tbl1, tbl2) %>%
  tsibble::as_tsibble(index = index, key = key)
#> Error: A valid tsibble must have distinct rows identified by key and index.
#> Please use `duplicates()` to check the duplicated rows.

Thank you.

Answer 1

It appears that using bind_rows to combine the tsibbles is what doesn't work. Using bind_rows and setting validate = FALSE in the as_tsibble function, creates a tsibble alright but it displays the tsibble as a daily series instead of monthly (which is what it should be). However, using rbind with the same argument setting, creates the desired tsibble.

rbind(tbl1, tbl2) %>%
  tsibble::as_tsibble(index = index, key = c(key, .id), validate = F)

Thanks.

Answer 2

Rather than splitting the data manually by key, you can compute your slices on groups of the tsibble. group_by_key() is a convenience function (with better performance) that is equivalent to group_by(key) . The n() function is a group aware dplyr function which gives the number of observations for the current group.

library(dplyr)
library(tibble)
library(tsibble)

# get data
raw_tsbl <- tibble::tribble(
  ~index,      ~key,    ~Revenue,     ~Claims,
  20160101, "series1",  11011836.1, 5386836.696,
  20160201, "series1", 11042641.16, 9967325.715,
  20160301, "series1", 11445687.52, 10947197.89,
  20160401, "series1", 11252943.11, 6980431.415,
  20160101, "series2",    12236155,    12526224,
  20160201, "series2",     8675364,     9812904,
  20160301, "series2",    10081130,     8423497,
  20160401, "series2",    14840111,     8079813
) %>%
  dplyr::mutate(index = tsibble::yearmonth(as.character(index))) %>%
  tsibble::as_tsibble(index = index, key = key)

forecast_horizon <- 1

raw_tsbl %>% 
  group_by_key() %>% 
  slice(1:(n() - forecast_horizon)) %>% 
  ungroup() %>% 
  stretch_tsibble(.init = 2, .step = 1)
#> # A tsibble: 10 x 5 [1M]
#> # Key:       .id, key [4]
#>       index key       Revenue    Claims   .id
#>       <mth> <chr>       <dbl>     <dbl> <int>
#>  1 2016 Jan series1 11011836.  5386837.     1
#>  2 2016 Feb series1 11042641.  9967326.     1
#>  3 2016 Jan series2 12236155  12526224      1
#>  4 2016 Feb series2  8675364   9812904      1
#>  5 2016 Jan series1 11011836.  5386837.     2
#>  6 2016 Feb series1 11042641.  9967326.     2
#>  7 2016 Mar series1 11445688. 10947198.     2
#>  8 2016 Jan series2 12236155  12526224      2
#>  9 2016 Feb series2  8675364   9812904      2
#> 10 2016 Mar series2 10081130   8423497      2

^{Created on 2020-05-08 by the reprex package (v0.3.0)}

A slight difference in this code is that .init is set to 2, rather than nrow(data)-2*forecast_horizon . For this data it gives the same result, however the number of observations for each key differs it won't. Once dplyr v1.0.0 is released, it will be easier to use tools like group_map() or bind_rows() to use a split-apply-combine approach necessary to specify different window parameters for each key.