Rolling Origin & Initial Time Split: Lag parameter

[mdancho84]

As described in #135, the rolling_origin() and initial_time_split() functions have a proposed modification to add an overlap parameter, which is necessary to perform resamples and train/test validation with lagged predictors when sample sizes are very low.

# - Add overlap parameter to initial_time_split() & rolling_origin()

# Libraries
library(recipes)
library(rsample)
library(tidyverse)

# 3 years of data
drinks_subset <- drinks %>% tail(36)

splits_no_overlap <- initial_time_split(drinks_subset, prop = 2/3, overlap = 0)

training(splits_no_overlap)
#> # A tibble: 24 x 2
#>    date       S4248SM144NCEN
#>    <date>              <dbl>
#>  1 2014-10-01          11817
#>  2 2014-11-01          10470
#>  3 2014-12-01          13310
#>  4 2015-01-01           8400
#>  5 2015-02-01           9062
#>  6 2015-03-01          10722
#>  7 2015-04-01          11107
#>  8 2015-05-01          11508
#>  9 2015-06-01          12904
#> 10 2015-07-01          11869
#> # … with 14 more rows

testing(splits_no_overlap)
#> # A tibble: 12 x 2
#>    date       S4248SM144NCEN
#>    <date>              <dbl>
#>  1 2016-10-01          11914
#>  2 2016-11-01          13025
#>  3 2016-12-01          14431
#>  4 2017-01-01           9049
#>  5 2017-02-01          10458
#>  6 2017-03-01          12489
#>  7 2017-04-01          11499
#>  8 2017-05-01          13553
#>  9 2017-06-01          14740
#> 10 2017-07-01          11424
#> 11 2017-08-01          13412
#> 12 2017-09-01          11917

# Without overlap - Get missing values in testing dataset 
rec_lag <- recipe(~ ., data = training(splits_no_overlap)) %>%
  step_lag(S4248SM144NCEN, lag = 12) 

bake(prep(rec_lag), testing(splits_no_overlap)) %>% tail(12) # Missing values
#> # A tibble: 12 x 3
#>    date       S4248SM144NCEN lag_12_S4248SM144NCEN
#>    <date>              <dbl>                 <dbl>
#>  1 2016-10-01          11914                    NA
#>  2 2016-11-01          13025                    NA
#>  3 2016-12-01          14431                    NA
#>  4 2017-01-01           9049                    NA
#>  5 2017-02-01          10458                    NA
#>  6 2017-03-01          12489                    NA
#>  7 2017-04-01          11499                    NA
#>  8 2017-05-01          13553                    NA
#>  9 2017-06-01          14740                    NA
#> 10 2017-07-01          11424                    NA
#> 11 2017-08-01          13412                    NA
#> 12 2017-09-01          11917                    NA

# With overlap - No missing values
splits_with_overlap <- initial_time_split(
  drinks_subset, 
  prop    = 2/3, 
  overlap = 12 # New parameter
)

rec_lag <- recipe(~ ., data = training(splits_with_overlap)) %>%
  step_lag(S4248SM144NCEN, lag = 12) 

bake(prep(rec_lag), testing(splits_with_overlap)) %>% tail(12) # Get the right values
#> # A tibble: 12 x 3
#>    date       S4248SM144NCEN lag_12_S4248SM144NCEN
#>    <date>              <dbl>                 <dbl>
#>  1 2016-10-01          11914                    11983
#>  2 2016-11-01          13025                    11506
#>  3 2016-12-01          14431                    14183
#>  4 2017-01-01           9049                    8650
#>  5 2017-02-01          10458                    10323
#>  6 2017-03-01          12489                    12110
#>  7 2017-04-01          11499                    11424
#>  8 2017-05-01          13553                    12243
#>  9 2017-06-01          14740                    13686
#> 10 2017-07-01          11424                    10956
#> 11 2017-08-01          13412                    12706
#> 12 2017-09-01          11917                    12279

^{Created on 2020-03-26 by the reprex package (v0.3.0)}

[topepo]

Looks good but can you make these changes:

• Change the argument name to lags instead of overlap. I think that's more literal (but make suggestions if you have another idea).

• Add some input validation for overlap/lags.

[mdancho84]

@topepo Sounds good to me. I will:

1. Change argument to lags.
2. Add input validation - Checks for whole number and not greater than the training data size.

[mdancho84]

@topepo I've implemented the change to lag (singular, not plural since only one lag should be added not multiple) and I've added 2 checks to make sure a whole number that does not exceed the number of training observations. Should be good to go.

[topepo]

Thanks!

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex https://reprex.tidyverse.org) and link to this issue.