Holt forecast for multiple timeseries in R

Question

I am trying to do Holt's forecast for multiple timeseries and combine them with my original data.frame. Consider the following data.frame, where I have two population groups:

library("forecast")

d <- data.frame(SEX       = c("MALE","MALE","MALE","FEMALE","FEMALE","FEMALE"),
                EDUCATION = c("01","01","01","01","01","01"),
                TIME      = c("2000","2001","2002","2000","2001","2002"),
                VALUE     = c(120,150,140,90,75,60))

Then I am doing the Holt's forecast for the two time series:

male   <- ts(as.numeric(d[1:3,]$VALUE),start=c(2000))
female <- ts(as.numeric(d[4:6,]$VALUE),start=c(2000))

forecastmale   <- holt(male,h = 3,damped = FALSE)
forecastfemale <- holt(female,h = 3,damped = FALSE)

Then I save the result and combine with my original data.frame:

forecastmale   <- data.frame(forecastmale[["mean"]])
forecastfemale <- data.frame(forecastfemale[["mean"]])

forecastmale$SEX            <- c("MALE","MALE","MALE")
forecastmale$EDUCATION      <- c("01","01","01")
forecastmale$TIME           <- c("2003","2004","2005")
colnames(forecastmale)[1]   <- "VALUE" 
forecastmale                <- forecastmale[, c(2,3,4,1)]

forecastfemale$SEX          <- c("FEMALE","FEMALE","FEMALE")
forecastfemale$EDUCATION    <- c("01","01","01")
forecastfemale$TIME         <- c("2003","2004","2005")
colnames(forecastfemale)[1] <- "VALUE"
forecastfemale              <- forecastfemale[, c(2,3,4,1)]

d <- rbind(d,forecastmale,forecastfemale)

This works when I only have two time series. But if I have like 100 time series that has to be forecasted, then it is not a very efficient way do to it. Can anyone help with make the coder more efficient, so if I for instance include an extra population group in my data.frame, then I do not have change anything in the code?

Answer 1

This is what the fable package is designed to handle. Here is an example using the same data structure that you have.

library(dplyr)
library(tsibble)
library(fable)

# Artifical data
df <- expand.grid(
    education = 1:3,
    sex = c("male","female"),
    year = 1990:2002
  ) %>%
  as_tsibble(index=year, key=c(sex,education)) %>%
  mutate(value = rnorm(78))

# Fit Holt's method to each series and forecast 3 years ahead
df %>%
  model(holt = ETS(value ~ trend("A"))) %>%
  forecast(h=3)
#> # A fable: 18 x 6 [1Y]
#> # Key:     sex, education, .model [6]
#>    sex    education .model  year         value  .mean
#>    <fct>      <int> <chr>  <dbl>        <dist>  <dbl>
#>  1 male           1 holt    2003  N(0.14, 1.7)  0.137
#>  2 male           1 holt    2004  N(0.17, 1.7)  0.171
#>  3 male           1 holt    2005  N(0.21, 1.7)  0.205
#>  4 male           2 holt    2003 N(-0.75, 1.5) -0.749
#>  5 male           2 holt    2004 N(-0.84, 1.8) -0.837
#>  6 male           2 holt    2005   N(-0.93, 2) -0.926
#>  7 male           3 holt    2003  N(0.51, 0.7)  0.514
#>  8 male           3 holt    2004  N(0.53, 0.7)  0.530
#>  9 male           3 holt    2005  N(0.55, 0.7)  0.546
#> 10 female         1 holt    2003 N(0.44, 0.98)  0.445
#> 11 female         1 holt    2004 N(0.47, 0.98)  0.470
#> 12 female         1 holt    2005  N(0.5, 0.98)  0.495
#> 13 female         2 holt    2003 N(0.13, 0.89)  0.127
#> 14 female         2 holt    2004 N(0.15, 0.89)  0.148
#> 15 female         2 holt    2005 N(0.17, 0.89)  0.168
#> 16 female         3 holt    2003  N(0.78, 1.8)  0.781
#> 17 female         3 holt    2004  N(0.88, 1.8)  0.880
#> 18 female         3 holt    2005  N(0.98, 1.8)  0.978

^{Created on 2020-09-05 by the reprex package (v0.3.0)}