forecast ARIMA by ID after after finding parameters in train data

I have around 14k timeseries, where each time serie its assigned to an ID.

I'm trying to forecast from February 2020 to January 2021 using ARIMA by using the hyperparameters PDQ that are found using 2018 as a train data. This is a sample of my data:

      ~ID, ~Period,  ~Value,
      1L, 201801L,   7713L,
      1L, 201802L,   4506L,
      1L, 201803L,  24475L,
      1L, 201804L,  12418L,
      1L, 201805L,  14545L,
      1L, 201806L,  14233L,
      1L, 201807L,   1271L,
      1L, 201808L,  19064L,
      1L, 201809L,   3018L,
      1L, 201810L,  13291L,
      1L, 201811L,  47111L,
      1L, 201812L,  16961L,
      1L, 201901L,  32442L,
      1L, 201902L,  16861L,
      1L, 201903L,  31819L,
      1L, 201904L,  38759L,
      1L, 201905L,  29220L,
      1L, 201906L,  19786L,
      1L, 201907L,  28620L,
      1L, 201908L,  47736L,
      1L, 201909L,  32586L,
      1L, 201910L,  12347L,
      1L, 201911L,  19758L,
      1L, 201912L,  14669L,
      1L, 202001L,   1499L,
      2L, 201801L,   1660L,
      2L, 201802L,   1857L,
      2L, 201803L,   3221L,
      2L, 201804L,  11009L,
      2L, 201805L,  11945L,
      2L, 201806L,   7152L,
      2L, 201807L,   3201L,
      2L, 201808L,  13226L,
      2L, 201809L,  13568L,
      2L, 201810L,  11952L,
      2L, 201811L,   1276L,
      2L, 201812L,  20049L,
      2L, 201901L,   7576L,
      2L, 201902L,  10370L,
      2L, 201903L,  47760L,
      2L, 201904L,  37809L,
      2L, 201905L,   9232L,
      2L, 201906L,  18635L,
      2L, 201907L,   6548L,
      2L, 201908L,  29065L,
      2L, 201909L,   2225L,
      2L, 201910L,   3613L,
      2L, 201911L,  11113L,
      2L, 201912L,   4626L,
      2L, 202001L,  12083L,
      3L, 201801L,  16850L,
      3L, 201802L,   9559L,
      3L, 201803L,   6727L,
      3L, 201804L,  29877L,
      3L, 201805L,   7453L,
      3L, 201806L,  11100L,
      3L, 201807L,  14289L,
      3L, 201808L,  16686L,
      3L, 201809L,  17925L,
      3L, 201810L,   2381L,
      3L, 201811L,  25015L,
      3L, 201812L,  20258L,
      3L, 201901L,  12875L,
      3L, 201902L,   8534L,
      3L, 201903L,   3880L,
      3L, 201904L,  27034L,
      3L, 201905L,  13624L,
      3L, 201906L,  29521L,
      3L, 201907L,   4933L,
      3L, 201908L,   5963L,
      3L, 201909L,  15193L,
      3L, 201910L,   2960L,
      3L, 201911L,   6150L,
      3L, 201912L,  18957L,
      3L, 202001L, 10326L,
        4L, 201801L,   85837L,
        4L, 201802L,   90903L,
        4L, 201803L,  110829L,
        4L, 201804L,   67992L,
        4L, 201805L,  117665L,
        4L, 201806L,  136909L,
        4L, 201807L,  -23708L,
        4L, 201808L,  196362L,
        4L, 201809L,  -28869L,
        4L, 201810L,  114243L,
        4L, 201811L,  113408L,
        4L, 201812L,   18932L,
        4L, 201901L,  254189L,
        4L, 201902L, -151225L,
        4L, 201903L,  103182L,
        4L, 201904L, -242319L,
        4L, 201905L,  111250L,
        4L, 201906L,  449959L,
        4L, 201907L,  105185L,
        4L, 201908L,  103575L,
        4L, 201909L,  214451L,
        4L, 201910L,   99015L,
        4L, 201911L,  280420L,
        4L, 201912L,  -15325L,
        4L, 202001L,  199340L
    df$year<-as.numeric(substr(df$Period,start = 1,stop = 4))

Some treatment to use ARIMA from library fable:

df <- df %>% 

df<-df %>% 
  mutate(YearMonth = tsibble::yearmonth((ymd(date)))) %>%
  as_tsibble(key=ID,index = YearMonth)

df_train<- df %>% 
  filter(YearMonth <= yearmonth("2018 Dec") & YearMonth>=yearmonth("2018 Jan")) %>%
  model(selected_model=ARIMA(Value ~ PDQ(0,0,0), stepwise=FALSE, approximation=FALSE))

Im saving the parameters with a map:

PDQ<-df_train$selected_model %>%

# A tibble: 4 x 8
#      p     d     q     P     D     Q constant period
#  <int> <int> <int> <dbl> <dbl> <dbl> <lgl>     <dbl>
#1     0     0     0     0     0     0 TRUE         12
#2     0     0     0     0     0     0 TRUE         12
#3     0     0     0     0     0     0 TRUE         12
#4     2     0     0     0     0     0 TRUE         12

And some metrics:

df_train %>%
  forecast(h = 13) %>%
  accuracy(df) %>%
  select(ID, RMSE, MAPE)

For example, for ID nª1 it has an ARIMA (0,0,0) and ID nº4 has an ARIMA (2,0,0).

Now I need to use the parameters of each ID for each time series.

I dont know how to loop it through the ids and using the pdq values, or if there's a function that can be useful for this problem. Does anyone know how to solve it? Thanks!

The simplest approach here is to re-estimate your models on the new data using the refit() function. By setting reestimate = TRUE, the coefficients will be re-estimated to suit the new data provided, but the hyperparameters (number of terms) will not change.

df_train %>% 
  # Re-estimate the coefficients of df_train with data up to and including Jan 2020.
  refit(df %>% filter(YearMonth <= yearmonth("2020 Jan")), reestimate = TRUE) %>% 
  # Forecast one year ahead (Feb 2020 - Jan 2021)
  forecast(h = "1 year")
#> # A fable: 48 x 5 [1M]
#> # Key:     ID, .model [4]
#>       ID .model         YearMonth             Value  .mean
#>    <int> <chr>              <mth>            <dist>  <dbl>
#>  1     1 selected_model  2020 Feb N(20188, 1.7e+08) 20188.
#>  2     1 selected_model  2020 Mar N(20188, 1.7e+08) 20188.
#>  3     1 selected_model  2020 Apr N(20188, 1.7e+08) 20188.
#>  4     1 selected_model  2020 May N(20188, 1.7e+08) 20188.
#>  5     1 selected_model  2020 Jun N(20188, 1.7e+08) 20188.
#>  6     1 selected_model  2020 Jul N(20188, 1.7e+08) 20188.
#>  7     1 selected_model  2020 Aug N(20188, 1.7e+08) 20188.
#>  8     1 selected_model  2020 Sep N(20188, 1.7e+08) 20188.
#>  9     1 selected_model  2020 Oct N(20188, 1.7e+08) 20188.
#> 10     1 selected_model  2020 Nov N(20188, 1.7e+08) 20188.
#> # … with 38 more rows

Created on 2020-08-02 by the reprex package (v0.3.0)

1 Like