CODEWITHSUNDEEP
rtime-seriesarima
AWM
AWM

Reputation: 3

ARIMAX with grouped data R

Im new with R and I need to compare the accuracy of ARIMAX and ARIMA. This is a sample of my data and what I've done to do the ARIMA model: 

library(dplyr)
library(forecast)
library(lubridate)
data<-tibble::tribble(
        ~id, ~day, ~month, ~year, ~value, ~reg1, ~reg2,
         1L,   1L,     1L, 2019L,  4.634, 0.626, 0.684,
         1L,   1L,     2L, 2019L,  2.969, 0.698, 0.049,
         1L,   1L,     3L, 2019L,  1.885,  0.62, 0.155,
         1L,   1L,     4L, 2019L,  2.415, 0.553, 0.959,
         1L,   1L,     5L, 2019L,  2.215, 0.598, 0.065,
         1L,   1L,     6L, 2019L,  1.805, 0.454,  0.07,
         1L,   1L,     7L, 2019L,  4.682, 0.045, 0.376,
         1L,   1L,     8L, 2019L,  4.248, 0.087, 0.094,
         1L,   1L,     9L, 2019L,   0.55, 0.523,  0.86,
         1L,   1L,    10L, 2019L,  0.109, 0.176, 0.591,
         2L,   1L,     1L, 2019L,  2.918, 0.442, 0.956,
         2L,   1L,     2L, 2019L,  3.083, 0.233, 0.388,
         2L,   1L,     3L, 2019L,  3.271, 0.652, 0.946,
         2L,   1L,     4L, 2019L,  2.175, 0.704, 0.902,
         2L,   1L,     5L, 2019L,   4.51, 0.851, 0.533,
         2L,   1L,     6L, 2019L,  4.178, 0.655, 0.614,
         2L,   1L,     7L, 2019L,  1.956, 0.434, 0.977,
         2L,   1L,     8L, 2019L,  3.219, 0.418,   0.4,
         2L,   1L,     9L, 2019L,   2.72, 0.335, 0.096,
         2L,   1L,    10L, 2019L,  4.519, 0.534, 0.388,
         3L,   1L,     1L, 2019L,  2.969, 0.707, 0.752,
         3L,   1L,     2L, 2019L,  2.456, 0.085, 0.651,
         3L,   1L,     3L, 2019L,  0.418, 0.851, 0.399,
         3L,   1L,     4L, 2019L,  2.324, 0.626, 0.317,
         3L,   1L,     5L, 2019L,  3.548, 0.175, 0.081,
         3L,   1L,     6L, 2019L,   3.74, 0.667, 0.691,
         3L,   1L,     7L, 2019L,   4.48, 0.853, 0.259,
         3L,   1L,     8L, 2019L,   0.18, 0.016, 0.489,
         3L,   1L,     9L, 2019L,  3.028,  0.51, 0.741,
         3L,   1L,    10L, 2019L,  4.652, 0.916, 0.953
        )
data<-data %>% 
  mutate(date=as.character(make_date(year,month,day)),YearMonth = tsibble::yearmonth((ymd(date)))) %>%
  as_tsibble(key=id,index = YearMonth)


fit <- data %>% 
  filter(YearMonth <= yearmonth("2019 Aug")) %>%
  model(ARIMA(value ~ PDQ(0,0,0), stepwise=FALSE, approximation=FALSE))

# Now forecast the test set and compute RMSE and MSE
fit %>%
  forecast(h = 2) %>%
  accuracy(data)

Now I need to do this but with an ARIMAX: 

covariates <- c("reg1","reg2")
fit_arimax <- data %>% 
  filter(YearMonth <= yearmonth("2019 Aug")) %>%
  group_by(id) %>% 
  do(autoarima=auto.arima(.$value,xreg = as.matrix(data[,covariates])))

and I get the following error: 

Error in model.frame.default(formula = x ~ xregg, drop.unused.levels = TRUE) : 
  variable lengths differ (found for 'xregg')

In addition: Warning message: In !is.na(x) & !is.na(rowSums(xreg)) : longer object length is not a multiple of shorter object length

I saw this answer but I couldn't do it, as I'm a beginner in R. So I want to know if ARIMA has something to work with the regressors or how to solve it with auto.arima, and then compare the accuracy by ID in ARIMA and ARIMAX. Does anyone know how to? thanks !

Upvotes: 0

Views: 10166

Answers (1)

Rob Hyndman
Rob Hyndman

Reputation: 31820

You've switched from using the tsibble and fable packages to using the forecast packages. These use different data structures and should not generally be mixed.

You can easily fit a regression model with ARIMA errors using fable as follows.

fit_arimax <- data %>% 
  filter(YearMonth <= yearmonth("2019 Aug")) %>%
  model(
    ARIMA(value ~ reg1 + reg2 + PDQ(0,0,0))
  ) 

fc <- fit_arimax %>%
  forecast(new_data = filter(data, YearMonth > yearmonth("2019 Aug"))) 
fc %>% accuracy(data)

Note that this is not actually an ARIMAX model -- see https://robjhyndman.com/hyndsight/arimax/

Upvotes: 3

Related Questions