How to deal with NA's in timeseries in R?

Question

I'm trying to build an ARIMA-model, to forecast the occupancy rates in an office. There are some NA's in the data, those dates are national holidays, which means no one is in the office and therefore there is no data. How can I deal with those NA values to build an ARIMA-model?

Example of NA:

    2019-04-19 09:00:00      12.878788
    2019-04-19 10:00:00      19.848485
    2019-04-19 11:00:00      21.969697
    2019-04-19 12:00:00      11.212121
    2019-04-19 13:00:00      14.090909
    2019-04-19 14:00:00      16.363636
    2019-04-19 15:00:00      22.727273
    2019-04-19 16:00:00       7.727273
    2019-04-22 09:00:00             NA
    2019-04-22 10:00:00             NA
    2019-04-22 11:00:00             NA
    2019-04-22 12:00:00             NA
    2019-04-22 13:00:00             NA
    2019-04-22 14:00:00             NA
    2019-04-22 15:00:00             NA
    2019-04-22 16:00:00             NA
    2019-04-23 09:00:00      23.636364
    2019-04-23 10:00:00      49.545455
    2019-04-23 11:00:00      57.575758
    2019-04-23 12:00:00      48.030303
    2019-04-23 13:00:00      45.151515
    2019-04-23 14:00:00      35.606061
    2019-04-23 15:00:00      25.151515
    2019-04-23 16:00:00       8.333333

I tried using this code:

    plot(stl(ts, na.action = na.omit))

But I got this error:

    Error in na.omit.ts(as.ts(x)) : time series contains internal NAs

Answer 1

ARIMA models in R handle NA's without any problem. STL decompositions do not handle NAs, which is where your error is coming from.

If you want to do an STL, you could use mstl from the forecast package which estimates the missing values for you.

library(forecast)
library(ggplot2)
USAccDeaths[20:23] <- NA
USAccDeaths %>%
  mstl(s.window="periodic") %>%
  autoplot()

USAccDeaths %>%
  auto.arima() %>%
  forecast(h=24) %>%
  autoplot()

^{Created on 2019-11-30 by the reprex package (v0.3.0)}