2

does any one have the dataset requirements for fable package in R here some problems I had, any one can give any suggestion will be nice. my R version is

platform x86_64-w64-mingw32
arch x86_64
os mingw32
system x86_64, mingw32
status
major 3
minor 5.2
year 2018
month 12
day 20
svn rev 75870
language R
version.string R version 3.5.2 (2018-12-20) nickname Eggshell Igloo

  1. could I use irregular time data for modeling? say: I have price data for 250 days out of 365 days but I still want to use fable to model ARIMA of the price. is that possible? the example form github is uesing tsibbledata::ausretail has no missing date in the data set

  2. seems the commend the fable pacakge grammar changed I was using the example from this page 2018-12 it was fine https://github.com/mitchelloharawild/fable-tfeam-2018/blob/master/index.Rmd

but now I am not able to use the code. e.g. the ETS was using

fbl_cafe_fit <- vic_cafe %>%
  fable::ETS(Turnover ~ season("M"))

now from this page, people need to put extra 'model' outside?? https://github.com/tidyverts/fable

    UKLungDeaths %>%
  model(ets = ETS(log(mdeaths))) %>%
  forecast

is that new grammar or my understanding is wrong?

  1. Seems now i do not have auto.arima () option from fable any more??? i need to specify pdq() and PDQ()

    USAccDeaths %>% as_tsibble %>% model(arima = ARIMA(log(value) ~ pdq(0,1,1) + PDQ(0,1,1)))

  2. after i fit the arima model, i also have problem use the fitting model to predict next period this grammar not work any more:

    fbl_cafe_fc <- fbl_cafe_fit %>% forecast(h=24)

Steffen Moritz
  • 7,277
  • 11
  • 36
  • 55
CloverCeline
  • 511
  • 3
  • 18

1 Answers1

2

  1. ARIMA requires a regular time series, however it will also work in the presence of missing values. You can use tsibble::fill_gaps() to convert implicit missing values to explicit.

  2. Correct, the fable package is currently experimental and changes to the interface are expected to continue. These changes will likely have a relatively minor impact on users. Since the fable TFEAM talk, we now support multiple model columns in a mable. To achieve this, we now use model() to specify models. Previously, if you wanted to model data %>% ETS(log(y) ~ season("A")), this is now data %>% model(ETS(log(y) ~ season("A")).

  3. Automatic model selection (such as forecast::auto.arima()) is contained within the same function in fable (ARIMA()). When estimating a model, if the right-hand-side is left empty, a model will be chosen automatically from the defaults. For ARIMA models, if you used data %>% model(ARIMA(y)), an appropriate model will be automatically chosen (same as forecast::auto.arima()). You can also now estimate an ARIMA(p,0,0)(2,1,Q)[12] model, where p and Q are unknown between 0 and 3. To do this, you would use data %>% model(ARIMA(y ~ pdq(0:3, 0, 0) + PDQ(2, 1, 0:3, period = 12))).

  4. That code looks correct, and should still work. Perhaps you need to update the packages.

Mitchell O'Hara-Wild
  • 2,174
  • 5
  • 9
  • One more question here,@Mitchell O'Hara-Wild, if i want to use weekly data for prediction(I have tsibble showing it is [7d] ts already), if i want to use auto ARIMA model. should i use: `data %>% model(ARIMA(y))`, or should I use `data %>% model(ARIMA(y),period=52)`? Thank you! – CloverCeline Mar 07 '19 at 12:23
  • 2
    If you have weekly data, I'd first suggest using the yearweek() class for your index. An annual seasonal pattern for your weekly data will be selected by default in ARIMA, however to make this more explicit you could use `data %>% model(ARIMA(y ~ PDQ(period = 52)))`. Note that you could also use `period = "year"`. – Mitchell O'Hara-Wild Mar 07 '19 at 12:43