(updated) ARIMA gives subscript out of bounds when forecasting with fable/tsibble

[xzeng76]

Hi, I am working on a hierarchical arima model to predict store transactions. The data is of tsibble format and I set key on store_nbr and family. The index is date. This morning I tried to fit the model and it gave NULL ARIMA MODELs. Then I contraint the dataset and tried to study on specific store_nbr and family. It showed an error of subscript out of bounds. could you give me some clue how to debug such problems? thanks!

The dataset looks like this:

trans_tbl1 <- trans_tbl %>%
            filter(store_nbr %in% c(1) & family=='AUTOMOTIVE')
head(trans_tbl1)

the output is:
A grouped_ts: 6 × 8

Blockquote
|date|store_nbr|family|transactions|onpromotion|log.sales|sales|wday|
| --- | --- | --- | --- | --- | --- | --- | --- |
|||||||||
|2013-01-02|1|AUTOMOTIVE|2111|0.01|0.6931472|2.00|4|
|2013-01-03|1|AUTOMOTIVE|1833|0.01|1.0986123|3.00|5|
|2013-01-06|1|AUTOMOTIVE|520|0.01|0.6931472|2.00|1|
|2013-01-07|1|AUTOMOTIVE|1807|0.01|-4.6051702|0.01|2|
|2013-01-10|1|AUTOMOTIVE|1679|0.01|0.6931472|2.00|5|
|2013-01-15|1|AUTOMOTIVE|1680|0.01|0.0000000|1.00|3|

The model fitting is as:

model(trans_tbl1,
    arima = ARIMA(transactions ~ onpromotion + wday +
                  pdq(2,1,0) + PDQ(1,1,1),
                approximation=FALSE, order_constraint = p+q+P+Q <= 10,na.action=na.omit)
  )

and it gave an error:

Error in [[.default(.data[[index_var(.data)]], 1): subscript out of bounds
Traceback:

1. model(trans_tbl1, arima = ARIMA(transactions ~ onpromotion +
   . wday + pdq(2, 1, 0) + PDQ(1, 1, 1), approximation = FALSE,
   . order_constraint = p + q + P + Q <= 10, na.action = na.omit))
2. model.tbl_ts(trans_tbl1, arima = ARIMA(transactions ~ onpromotion +
   . wday + pdq(2, 1, 0) + PDQ(1, 1, 1), approximation = FALSE,
   . order_constraint = p + q + P + Q <= 10, na.action = na.omit))
3. eval_models(models, .data[["lst_data"]])
4. map(models, function(model) {
   . map(lst_data, estimate_progress, model)
   . })
5. lapply(.x, .f, ...)
6. FUN(X[[i]], ...)
7. map(lst_data, estimate_progress, model)
8. lapply(.x, .f, ...)
9. FUN(X[[i]], ...)
10. estimate(dt, mdl)
11. fabletools::estimate(dt, null_model(!!f))
12. estimate.tbl_ts(dt, null_model(!!f))
13. .model$add_data(.data)
14. .data[[index_var(.data)]][[1]]
15. [[.Date(.data[[index_var(.data)]], 1)
16. .Date(NextMethod("[["), oldClass(x))
17. NextMethod("[[")

Thanks!

Hi, I am working on a store transaction prediction problem. I converted the dataset into tsibble format. And try to predict store transactions depending on store number(store_nbr) and product categories (family). 

`head(trans_df)`
gives:

```
A grouped_ts: 6 × 8|date|store_nbr|family|transactions|onpromotion|log.sales|sales|wday|
| --- | --- | --- | --- | --- | --- | --- | --- |
|<date>|<fct>|<fct>|<int>|<dbl>|<dbl>|<dbl>|<dbl>|
|2013-01-02|1|AUTOMOTIVE|2111|0.01|0.6931472|2.00|4|
|2013-01-03|1|AUTOMOTIVE|1833|0.01|1.0986123|3.00|5|
|2013-01-06|1|AUTOMOTIVE|520|0.01|0.6931472|2.00|1|
|2013-01-07|1|AUTOMOTIVE|1807|0.01|-4.6051702|0.01|2|
|2013-01-10|1|AUTOMOTIVE|1679|0.01|0.6931472|2.00|5|
|2013-01-15|1|AUTOMOTIVE|1680|0.01|0.0000000|1.00|3|
```
and I fit the model with the following codes:

```
fit <- trans_tbl %>%
  model(
    arima = ARIMA(transactions ~ onpromotion + wday +
                  pdq(2,1,0) + PDQ(1,1,1),
                approximation=FALSE, order_constraint = p+q+P+Q <= 10)
  )
fit
```
it showed all fitted models are null. Do you know what are the possibilities to cause an null model? Thanks!

```
A mdl_df: 6 × 3|store_nbr|family|arima|
| --- | --- | --- |
|<fct>|<fct>|<model>|
|1|AUTOMOTIVE|<NULL model>|
|1|BEAUTY|<NULL model>|
|1|BEVERAGES|<NULL model>|
|1|BREAD/BAKERY|<NULL model>|
|1|CELEBRATION|<NULL model>|
|1|CLEANING|<NULL model>|
```