Cannot forecast of a reconciled ensemble model which includes prophet()

[ewlundstrom]

I find that while I am able to forecast a reconciled ensemble forecast that does not include a prophet model, an error occurs when a prophet model is included.

Reproducible example:

library(tsibble)
library(tsibbledata)
library(remotes)
remotes::install_github("tidyverts/fable")
library(fable)
library(fabletools)
library(fable.prophet)

## Creating aggregate TS
Agg_Retail <- aus_retail %>% aggregate_key(State,
                                               Turnover = sum(Turnover))

## Ensemble Model without Prophet

Agg_Retail %>%
  model(Ensemble = combination_ensemble(ARIMA(Turnover),
                             ETS(Turnover),
                             weights = "inv_var")) %>%
  reconcile(min_trace(Ensemble)) %>%
  forecast(h = 12)


## Ensemble Model with Prophet


Agg_Retail %>%
  model(Ensemble = combination_ensemble(ARIMA(Turnover),
                                        ETS(Turnover),
                                        prophet(Turnover),
                                        weights = "inv_var")) %>%
  reconcile(min_trace(Ensemble)) %>%
  forecast(h = 12)

Error encountered:

! Problem with `mutate()` input `min_trace(Ensemble)`.
x min_trace needs covariance matrix to be positive definite.
i Input `min_trace(Ensemble)` is `(function (object, ...) ...`.
Run `rlang::last_error()` to see where the error occurred.
Warning message:
9 errors (1 unique) encountered for Ensemble
[9] .onLoad failed in loadNamespace() for 'processx', details:
  call: NULL
  error: 'ps_is_supported' is not an exported object from 'namespace:ps'```

[mitchelloharawild]

This should now be possible since ensemble models can now produce bootstrapped forecasts. This was added recently in dfc4011.

To produce ensemble forecasts from models with different distributions, produce bootstrap or simulated sample paths with forecast(bootstrap = TRUE) or forecast(simulate = TRUE).