Q: Why do I get a different shape of output when producing bootstrap-simulated forecasts from bottom-up vs top-down hierarchical reconciliation procedures in the R tidyvets/fable framework?

Borrowing from Forecasting Principles and Practice, Ch. 11.2

library (fpp3)

# making sure to set seed, so that the bottom-up reconciled forecasts `bu` use 
# the same bootstrapped sample as the baseline forecasts `ets`.
B -> 2
tourism_states %>%
  model(ets = ETS(Trips)) %>%
  reconcile(bu = bottom_up(ets),
            tpd = top_down(ets)) %>%
  forecast(bootstrap=TRUE,times=B,seed=1) ->
  fcs

fcs %>% # to display
  filter(State=='ACT' | is_aggregated(State)) %>%
  group_by(State,.model) %>%
  slice(1:2)

# # A fable: 12 x 5 [1Q]
# # Key:     State, .model [6]
# # Groups:  State, .model [6]
# State        .model Quarter     Trips  .mean
# * <chr*>       <chr>    <qtr>    <dist>  <dbl>
# 1 ACT          bu     2018 Q1 sample[2]   600.
# 2 ACT          bu     2018 Q2 sample[2]   730.
# 3 ACT          ets    2018 Q1 sample[2]   600.
# 4 ACT          ets    2018 Q2 sample[2]   730.
# 5 ACT          tpd    2018 Q1  598.2936   598.
# 6 ACT          tpd    2018 Q2  753.1552   753.
# 7 <aggregated> bu     2018 Q1 sample[2] 28832.
# 8 <aggregated> bu     2018 Q2 sample[2] 27310.
# 9 <aggregated> ets    2018 Q1 sample[2] 28753.
# 10 <aggregated> ets    2018 Q2 sample[2] 28181.
# 11 <aggregated> tpd    2018 Q1  28753.29 28753.
# 12 <aggregated> tpd    2018 Q2  28181.15 28181.

As expected, bottom-up bu and baseline ets methods produce a length=2 bootstrapped forecast in Trips. But, to my surprise, top-down tpd only produces a single estimate, the mean. This is even the case at the aggregated level, where the tpd bootstrapped forecast would simply be exactly the same as the baseline forecasts ets.

What’s going on here? Why doesn’t top-down work the way I’d expect? I’d like to produce bootstrapped forecast simulation paths for the top-down methodology. How can I do that here? If it’s not done easily, is there a hack I can use? E.g., would I have to loop over the total number of bootstrap simulations B and, instead, forecast with forecast(bootstrap=TRUE,times=1,seed=1) within each iteration?

Note, this issue I’m having with top-down also goes for middle-out.