Extract elements from forecast, non-bootstrapped and bootstrapped in fable

[Amarjit_Chandhial]

Hi,

I can extract elements from non-bootstrapped and bootstrapped forecasts in fable using matrix and unlist, but would like a tidyverse version.

For non-bootstrapped forecasts, I want to extract the mean and variance (or sd) of the forecast distribution at EACH time-step, as individuals numbers e.g. mu1, sd1, mu2, sd2, etc.
For bootstrapped forecasts, I want to extract the 5000 bootstrapped samples at EACH time-step and stack them ALL in a tibble.

Here's a reproducible example

library(fpp3)

caf_fit <- global_economy |>
  filter(Code == "CAF") |>
  model(arima210 = ARIMA(Exports ~ pdq(2,1,0)))

caf_fit |> report()


# Non-Bootstrapped forecasts
# --------------------------
fc <- caf_fit |>
  forecast(h = 5, bootstrap = FALSE)

fc |> pull(Exports)

fc[[4]][[1]]
fc[[4]][[2]]
fc[[4]][[3]]
fc[[4]][[4]]
fc[[4]][[5]]

mu1 <- t(matrix(unlist(fc[[4]][1]), ncol = 1))[1]
sd1 <- t(matrix(unlist(fc[[4]][1]), ncol = 1))[2]
mu2 <- t(matrix(unlist(fc[[4]][2]), ncol = 1))[1]
sd2 <- t(matrix(unlist(fc[[4]][2]), ncol = 1))[2] 
# etc


# Bootstrapped forecasts
# ----------------------
fc_B <- caf_fit |>
  forecast(h = 5, bootstrap = TRUE)
fc_B

str(fc_B)

# make tb_B bootstrap forecasts
fc.B.1 <- rep(0, 5000) 
fc.B.2 <- rep(0, 5000) 
fc.B.3 <- rep(0, 5000) 
fc.B.4 <- rep(0, 5000) 
fc.B.5 <- rep(0, 5000) 

for (i in 1:5000) {
  fc.B.1[i] <- t(matrix(unlist(fc_B[[4]][1]), ncol = 1))[i]
  fc.B.2[i] <- t(matrix(unlist(fc_B[[4]][2]), ncol = 1))[i]
  fc.B.3[i] <- t(matrix(unlist(fc_B[[4]][3]), ncol = 1))[i]
  fc.B.4[i] <- t(matrix(unlist(fc_B[[4]][4]), ncol = 1))[i]
  fc.B.5[i] <- t(matrix(unlist(fc_B[[4]][5]), ncol = 1))[i]
}

fc_B_1 <- as_tibble(fc.B.1)
fc_B_2 <- as_tibble(fc.B.2) 
fc_B_3 <- as_tibble(fc.B.3) 
fc_B_4 <- as_tibble(fc.B.4)
fc_B_5 <- as_tibble(fc.B.5) 

tb_B <- bind_rows(
  list(
    f01_B = fc_B_1,
    f02_B = fc_B_2,
    f03_B = fc_B_3,
    f04_B = fc_B_4,
    f05_B = fc_B_5
  ),
  .id = "id"
)
head(tb_B)

Is there a more efficient way?

thanks,
Amarjit

[Amarjit_Chandhial]

I have managed to do this using pull and purrr::pluck

# Non-Bootstrapped forecasts
# --------------------------
fc |> pull(Exports) |> parameters(dist)

fc |> pull(Exports) |> mean()
mu1 <- fc |> pull(Exports) |> mean() |> first()
mu2 <- fc |> pull(Exports) |> mean() |> nth(2)
mu3 <- fc |> pull(Exports) |> mean() |> nth(3)
mu4 <- fc |> pull(Exports) |> mean() |> nth(4)
mu5 <- fc |> pull(Exports) |> mean() |> nth(5)

fc |> pull(Exports) |> variance()
var1 <- fc |> pull(Exports) |> variance() |> first()
var2 <- fc |> pull(Exports) |> variance() |> nth(2)
var3 <- fc |> pull(Exports) |> variance() |> nth(3)
var4 <- fc |> pull(Exports) |> variance() |> nth(4)
var5 <- fc |> pull(Exports) |> variance() |> nth(5)


# Bootstrapped forecasts
# ----------------------
fc_B
fc_B |> pull(Exports)
fc_B |> pull(Exports) |> first()
str(fc_B |> pull(Exports) |> first())
fc_B |> pull(Exports) |> first() |> purrr::pluck("x")
fc_B |> pull(Exports) |> nth(1) |> purrr::pluck("x")
fc_B |> pull(Exports) |> nth(2) |> purrr::pluck("x")
fc_B |> pull(Exports) |> nth(3) |> purrr::pluck("x")
fc_B |> pull(Exports) |> nth(4) |> purrr::pluck("x")
fc_B |> pull(Exports) |> nth(5) |> purrr::pluck("x")