Search code examples
rtime-seriesforecastfable-rtsibble

BoxCox.lambda not returning same results as fable features

I am comparing the results of the automatic lambda selection function BoxCox.lambda from the forecast package VS the fable package automatic lambda selection features

As you can see below, the two functions did not return the same results. Moreover, when I apply the BoxCox.lamda on a the same data, but once on a ts object and once on vector, the results are different.

Some body can explain me why it behaves like this ?

library(tidyverse)
library(tsibble)
library(lubridate)
library(fable)
library(tsibbledata)

library(forecast)


vic_cafe <- tsibbledata::aus_retail %>%
  filter(
    State == "Victoria",
    Industry == "Cafes, restaurants and catering services"
  ) %>%
  select(Month, Turnover)


lambda_fable <- vic_cafe %>% features(Turnover, guerrero) %>% pull(lambda_guerrero)
lambda_fable
#> [1] 0.1240828

lambda_forecast <- BoxCox.lambda(vic_cafe$Turnover, method = "guerrero")
lambda_forecast
#> [1] 0.02686482

lambda_forecast_ts <- BoxCox.lambda(as.ts(vic_cafe), method = "guerrero")
lambda_forecast_ts
#> [1] 0.1734189
Solution

  • The feasts::guerrero() function from uses all of the data available, whereas forecast::BoxCox.lambda() ignores data that doesn't fit exactly within the year.

    Your vic_cafe dataset starts in April, and so the feasts package will group seasons as April-March, but forecast will use January-December and remove the first part of the data.

    Some more discussion can be found here, and I've added a mention of these differences in the documentation: https://github.com/tidyverts/feasts/commit/830fe4095cf6231e7bb179519cddfeadd9cd7531