Forecasting using transformations

[newsagent]

Hi all,

In section 3.1 of the Principles for forecasting (https://otexts.com/fpp3/transformations.html#eq:boxcox), the Box-Cox transformation is described in relation to forecasting. Since we can pull out the optimal value of lambda, using the guerrero feature for a given series, why not work always with this type of transformation. For example, if I want to analyze any series, say, y_t, I can "filter" it with the Box-Cox transformation and then work with the transformed series.

So I can always(?) run this hypothetical code

lambda <- df %>%
  features(y, features=guerrero) %>%
  pull(lambda_guerrero)

df_bc <- df %>%  mutate(box_cox = box_cox(y, lambda))

and then I can use the ARIMA, ETS, etc, based on the y_bc.

So my question is: can I always follow this process to be safe? Or is this approach applied to series with particular characteristics that
are observed by plotting the series? In other words, under which circumstances, can I use this approach?

Best regards

[robjhyndman]

The Guerrero approach is one way to automate the chosen value of lambda, but it is not the only way, and it is not always reliable. That is why we do not use it automatically in ARIMA. It is important to check the plots and ensure that the resulting forecasts look sensible. Sometimes the Guerrero method will choose a value that is lower than what looks sensible to the eye, and then the forecasts can explode.

[newsagent]

Thank you!