Is it possible to avoid negative prediction intervals in fable with non-negative time series that include natural zeros? Box-Cox transformation with lambda=0 can be used when there is no zero but what if it includes many zeros.
You could use a log(x+c) transformation where c is some small number. It doesn't quite solve the problem, but often works well enough.
Otherwise you could try a different transformation. Some are discussed here: https://stats.stackexchange.com/q/1444/159
Thank you Rob, I found the solution using scaled_logit() in fpp3 very useful too.
This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.