SaintRod
commented
1 year ago Greetings. First time Nixtla user - trying to port work from R fable. Adding support for blambda in ARIMA.
I'm currently log transforming Y before passing it to the model and back transforming Y to it's original scale via exp($\hat{y}$). What gives me pause is in regression context when back-transforming a logged variable I was taught to do exp($\hat{y} + \frac{\sigma^2}{2}$). I'm unsure if a similar back-transformation is required in time series. Based on the below link on Rob Hyndman's blog I think it's not as simple as exp($\hat{y}$). I think the fable models handle this in the background but I'm not sure.
For context Rob Hyndman link
Thanks for the hard work!
Description
Currently, the blamda is raising an expectation, when trying to use it, would be nice to have this enabled, I am not sure if the commented code just needs to be enabled again or if there is a reason for this?
https://github.com/Nixtla/statsforecast/blob/cb80fffc33655f5c7f6cbf44bb3baa4347e431d1/statsforecast/arima.py#L1425
Use case
No response