I am forecasting a collection of different types of items, using many different forecasting techniques. Some of the techniques I use take the input data as is to produce a distributional forecast. These are easy to combine and produce improved distributions from any model on their own.

For other forecasts, I use different types of data transformations prior to fitting models (box-cox, logs, IHS, etc.) I have noticed that you cannot combine forecast results which produce non-normal distribution estimates. Is there an explanation for this? Is it possible to combine forecasts with the same exact data transformations? (ex: an ARIMA and ETS model that take the natural log of the input data.)

Most of my forecasting and combining is done using the fable package in R.

  • 2
    $\begingroup$ What do you mean by "cannot" combine? Why couldn't you? $\endgroup$
    – Chris Haug
    Mar 4, 2021 at 23:54

1 Answer 1


You can certainly combine non-normal distributional (density) forecasts.

As a very simple example, your density forecasts may be empirical, derived by resampling in-sample residuals for various models. Then you can simply throw all these resamples into one big pile and get a large empirical predictive distribution.

In the parametric case, you can calculate mixtures of different types of distributions.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.