statsforecast coherence with R solution

[ncooder]

Description

Are there any tests that show that statsforecast gives exactly the same results as the R packages? I see a difference in parameter selection and values for the AutoCES.

Link

No response

[MMenchero]

Hi @ncooder, there are some slight variations in the R results, mainly due to the optimization method used. I have observed similar discrepancies when working on TBATS. In our case, we use the minimize function from scipy.optimize, while the forecast R package uses optim from stats. The results are identical to those in R until we optimize the likelihood; then, some variations appear. So the differences you've seen in parameter selection and values for the AutoCES are to be expected.

If you think the differences are too significant, please provide us with a reproducible example, and we will take a closer look at it.

[ncooder]

@MMenchero I will try to demonstrate the difference between the statsforecast and the smooth package for the AutoCES model. You are right that the problem could be related to optimization. In any case, the complex numbers I get when using the statsforecast and the smooth in R are significantly different.