@Valeriy: its settled, NeuralProphet is also sort of broken.

As everyone knows I am a huge sceptic of Prophet, I have done some initial limited explorations of Neural Prophet beginning of this year but then didn't have time as started new role. https://valeman.medium.com/benchmarking-facebook-prophet-53273c3ee9c6 https://valeman.medium.com/benchmarking-neural-prophet-part-i-neural-prophet-vs-prophet-252990763468 https://valeman.medium.com/benchmarking-neural-prophet-part-ii-exploring-electricity-dataset-4cade7b2431c TL;DR Results were somewhat better than Prophet as including autoregressive terms helped, but not conclusive. Great comprehensive work evaluating it across multiple datasets.

@Max (Nixtla) it is settled bigly, I didn't pin big hopes on it as it is still curve fitter but AR terms were improvement even then my own tests indicated “deepAR” didn't improve much compared to linear AR. Even then to see it losing to ES is big news.

My Twitter thread on it, https://threadreaderapp.com/thread/1559948624548741123.html

@Max (Nixtla) @Kin Gtz. Olivares quick question - what sort of ES model is under the hood? Is there some document describing it, presumably it is some sort of similar to auto ARIMA with automatic hyperparameter optimisation?

I was editing the post and I erased the other thing I wrote about the nelder mead. Either way, we are preparing a post explaining that particular aspect and making that particular optimization method available to the community.