Hey, how did you guys make neural forecast so incredibly fast? Rn I'm working on my msc thesis where I make the models by myself using torch and it doesn't even come close. I'm scared to use Nixtla as not everything I want is supported, so I wonder if it's possible to learn to know how you guys made it so fast :). Thanks in advance.

Hi @Christiaan can you tell us what you need that is not currently supported? That will help us plan future developments better. And maybe what you need is already in the works.

E.g. Probabilistic methods such as using gaussian mixtures, MC dropout (even though not a great method for probabilistic forecasts it can still serve as a baseline). But how did you make it so fast?

@Christiaan why would you want these methods in Nixtla? Nixtla stack includes things that work and such methods don’t have coverage and instead of user specified coverage of 90%-95% they show coverage as low as 20%. MC dropout is not a baseline method for probabilistic forecasting it is a model. If you would like a baseline I suggest NPTS https://docs.aws.amazon.com/forecast/latest/dg/aws-forecast-recipe-npts.html

Fair reasoning and thank you a lot for that method!