Hi team! We received a question from a user (Roberto McLean, economist, non-paying) asking about the specific variant of conformal prediction implemented in our open-source libraries (StatsForecast, MLForecast, NeuralForecast). He’s wondering whether we’re using a variant such as *S*plit Conformal, Inductive Conformal, ACI, EnbPI or CQR. Is there any documentation or paper I could point him to?

I think @Mariana Menchero or @Marco could help with that.

Well, I'm really not super well versed in conformal prediction, and I find the papers to be really hard to understand. So, to the best my knowledge, our method is more similar to a bootstrap method (so I don't think we use any of the methods mentioned above). You can refer to this paper: https://arxiv.org/pdf/2109.12156 However, I think the implementation is more simplistic and not as robust as the actual methods described in the papers. That way, we get a good enough estimation and keep things fast. We use the same method in all packages. @Mariana Menchero, if I'm wrong, or if you have something to add, please let us know 😅

thanks @Marco you are correct in that this is the "vanilla" (?) version of conformal prediction, which is described in Rob Hyndman's book here. Maybe we can share this reference with the user @Yibei?

Got you, thanks!! I’ll update him with the reference in Hyndman book:)