This message was deleted.

Hi @Manuel. We dont have conformal predictions in the library yet. The only other alternative is to use the multi-quantile losses, such as

MQLoss

and target specific quantiles directly.

@Cristian (Nixtla) Thanks! Unfortunately, even MQLoss does not give me good results. Regarding conformal prediction, do you think it's feasible to train a single model and then rely on the predict_insample() output, or would that be considered cheating? I guess using predictions made on training data can be a problem.