Often we want our models to output a prediction *interval* rather than a point estimate. But did you know that quantile regression (classic approach) implicitly assumes the noise distribution to be symmetric?
In our new ICML paper we investigate and propose a solution.
🧵[1/N]