Skip to content

feat: Add quantile (pinball) loss - #672

Open
georgeguimaraes wants to merge 2 commits into
elixir-nx:mainfrom
georgeguimaraes:quantile-loss
Open

georgeguimaraes wants to merge 2 commits into
elixir-nx:mainfrom
georgeguimaraes:quantile-loss

Conversation

@georgeguimaraes

@georgeguimaraes georgeguimaraes commented Sep 15, 2026

Copy link
Copy Markdown
Member

Adds Axon.Losses.quantile/3, the pinball loss. Same shape as huber: takes :reduction plus a :quantile option (default 0.5, where it's half the MAE).

It's the loss you train quantile regression heads with, so a model predicts the 10th and 90th percentiles instead of the mean and you get prediction intervals out of it. I've been using it in soothsayer for NeuralProphet-style quantile outputs and it felt like it belonged next to the other regression losses here.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant