Describe the workflow you want to enable

I'd like to evaluate and compare the predictive performance of (conditional) quantiles as predicted by GradientBoostingRegressor(loss='quantile', alpha=0.9) for example.

Describe your proposed solution

Implement a new metric quantile_loss(y_true, y_pred, alpha=0.5), Eq. (24) of https://arxiv.org/pdf/0912.0902.pdf.

This is the same loss as in Koenker's book "Quantile Regression" and in: