# Evaluation metric for prediction interval?

In quantile regression (https://en.wikipedia.org/wiki/Quantile_regression), what are some suitable evaluation metrics?

Intuitively, I think a good model should have:

• good accuracy, i.e. the prediction interval (say at 90% confidence) should contain the true value of y.
• narrow width, i.e. the interval should not be large (if we expand the interval to [-Inf, Inf] definitely this interval will cover the true of y).

Could you recommend some ways to evaluate a predictor that outputs the prediction interval?

Many thanks