Custom evaluation metric

[dirkschumacher]

It would be nice to have a metric that takes into account the prediction error and also the proportion of observations within the prediction interval. Since we are mostly interested in a good prediction interval than a point estimate.