How Do You Measure the Accuracy of a Probabilistic Forecast?

The probability distributions tend to be extremely skewed and the hard part actually is estimating the sort of the upper tail. So how do you measure the accuracy of a probabilistic forecast? We use CRPS the continuous rank probability scores for all of the component models to see which ones are doing best, we run that retrospectively they're measuring both how it's calibrated and how sharp it is so sharpness has had narrow the distribution.