On the ordering of probability forecasts