This suggests that the best way to grade predictions is some sort of weighting of how unlikely they were at the time. Like, if you were to open a prediction market for statement X, some sort of grade of the delta between your confidence of the event and the “expected” value, summed over all your predictions.