Statistics of History as 'retrospective prediction'? On the limits of predicting the future by learning from the past