Reinterpreting causal discovery as the task of predicting unobserved joint statistics