Predictions over historical data in a landscape with fragile priors doesn't seem like a strong metric to me (it's a useful approximation at best)