We Trained a Neural Network for Gold, Then Deleted It: What the Out-of-Sample Years Showed