From predictions to decisions: Using lookahead regularization