Using RMSE doesn't make sense to me, your losses are linear to how wrong you are. Shouldn't the error just be how far off you were?