Beyond Random Split for Assessing Statistical Model Performance