The Optimal Use of Return Predictability: An Empirical Study