Prediction power of high-frequency based volatility measures: a model based approach