Recent advances in the measurement of beta (systematic return risk) and volatility (total return risk), demonstrate substantial advantages in utilizing high frequency return data in a variety of settings. These advances in the measurement of beta and volatility have resulted in improvements in the evaluation of alternate beta and volatility forecasting approaches. In addition, more precise measurement has also led to direct modeling of the time variation of beta and volatility. Both the realized beta and volatility literature have most commonly modeled with an autoregressive process. In this paper we evaluate constant beta models, against autoregressive models of time-varying realized beta. We find that a constant beta model computed from daily returns over the last 12 months generates the most accurate quarterly forecast of beta and dominates the autoregressive time series forecasts. It also dominates (dramatically) the popular Fama-MacBeth constant beta model which uses 5 years of monthly returns.