The large increase in the number of traded assets in the portfolio of most financial institutions has made the measurement of risk exposure a primary concern for regulators and for internal risk control. The key ingredient for all relevant risk measures is a volatility forecast.

The focus of this diploma thesis is in finding an optimal forecasting procedure for the volatility on the basis of high frequency data. In the first few sections, the properties of financial time series and the issue of volatility measurement are discussed. Next, suitable prediction models are identified by an explorative analysis of high frequency realized volatilities. Thereafter those models are thoroughly explained. The next sections contain an empirical study, where the performance of those forecasting procedures is explored. As the true volatility is latent, measuring the accuracy of the predictions is difficult. In order to avoid these problems, a new evaluation method is proposed. Last but not least, the question whether a more reliable volatility forecast can improve the quality of the risk measures is answered.