Efficient daily volatility forecasts and forecast revisions using intra-day data

The objective of this thesis is to examine the predictive performance of begin-of-the-day volatility in forecasting end-of-the-day volatility using intra-day data; to search for an efficient timing window to make our forecast and to assess the process of forecast revisions during the day. For this we develop three methods that make use of diurnal patterns: the average seasonal over the last K days, an exponentially weighted average seasonal and a Flexible Fourier Form alternatively we build a method on Mincer-Zarnowitz bias adjustments. As the actual volatility is an unobserved variable several high-frequency volatility estimators: Realized Variance, Bipower variation, Realized Range, Two Time Scales and Kernel estimation are proposed to base ex-ante forecasts and measure ex-post forecasting performance. We conclude that begin-of-the-day volatility is a highly predictive measure for end-of-the-day volatility and can be improved by scaling and bias adjustments. The assessment of efficient timing samples leads us to believe that roughly the first 15-20 min are most informative, with the remark that omitting the first 0-10 minutes of return data could be beneficial as noise prevails during this interval. Furthermore, correlations between lagged and current volatility forecast revisions is found to be small yet statistically significant, implying our models inherited some build in smoothing property.