On modeling and forecasting stock volatility

Accurate measures and forecasts of volatility are indispensable for asset pricing, portfolio selection and risk management. Instrumental for these tasks have been statistical approaches such as Autoregressive Conditional Heteroskedastic (ARCH) and Stochastic Volatility models as well as economic formulations that imply future volatility estimates using the information contained in options. The resulting volatility estimates are however model driven and therefore sensitive to the particular specification chosen. Our first chapter introduces this dissertation and outlines in greater detail the limitations of the existing methods when measuring, modeling and/or forecasting volatility. In the second chapter we derive that under general conditions daily volatility can be measured model-free and to any degree of accuracy from intradaily return observations. Using the record of each transaction underlying the Dow Jones Industrials Average portfolio, we next document the empirical properties of such ‘realized’ volatility measure and capture its characteristics using a time-series model. On the basis of ex ante one-day-ahead prediction criteria we find that this specification yields unbiased and accurate volatility predictions and that these clearly improve upon the ones obtained by various ARCH models, including those that closely match the volatility regularities we document. The third chapter concerns the identification of models that provide good volatility forecast over short and long horizons ex ante and ex post. We also examine whether the very ‘model-free’ advantage of the employed realized volatility measure makes it inefficient as it ignores any structural dependence in the intradaily data. As this structure turns out to be quite complex, we extend traditional ARCH specifications via semi-parametric methods to model intradaily returns. We find that the realized volatility and semi-parametric specifications perform equally well and that, for various in-sample and out-of-sample horizons, both of them yield far better forecasts than the ones that are obtained using numerous daily ARCH models. The last chapter concludes the dissertation with recommendations for future research.