Exploiting high frequency data for volatility forecasting and portfolio selection

An instant may matter for the course of an entire life. It is with this idea that the present research had its inception. High frequency financial data are becoming increasingly available and this has triggered research in financial econometrics where information at high frequency can be exploited for different purposes. The most prominent example of this is the estimation and forecast of financial volatility. The research, chapter by chapter is summarized below. Chapter 1 provides empirical evidence on univariate realized volatility forecasting in relation to asymmetries present in the dynamics of both return and volatility processes. It examines leverage and volatility feedback effects among continuous and jump components of the S&P500 price and volatility dynamics, using recently developed methodologies to detect jumps and to disentangle their size from the continuous return and the continuous volatility. The research finds that jumps in return can improve forecasts of volatility, while jumps in volatility improve volatility forecasts to a lesser extent. Moreover, disentangling jump and continuous variations into signed semivariances further improves the out-of-sample performance of volatility forecasting models, with negative jump semivariance being highly more informative than positive jump semivariance. A simple autoregressive model is proposed and this is able to capture many empirical stylized facts while still remaining parsimonious in terms of number of parameters to be estimated. Chapter 2 investigates the out-of-sample performance and the economic value of multivariate forecasting models for volatility of exchange rate returns. It finds that, when the realized covariance matrix approximates the true latent covariance, a model that uses high frequency information for the correlation is more appropriate compared to alternative models that uses only low-frequency data. However multivariate FX returns standardized by the predicted high frequency multivariate volatility are not normally distributed. Using high frequency information for the correlation, the multivariate model tends to produce forecasts of tail risk which are lower than the realized tail risk, under the normality assumption. Chapter 3 finds that a risk factor, constructed from high frequency market price data and representing the volatility asymmetry, is able to explain a significant proportion of the time and cross sectional variation of average returns. Small, growth and high beta portfolios are particularly subject to the asymmetry risk. A multi-factor pricing model with the high frequency volatility asymmetry performs well in pricing tests. Realized volatility ; forecasting ; high frequency data ; volatility asymmetry ; mixed frequency model ; conditional correlation ; risk evaluation