Component dynamic models for realized covariance matrices

The increasing availability of high-quality transaction data across many financial assets, allow the construction of estimates of ex-post daily realized volatility and co-volatility by summing squares and cross-products of intraday high-frequency returns. This exhaustive set of data has stimulated the development of a lively field of econometric analysis with the proposition of models directly fitted to time series of realized measures. One interest of these models is that they can be used for forecasting future values, which is typically of use in financial applications such as hedging, option pricing, risk management, and portfolio allocation. (Multivariate) GARCH models can be used for the same purposes but since they rely on daily observed returns, in principle they provide less precise estimates and forecasts of variances and covariances than measures based on intraday data. The present thesis tries to contribute to the existing literature by proposing feasible models for realized covariance matrices that can be readily applied in financial applications. The common feature of the novel specifications is a dynamic temporal structure of the covariance matrix consisting of a short-term and a long-term component, whereby the long-term component can be shaped differently depending on the specific problem at hand. Throughout the thesis we firstly embrace the common belief of a constant long-run component and then we relax it, thus illustrating the main strengths and weaknesses of both approaches. Furthermore, in each case, we propose our own methodology to allow estimation for large cross-sections, thus mitigating the common issue of parameter proliferation that prevents the application of many models to realistic high-dimensional frameworks. The usefulness of the proposed procedures is supported by extensive empirical evidence.(ECGE - Sciences économiques et de gestion) -- UCL, 201