Book excerpt: We address three key aspects of optimal portfolio construction: expected return, variance-covariance modeling and optimization in presence of cardinality constraints. On expected return modeling, we extend the self-excited point process framework to model conditional arrival intensities of bid and ask side market orders of listed stocks. The cross-excitation of market orders is modeled explicitly such that the ask side market order size and bid side probability weighted order book cumulative volume can affect the ask side order intensity, and vice versa. Different variations of the framework are estimated by using method of maximum likelihood estimation, based on a recursive application of the log-likelihood functions derived in this thesis. Results indicate that the self-excited point process framework is able to capture a significant amount of the underlying trading dynamics of market orders, both in-sample and out-of-sample. A new framework is introduced, Realized GARCH, for the joint modeling of returns and realized measures of volatility. A key feature is a measurement equation that relates the realized measure to the conditional variance of returns. The measurement equation facilitates a simple modeling of the dependence between returns and future volatility. Realized GARCH models with a linear or log-linear specification have many attractive features. They are parsimonious, simple to estimate, and imply an ARMA structure for the conditional variance and the realized measure. An empirical application with DJIA stocks and an exchange traded index fund shows that a simple Realized GARCH structure leads to substantial improvements in the empirical fit over standard GARCH models. Finally we describe a novel algorithm to obtain the solution of the optimal portfolio problem with NP-hard cardinality constraints. The algorithm is based on a local relaxation that exploits the inherent structure of the objective function. It solves a sequence of small, local, quadratic-programs by first projecting asset returns onto a reduced metric space, followed by clustering in this space to identify sub-groups of assets that best accentuate a suitable measure of similarity amongst different assets. The algorithm can either be cold started using the centroids of initial clusters or be warm started based on the output of a previous result. Empirical result, using baskets of up to 3,000 stocks and with different cardinality constraints, indicates that the algorithm is able to achieve significant performance gain over a sophisticated branch-and-cut method. One key application of this local relaxation algorithm is in dealing with large scale cardinality constrained portfolio optimization under tight time constraint, such as for the purpose of index tracking or index arbitrage at high frequency.