This thesis investigates the applications of machine learning in Financial Portfolio and Risk Management. The focus is to customize machine learning algorithms to accommodate the intuitions or practical needs in the domain. Empirical experiments are carried out to examine the proposed customizations. An extensive breadth of machine learning topics are discussed, explored, and extended. The experiments in this thesis represent the customization of algorithms in three aspects of any portfolio and risk management system: 1. A generic prediction framework that automates predictions to provide insights for future expectations. 2. A risk-aware agent that controls the balance between actively shifting a portfolio and the transaction costs involved. 3. A robust dynamic portfolio selection algorithm that continually diversifies to track switching regimes. Experiment 1: Practical Bayesian support vector regression for Financial Time Series Prediction The first experiment outlines a generic prediction framework that takes advantage of the powerful support vector regression. This framework introduces a faster and easier parameter selection process to determine the model that generates predictions and the corresponding uncertainty estimates. It is shown that the generalization performance of this parameter selection process can reach or sometimes surpass the computationally expensive cross-validation procedure. In addition, an ad-hoc adaptive calibration process is described to enable practical use of the prediction uncertainty estimates to assess the quality of predictions, which is also interpreted as a potential indicator of market condition changes. Experiment 2: Risk-Aware Reinforcement Learning Algorithm to Improve a Portfolio The model-free Monte Carlo control reinforcement learning algorithm is extended, by making use of its episodic nature, to allow consideration of "risk" when training the algorithm. The risk-aware reinforcement learning algorithm introduced allows the user to intuitively and flexibly incorporate any form(s) of risk consideration desired. A procedure is then suggested to filter out potentially unstable policies. The risk-aware mechanism is examined, and its abilities to control "risk" are demonstrated in empirical experiments. In addition, it is recommended to diversity out-of-sample by simultaneously following multiple policies with high in-sample Sharpe ratio. Experiment 3: Expert Advice Algorithms for Dynamic Portfolio Selection The connections between online machine learning and the sequential investment problem are explored in this experiment, and the Smart Switching Portfolio (SSP) Algorithm is proposed. It continually diversifies wealth to assets based on their previous performances to track switching regimes. A newly introduced scaling parameter illustrates the linkage between the learning rate and the action of leveraging. Moreover, the algorithm is theoretically generalized to select assets from a dynamic pool of investible assets. The behavior of the SSP Algorithm is examined. The effect of the new parameter under different volatility levels is also assessed. The proposed algorithm is shown to be the most robust. It outperforms some well-known algorithms, and is particularly impressive as transaction cost increases. A few ad-hoc methods are proposed to potentially enhance the algorithm further.