This thesis attempts to model and forecast realized volatility and stock market tail risk using hybrid models integrating Machine Learning algorithms with Financial Time Series models. One of the advantages of Machine Learning approaches is that it can well approximate a wide range class of linear and nonlinear functions, forming the input-output map by learning the data rather than assuming the data generating process. Traditional Time Series models, however, focus on reproducing the stylized facts of target variables through statistical modeling. By hybriding these two types of models, we find that Machine Learning approaches well complement Financial Time Series models in variable screening, complex relationship detection and nonlinearity modeling. In addition, it is found that instead of using raw data in the Machine Learning algorithms, Financial Time Series models generate more effective features that significantly improves learning ability of those algorithms.