This dissertation consists of two essays in financial econometrics. The first essay delves into the investigation of estimators for factor-model-based large covariance (and precision) matrices utilizing high-frequency data that are asynchronous and potentially affected by market microstructure noise. Our estimation approaches involve the pre-averaging method with refresh time to address microstructure challenges, along with three different factor model specifications incorporating various thresholding methods to combat the curse of dimensionality. In the second essay, we harness high-frequency data from over 10,000 stocks spanning more than two decades to employ machine learning algorithms in forecasting realized volatility. By capitalizing on the nonlinear relationships between realized volatility and a diverse set of features and panel information, machine learning algorithms—particularly neural networks—demonstrate significantly improved performance compared to traditional ordinary least square methods.