Essays on the use of MIDAS regressions in banking and finance

Economic time series are available at different frequencies due to their origin and data collection techniques. A mixed data sampling (MIDAS) regression is mainly a forecasting tool designed to harness mixed-frequency data. This dissertation proposes a computationally efficient estimation approach for MIDAS models and demonstrates its advantages in applications involving Big Data, including those related to banking and finance. In the first essay, the logit model is generalized by introducing the MIDAS logit model. It is applied to predict U.S. bank failures. The proposed model yields a strongly significant improvement in classification accuracy both in statistical and economic terms. Some of the largest bank failures previously misclassified are now correctly predicted. Furthermore, an algorithm is suggested to mitigate the class-imbalance problem that is pervasive across many datasets. The second essay is also related to the investigation of the banks' financial stability. The objective of this paper is to forecast bank closures costs in the United States. It is found that the primary federal regulators supervise banks inconsistently which affects the formation of costs. The accuracy of forecasting is low for banks supervised by some regulators. This evidence questions the quality of supervision and implies that the off-site monitoring might be blinded. In the third essay, MIDAS regressions are applied to predict the realized covariance matrices. A covariance matrix forecast is important in risk-management and portfolio applications. A study is performed across nearly all stocks in Dow Jones and it is discovered that the MIDAS model has superior forecasting accuracy as compared to the HAR model. The economic value of the proposed method is evaluated using portfolio management applications in terms of the investor's utility and annualized returns. The fourth essay aims at developing a computationally efficient approach to estimating MIDAS regressions and demonstrates its advantages in applications involving Big Data, in which estimation has previously not been feasible. I revise an optimization procedure for a MIDAS-NLS estimator and reach a major enhancement in its performance. In order to facilitate the use of my approach, that is the MIDAS-NLS with revised optimization, the "MIDAS Analytic" toolbox was released. Documentation for the toolbox is provided in Chapter 5.