Yu, Qilu2022-09-152022-09-152022-05Yu_cornellgrad_0058F_13069http://dissertations.umi.com/cornellgrad:13069https://hdl.handle.net/1813/111824159 pagesThere are three chapters in this dissertation. Chapter 1 introduces the machine learning and its advantages and disadvantages in the context of economic research. The machine learning algorithms can complement traditional econometric methods and expand the boundaries of research.Chapter 2 uses a novel deep learning approach, “Temporal Causal Discovery Framework (TCDF)” to uncover the causal graph structure on the European countries’ credit default swaps during the 2010-2013 eurozone crisis. TCDF uses attention-based convolutional neural networks combined with a causal validation step to learn the causal relationships and the time delay between a cause and the occurrence of its effect. This study provides a granular report of the eurozone crisis contagion and spillovers, adding new findings to the repository. The benchmark Granger causality tests are implemented by vector autoregression. The comparison between the two methods suggests the TCDF can filter the “real” cause-effect relationships using causality validation. Chapter 3 extends the famous Jordà-Schularick-Taylor Macrohistory database with a new crisis variable by referencing other crisis datasets. This new dataset contains 1570 observations of 17 countries from 1870 to 2016, of which 322 observations are crisis periods. XGBoost, random forest, and the logit model are applied to this dataset to establish early warning systems for financial crisis. Though XGBoost is a popular tool in applied ML, it has rarely been used in previous studies for early warning systems. This chapter shows that XGBoost outperforms the benchmark logit model, its performance is on a par with random forest. The two machine learning methods can achieve excellent prediction performance evaluated by the AUROC. Shapley values of the variables are calculated from the models to rank the variable importance in terms of predictive power.endissertation or thesishttps://doi.org/10.7298/hjcs-1c28