PREDICTING BANKING CRISES USING MACHINE LEARNING: THE CASE OF LEBANON

Abstract

This thesis revisits the existing empirical evidence on the predictability of systemic banking crises using machine learning methods with a particular emphasis on Lebanon's crisis of 2019.

More specifically, the dataset of Laeven and Valencia (2020) is extended by appending to it Lebanon's systemic banking crisis, and the predictive ability of machine learning techniques such as Logit, KNNs, SVMs, Trees, and XGBoost is assessed.

Evaluating the methods using the F-1 score and the ROC AUC suggests that the best-performing models over the testing period are the KNNs and XGBoost.