Predictive Modelling of Financial Crises: Machine Learning Algorithms in a Recursive Real-time Framework

The thesis explores the literature on quantitative methods employed in predicting financial crises, spanning from classical approaches like signals and regressions to the more recent and advanced machine learning algorithms, such as artificial neural networks. The document provides a comprehensive explanation of the algorithms, the incorporated variables, and the evaluation metrics. Following this, it outlines the outcomes achieved by earlier studies. The author conducts experiments with various algorithms, assessing their effectiveness through a recursive real-time test. The analysis involves comparing results across different methods and subsets of data. Special emphasis is placed on the objective of forecasting using exclusively retrospective information, aiming to simulate a real-world implementation.