Constructing a System of Early Warning Indicators for Financial Crises Using Machine Learning Techniques

The graduation thesis is dedicated to the study and analysis of machine learning methods for predicting economic crises. Due to increasing globalization and interdependence of the global economy, timely and accurate prediction of economic crises becomes a key task for reducing risks and maintaining economic stability. The research applies various statistical and computer technologies aimed at improving the ability to predict adverse economic events. The paper analyzes and compares different machine learning algorithms, including logistic regression, k-nearest neighbors, random forest, as well as advanced boosting methods such as XGBOOST, LIGHTGBM, and CATBOOST. The main focus is on evaluating and comparing the effectiveness of algorithms in predicting economic crises based on several indicators, including prediction accuracy, recall, F1-score, and the area under the receiver operating characteristic curve (ROC AUC). The results show that CATBOOST and LIGHTGBM demonstrate high performance in predicting the economic crisis state with higher accuracy compared to traditional statistical methods. The thesis also includes an analysis of the significance of economic variables and factors using SHAP analysis, which helps identify the most significant indicators contributing to the occurrence of crises. This research direction provides a deep understanding of risk factors and forms the basis for developing effective economic policy strategies in the regulation and monitoring of financial stability. The research findings provide valuable insights into the application of artificial intelligence and machine learning for analyzing economic data, contributing to the development of the scientific community.