A carregar...
Publicação
Modelling Default Risk Dynamics in Syndicated Loans: An Empirical Analysis with Macro-Financial Data and Machine Learning
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.32 MB | Adobe PDF |
Orientador(es)
Resumo(s)
This thesis explores the use of machine learning models to predict quarter-over-quarter
increases in defaults within the U.S. syndicated loan market, focusing exclusively on loans held
by depository institutions. The objective is to assess whether default volumes—either in raw
dollar amounts or as proportions of total credit exposure—are likely to rise in a given quarter,
based on a combination of macroeconomic indicators and credit exposure data broken down
by borrower risk class. A binary classification framework is employed, where models output
probabilities of default increases, which are subsequently transformed into an interpretable
risk index. The data, consisting of approximately 60 quarterly observations from 2009 to 2024,
is processed through a rigorous pipeline that includes feature transformation,
multicollinearity filtering, and greedy forward feature selection based on a composite
performance metric. Six classification models are evaluated—Logistic Regression, Decision
Tree, Random Forest, Gradient Boosting, Naive Bayes, and Linear Discriminant Analysis—
across both raw and proportional target formulations. Model performance is assessed using
a range of metrics, including AUC, LogLoss, F2-score, and accuracy, with robustness checks via
time-aware cross-validation and bootstrapping. The results suggest that macroeconomic
variables, particularly lagged unemployment and interest rate indicators, are consistently
among the most informative predictors. Gradient Boosting emerges as the most accurate and
robust model in both formulations, although simpler models like Naive Bayes and LDA also
demonstrate competitive performance in certain settings. Notably, credit portfolio variables
did not feature prominently in the selected predictors, highlighting the dominant role of
broader economic signals in anticipating shifts in aggregate credit risk. These findings have
implications for early-warning system design and systemic risk monitoring, offering a practical
approach to forecasting financial stress using public and aggregate data.
Descrição
Dissertation presented as the partial requirement for obtaining a Master's degree in Information Management, specialization in Business Intelligence
Palavras-chave
Credit risk Syndicated loans Default prediction Machine learning Early-warning systems