Decision intelligence in the credit qualification assessment process

Abstract

For a company considering automating its decisions with machine learning, there are two fundamental considerations. First, the associated risks and rewards of this technology need to be evaluated. Second, as businesses often refrain from adopting machine learning in their processes due to lack of understanding, the issue of the interpretability of ML models must be addressed. The aim of the thesis is to automate credit qualification assessment for increased decision-making speed and quality. To do that, we summarize the literature regarding the use of ML in credit qualification. We then build a prediction model by combining state-of-the-art Bagging and Boosting algorithms including XGBoost and LightGBM and demonstrate their performance on a large dataset from a German bank. The methodology consists of data preparation, model construction and state-of-the-art hyperparameter optimization. These models are then combined into a heterogenous ensemble through a Bagging algorithm. For performance evaluation, we use several state-of-the-art metrics. We find LightGBM to be the best performing individual classifier and the heterogenous ensemble to outperform the individual models. While the out-of-sample performance of the models show a noticeable decrease, we conclude that partial automation of the credit qualification assessment based on the model is feasible and profitable. To address the explainability of the Classifier’s decisions, state-of-the-art model interpretation software SHAP is utilized. We show that the overall model as well as individual decisions can be interpreted, also shedding light on implicit decision factors that are present in the current manual process.

Date of Award	17 Oct 2023
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Dan Tran (Supervisor)