Aplicación de técnicas de Inteligencia Artificial para predecir mora en créditos de una institución financiera en Ecuador

Abstract

INTRODUCTION. This research addresses the prediction of credit default through the implementation of models based on Artificial Intelligence techniques, specifically Machine Learning. The dependent variable is default, and the independent variables include demographic, socioeconomic, and credit history characteristics. OBJECTIVE. Implement and train predictive models using supervised machine learning techniques, with the aim of anticipating possible loan defaults and supporting decision-making. METHOD. The stages of the CRISP-DM methodology were applied, starting with data extraction, transformation, and loading, followed by exploratory analysis, cleaning, correlation verification, supervised algorithm training, and performance evaluation. RESULTS. The highest recall rate of 0.68, a key indicator for identifying defaults, was obtained with the Logistic Regression algorithm using the SMOTE balancing technique. DISCUSSION AND CONCLUSIONS. The result contrasts with other studies that adopt the Random Forest model in default prediction problems, in which case the recall values obtained were not significant. An important limitation was the imbalance in the variable to be predicted, which was addressed using balancing techniques. Finally, the importance of empirically validating the results according to the data and the specific context of application is evident.