Análisis de sesgo de modelos de aprendizaje automático en la predicción de riesgo de crédito en la banca en el Ecuador

Abstract

Assessing bias in machine learning models designed to predict credit risk in banking is critical to ensuring fair and transparent decisions. The purpose of this analysis is to investigate the potential for bias in the algorithms used to evaluate the reliability and sensitivity of the data needed in the credit rating process. One of the fundamental aspects covered in this study is the need to correct identified errors and take specific actions to reduce the presence of erroneous data in forecast models. This includes the introduction of capital metrics when evaluating model performance to improve transparency and accountability in the use of algorithms in credit decisions. The detailed analysis aims to identify potential sources of bias, such as unbalanced data sets or discriminating variables, that could affect the accuracy and reliability of the forecasts. Having unbalanced data can lead to biased results because the model may favor certain groups over others due to a lack of equal representation in the training data set. It also emphasizes the importance of eliminating discriminatory variables that may introduce biases into the credit rating process. These variables can include sensitive information such as age, gender, or race, which if not used appropriately can negatively affect the model's decisions. To achieve the goal of correct and accurate outcome prediction, it is proposed to implement bias reduction techniques, such as sample weighting or incorporating fairness characteristics in the model training process. Models must also be continually monitored and updated to ensure reliability and accuracy as data sets and environmental conditions change. In summary, the article emphasizes the importance of analyzing and eliminating errors in machine learning-based prediction models for credit risk.