Modelo de clasificación de riesgo crediticio utilizando Random Forest en financiera del Ecuador

Abstract

This research focuses on Insotec, a medium-sized entity that dedicates to granting microcredits. The organization is looking forward to enhancing different areas procedures, including the credit evaluation, which is carried out manually. Currently, analysts classify risk on a case-by-case basis, and applications are approved depending on the credit rating. This process generates many drawbacks, such as high operational load and the probability of making wrong evaluations. This study proposes to enhance the risk evaluation process by creating a model based on an artificial intelligence algorithm that classifies the organization's customers as good or bad payers based on the different variables considered essential for the analysis. Furthermore, the study uses the organization's credit customer database with a historic from 2017 to 2020. This dissertation proposes creating an algorithmic model using the CRISP-DM methodology to classify the entity's clients. First, the business and the data set are understood. Second, the information is prepared within the data preprocessing phase. This step applies the outlier elimination technique. In this phase, the dataset reduces from 63,896 to 58,247 records. Afterward, the chi-square technique selects the variables with the most significant importance method, which sets eight variables for the analysis. When the dataset is ready, the CRISP-DM methodology suggests selecting the techniques most suitable for the model. In this case, the Random Forest algorithm obtains the best precision results compared against other algorithms such as neural networks and decision trees. The Random Forest model throws a precision of 97%, an error of 2.8%, and a rate of 2.1% false positives and 11.1% false negatives to predict. Additionally, the analysis implements ten models with the cross-validation method for the model evaluation phase. Finally, creating the model leads to having an additional tool, which automates the clients' classification process as good and bad payers. This automation can be used to deliver loans faster and with less risk. However, the deployment phase needs to be executed as mentions de CRISP-DM methodology states, so that the client can make the most of the knowledge obtained by the model.