Aplicação do aprendizado de máquinas para classificação de cooperativas de crédito em risco de encerramento

Abstract

Introduction Credit cooperatives, entities formed by individuals with common economic interests, aim to achieve both the economic and social goals of their members (McKillop & Wilson, 2011). Given this, it is essential to monitor the risks associated with these entities, especially those related to credit granting, due to potential conflicts of interest. Determining the most relevant indicators for categorizing cooperatives with a potential for discontinuity can lead to more effective monitoring and support decision-making in the management process. Research Problem and Objective Due to the high risks involved in potential conflicts of interest, it is essential to monitor and assess capital adequacy, asset quality, management quality, profitability, and liquidity, which can be carried out using the CAMEL methodology. However, the large number of indicators can make it difficult to analyze the actual financial situation of these organizations. Thus, this study aimed to identify the most relevant accounting indicators for classifying Brazilian credit cooperatives at risk of future closure. Theoretical Framework Several studies (Silva, Santos, Ranciaro, 2023; Vieira, Bressan V., Bressan A., 2018) highlight the importance of performance analysis in credit cooperatives. Although their objective is not profit generation, cooperatives need to manage their resources efficiently to provide members with higher returns on their investments and better rates for financial operations. Understanding these results is essential to support their social and economic functions and ensure their survival in the financial system. Methodology To achieve this, the study employed the machine learning methodologies Random Forest and K-Nearest Neighbors (KNN). The Random Forest model enabled the identification of the most relevant indicators, while the KNN model was used to classify cooperatives based on their potential risk of closure. Using a sample of 8,552 observations from single credit cooperatives between 2018 and 2022, the dataset was divided into a training set (80% of the sample) and a test set (remaining 20%). Results Analysis The results from non-parametric mean difference tests showed that the analyzed CAMEL indicators exhibited statistically significant differences in the years preceding closure. However, the model achieved 92.9% accuracy in predicting the future operational status of the cooperatives, with a 93% success rate for active cooperatives but only 46% for those classified as closed. Thus, while the model was ineffective in identifying closed cooperatives, it was efficient in classifying the most impactful indicators. Conclusion Due to the low accuracy in identifying closed cooperatives, the study concludes that using the six most impactful indicators provides a similar level of information to that obtained with the full set of CAMEL indicators. Additionally, given the high error rate in classifying credit cooperatives at risk of closure, the findings suggest the need to explore alternative models or techniques for predicting credit cooperative closures. Contribution / Impact This study aimed to identify the most relevant CAMEL indicators for classifying cooperatives at higher risk of closure and to predict closure probabilities using machine learning. The main contribution lies in identifying key financial indicators for monitoring the performance of credit cooperatives. It is expected that this study will contribute to improving the oversight of these institutions and strengthening the cooperative system, enabling proactive interventions to reduce their vulnerabilities.