Previsão de dificuldades financeiras em empresas latino-americanas via aprendizagem de máquina

• Ano de defesa: 2019
• Autor(a) principal: Marra, Vinicius Nogueira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Administração
• Palavras-chave em Português: Extreme Gradient Boosting,; Administração.; Análise de risco de crédito,; Administração de risco; Aprendizagem de máquina; Aprendizado por computador.; Administração de empresas - Aspectos econômicos.; Extreme Gradient Boosting; Credit risk analysis; Machine learning; CNPQ::CIENCIAS SOCIAIS APLICADAS::ADMINISTRACAO::ADMINISTRACAO DE EMPRESAS::ADMINISTRACAO FINANCEIRA
• Link de acesso: https://repositorio.ufu.br/handle/123456789/24750; http://dx.doi.org/10.14393/ufu.di.2019.947

The ability to foresee financial distress in business is paramount, as decisions regarding inappropriate credit concessions may have direct and indirect financial consequences throughout the economy. Predicting bankruptcies and measuring credit scores are two important research topics, in both accounting and finance, and this study aims to demonstrate the direction of academic researches on credit risk management from the perspective of machine learning. Interest in simultaneously examining these two topics emerged after noting that remnants and feelings of the last financial crisis still lingers in society, even a decade after it occurred. Paired with this fact, the increasing development and application of Artificial Intelligence in finance promises to increase the rigor and improvement of financial information analysis processes. Through a structured process to select scientifically relevant articles in the chosen research field, a final sample of 168 studies deemed able to guide the directions of use of machine learning - sub-area of Artificial Intelligence - applied to finance was obtained. From this sample of studies, a deeper analysis was carried out through their readings and it was identified that machine learning algorithms are being explored, improved and taken to extreme to detect subtle combinations and better describe credit risk. From the results and gaps found in the research, in this study proposed the use of a novel algorithm used to teach machines to seek the best classification according to the data previously provided. Xgboost is a model that appeared in 2016 and has been gaining notoriety due to its accuracy and its computational optimization. Because it is a relatively new model, studies with its application in finance, per se, have not been found. Thus, in this context, the proposed model was able to present better results when compared to Logistic Regression and Random Forest. The results were interesting as they identified other gaps and possibilities for future research. The intention to prove that the use of artificial intelligence adds value to credit analysis has been demonstrated and, in this way, it is believed that by advancing this research, interesting results should be achieved, both for suppliers, analysts, managers, and investors.