Tipo:	Trabalho de Conclusão de Curso
Título:	Análise preditiva de churn em uma fintech de pagamentos usando aprendizagem de máquina e redes neurais
Autor(es):	Costa, Arthur
Primeiro Orientador:	Vinhal, Gustavo Siqueira
metadata.dc.contributor.referee1:	Abadia, Fernando Gonçalves
metadata.dc.contributor.referee2:	Amaral, Nilson Cardoso
Resumo:	In payment fintechs, merchant retention is closely linked to processed volume sustainability and reacquisition costs. Unlike subscription-based services, churn rarely involves explicit cancellation and must be inferred from transaction dynamics affected by seasonality and operational fluctuations. This study frames churn prediction as a time-oriented supervised binary classification problem using merchant-level aggregated transactional time series. A dual labeling scheme combines hard churn (30 consecutive days without transactions) with imminent churn (sustained volume decline relative to a historical baseline), serving as an early-warning proxy for relationship deterioration. Temporal features from rolling windows of 7, 14, and 28 days capture recency, intensity, trend, and variability. A time-based train–test split and walk-forward cross-validation reduce leakage under non-stationarity. On the holdout set, XGBoost achieved ROC-AUC of 0.830, PR-AUC of 0.655 (≈2.02× above random), and Recall@10% of 0.223 (≈2.2× random). Calibration and interpretability analyses, including permutation importance and Shapley values, support threshold definition and businessoriented decision-making.
Palavras-chave:	Machine Learning, Churn prediction, Payment fintech.
CNPq:	CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO
Idioma:	eng
País:	Brasil
Editor:	Pontifícia Universidade Católica de Goiás
Sigla da Instituição:	PUC Goiás
metadata.dc.publisher.department:	Escola Politécnica
Tipo de Acesso:	Acesso Aberto
URI:	https://repositorio.pucgoias.edu.br/jspui/handle/123456789/10540
Data do documento:	12-Jun-2026
Aparece nas coleções:	TCC Engenharia de Computação

Os itens no repositório estão protegidos por copyright, com todos os direitos reservados, salvo quando é indicado o contrário.