Prototipo de modelo de redes neuronales para la gestión y priorización predictiva de cobranzas en empresas del sector financiero

Objective: The study aimed to develop a neural network-based prototype to estimate the probability of payment by customers with overdue obligations and to automate the prioritization of collection efforts in financial sector companies, to improve operational efficiency and portfolio recovery. Methodology: A quantitative, non-experimental, predictive approach was applied, structured according to CRISP-DM, which included business understanding, data analysis and preparation, modeling, evaluation, and simulated deployment. The model was implemented as a Multilayer Perceptron (MLP) with two hidden layers (ReLU) and a sigmoid output layer for binary classification; performance was evaluated using accuracy, recall, F1-score, and AUC-ROC, and its operational impact was compared against a traditional method in a simulation environment. Results: The model achieved an accuracy of 0.87, a recall of 0.82, an F1-score of 0.84, and an AUC-ROC of 0.90. In the simulated operational evaluation, a 25% improvement in efficiency (cases/day) and an 18% increase in the recovery rate were observed, along with a reduction in average handling time and less attention given to low-probability cases. Conclusions: The findings demonstrated the prototype’s viability in supporting data-driven prioritization decisions and optimizing resources in collections; as a future direction, expanding the volume and diversity of data and evaluating the model in real-world systems were proposed to strengthen its generalization and scalability.