Prediction and calibration of payment default risk using machine learning in residential water service customers

Valles-Coral M.A.; Cuello-Sangama R.; Rengifo-Amasifen R.; Vidaurre-Rojas P.; Prieto-Luna J.C.; Holgado-Apaza L.A.; Injante R. · Revista Cientifica de Sistemas e Informatica (2026) · DOI: 10.51252/rcsi.v6i2.1265

Resumen

Water service payment delinquency limits the availability of resources required to sustain operations and hinders the prioritization of collection activities. This study aimed to evaluate machine learning models for estimating the payment default risk of residential customer bills at EMAPAB S.A. A total of 264,382 bills from 6,358 customers were analyzed using 51 predictors derived from historical billing and payment records. Random Forest, CatBoost, and a feedforward neural network were compared using Optuna optimization, stratified and grouped ten-fold cross-validation, and an independent temporal test. Random Forest achieved the best performance on the temporal test, with an accuracy of 0.9037, an F1-score of 0.6333, a ROC-AUC of 0.8681, and a PR-AUC of 0.6892. Its probabilities were subsequently calibrated using Platt Scaling and stratified into low-, medium-, and high-risk levels, accounting for 74.61%, 13.96%, and 11.44% of the bills, respectively. The results show that the proposed approach can identify bills with higher default risk and support the prioritization of the collection portfolio, although its operational impact requires prospective validation. © The authors.

Instituciones: Universidad Nacional Amazónica de Madre de Dios; Universidad Nacional Autónoma de Alto Amazonas; Universidad Nacional de San Martín