Machine learning models for predicting payment status on an online car rental platform

Abstract

It has been demonstrated that the detailed data collected on online platforms are heterogeneous, semantically inconsistent, and weakly structured. Therefore, the use of machine learning for their aggregation, structuring, and analysis is well-justified. As a case study for developing machine learning models, the task of predicting the payment behavior of clients on an online car rental platform was considered. Input data were automatically generated based on users’ actions on the platform. Subsequently, the data were aggregated and structured through feature engineering, time field transformation, and the removal of redundant attributes to enhance model quality. Five classification models were developed: Support Vector Machine, Naive Bayes classifier, Logistic Regression, and two ensemble models (Soft Voting and Stacking). The results showed that Logistic Regression and ensemble models (particularly Stacking) achieved the best precision and recall, making them the most reliable for predicting on-time payments. Ensemble models, especially stacking, demonstrated high efficiency by combining the strengths of different base models. Although SVM can account for complex relationships between features, it showed the weakest performance in distinguishing payment statuses. The findings contribute to a better understanding of customer payment behavior and highlight the importance of choosing appropriate classification models for financial risk assessment. Future research will focus on improving model performance through enhanced feature selection, class imbalance correction, and the integration of additional data sources such as customer credit history. The use of such models can significantly improve automated risk management and enhance decision-making efficiency for companies dealing with payment obligations.