Comparing fare evasion severity by econometric and artificial intelligence models: An Italian case study

Fare evasion is a pressing issue in public transport networks, impacting the financial sustainability of Transit Agencies (TAs) and Public Transport Companies (PTCs). While prior studies have largely focused on the probability of fare evasion (or frequency), research on severity— e.g., the financial and operational impact of detected fare evasion cases—remains limited. This study addresses this gap by specifying, calibrating, and validating two prediction models for fare evasion severity using real-world survey data on passengers from a mid-sized Italian PTC. Two approaches are employed: an Econometric Approach (EA) that uses Logistic Regression Models (LRMs) and a Machine Learning Approach (MLA) leveraging an Artificial Neural Network Model (ANNM). Model performance is evaluated and compared using Confusion Matrices, Metrics robust to class imbalance (e.g., Area Under the Precision-Recall Curve, Balanced Accuracy), and Probability Calibration tools (e.g., reliability curves, Brier score). Probability thresholds (cut-offs) are enhanced to improve predictive performance under imbalanced conditions. Finally, each predictor effect is assessed for both models. Results indicate that the ANNM slightly outperforms the LRM in this case study, demonstrating higher predictive accuracy and a stronger ability to detect high-severity fare evasion cases. However, this gain entails a minor rise in false positives, reflecting the trade-off between predictive accuracy and calibration stability. The LRM remains valuable for policy analysis, offering consistent and interpretable probability estimates to help TAs/PTCs understand key factors influencing fare evasion severity. These findings provide critical insights for enhancing fare inspection policies and enforcement resource allocation.