Risk Prediction in Cash-in-Transit Security Using Machine Learning: A Data-Driven Performance Evaluations

Abstract

Cash-in-transit (CIT) transactions are the transportation of banknotes and cash between ATMs, cash processing centers, banks and retailer locations. Robbery, route compromise, leaking of information from inside the group, vehicle attacks, abnormal dwell times, and unsafe route deviations are all rare but high-impact threats that are faced during these trips. This study designs a machine learning-based framework to predict the risk level CIT for each trip by leveraging a synthetic operational benchmark of 30,000 CIT trip records of which 847 were high-risk or incident ones, corresponding to 2.82% CIT incident rate. In addition, the data set consisted of temporal variables, route variables, variables describing the type of vehicle, crew member, cash value, location risk, and journey anomaly, with these divided into 75% train and 25% test stratified samples.

Three algorithms were tested: logistic regression, random forest and gradient boosting. The following metrics were used to evaluate the model performance: accuracy, ROC-AUC, precision, recall, F1-score, brier score, Precision@5%, and Recall@5%. Random forest had the highest accuracy (0.9178) and the best calibration obtained on the probability of the model, with the lowest Brier score (0.0467). Logistic regression performed best across the board (ROC-AUC 0.8449 and recall 0.292), discriminating and detecting higher risk trips with greater ability. Gradient boosting performed best in terms of sensitivity (recall) with a score of 0.3868 (Recall@5%), capturing 38.68% of the estimated incidents when only the top 5% riskiest trips are marked, and in terms of precision (0.276), and F1 score (0.282). The findings demonstrate that machine-learning risk scoring could complement the CIT control rooms, allowing for prioritization of trips for review, trip assignment for escorts, adjustment of dispatch, and better oversight of security, while leaving room for human decision on trips' security to be made by the control room's team.