The current insurance premium zoning system in British Columbia, managed by the Insurance Corporation of British Columbia (ICBC), is based on fixed geographic boundaries that often fail to reflect actual crash risk patterns. This static approach can group together regions with significantly different risk profiles, leading to cross-subsidization and misaligned premium structures. This study proposes a data-driven alternative that leverages detailed territory level crash data to improve both fairness and accuracy in premium setting. Using a comprehensive dataset encompassing crash severity, frequency, and contextual variables, four premium generation methods such as structural, bimodal, normal, and uniform were tested.
To uncover regional risk patterns, clustering algorithms (K-Means, DBSCAN, and Hierarchical Clustering) were applied and compared against ICBC’s existing territorial zones to identify spatial inconsistencies. In parallel, supervised machine learning models (Random Forest, XGBoost, LightGBM, and CatBoost) were used to predict premiums and evaluated using standard performance metrics.
The results show that clustering reliably identifies territorial misalignments, and only the structurally generated data preserves realistic pricing relationships. These findings suggest that a machine learning driven rezoning framework can enhance actuarial precision, promote equity, and support more transparent premium allocation across British Columbia.
