Regional risk classification in emerging automobile insurance markets using Bayesian Hurdle models

Abstract

Abstract Accurate risk classification in automobile insurance claim frequency in emerging markets remains a critical challenge for actuarial practice and regulatory oversight. Claim data are characterized by sparsity, overdispersion, and structural zeros, reflecting low-frequency claim occurrences and heterogeneous risk exposures across regions. The study compares count models for sparse automobile insurance claims and developed a Bayesian hierarchical Negative Binomial (NB) Hurdle model using 10,149 policies from Tanzania (2014–2024). Poisson models were found inadequate due to their restrictive equidispersion assumption, while the NB regression effectively addressed overdispersion. To further account for excess zeros and regional heterogeneity, two-part Hurdle models, including Bayesian NB Hurdle, were employed. Model comparisons using log-likelihood, AIC, BIC, and cross-validation demonstrated that the Bayesian NB Hurdle model provided superior fit and predictive accuracy. Results reveal significant heterogeneity in claim behaviour across Tanzania. Dar es Salaam showed the highest claim frequencies, followed by Arusha, while Mbeya, Dodoma, and Mwanza exhibited lower claims, highlighting clear spatial variation in insurance risk. Covariates such as policy type, driver age, and vehicle age showed minimal influence on claim frequency. These results underscore the importance of accounting for regional heterogeneity in sparse claim data and demonstrates that the Bayesian NB Hurdle model effectively addresses both overdispersion and zero inflation, delivering robust and regionally sensitive predictions for automobile insurance claims. The findings further indicate that sparse stochastic systems in emerging insurance markets are best modelled using flexible two-part count frameworks, which provide reliable probabilistic forecasts, enhance risk segmentation, and support more informed actuarial decision-making.