Generalised Linear Models Versus Machine Learning in Claims Frequency Prediction: Nigerian Motor Insurance Evidence

This study provides a rigorous empirical comparison of generalized linear models and machine learning algorithms for predicting motor insurance claims frequency using a large Nigerian policy database, contributing original evidence on the relative merits of these approaches in an emerging market context. Claims frequency models are the foundation of risk-based insurance pricing, and improving their predictive accuracy directly benefits pricing fairness, adverse selection management, and insurer profitability. While machine learning comparisons with GLMs have been extensively studied in European and North American markets, Nigerian evidence is limited. This study uses a database of 78,000 motor insurance policy records from four Nigerian insurers, covering policy terms from 2019 to 2023 with matched claims outcomes. Predictive models compared include Poisson GLM, negative binomial GLM, gradient boosting machine, extreme gradient boosting, random forest, and neural network architectures. The dataset is split into training (70 percent) and holdout (30 percent) samples, with hyperparameter tuning using cross-validation. Model performance is evaluated using Poisson deviance, Gini coefficient, and calibration metrics on the holdout sample. Findings reveal that XGBoost achieves the highest predictive performance with a Gini coefficient 22 percent higher than the best-performing GLM. However, GLMs with well-engineered interaction terms close approximately 60 percent of the performance gap. The study concludes that ML models offer a meaningful accuracy advantage in the Nigerian motor insurance context and recommends regulatory guidance on algorithmic pricing fairness standards.

Keywords: GLM, machine learning, motor insurance, claims frequency, actuarial pricing.