Deep Learning Architectures for Non-Life Insurance Claim Severity Modelling in Emerging Markets

This study develops and validates deep learning neural network architectures for non-life insurance claim severity modelling using Nigerian insurance claims data, contributing original methodological and empirical insights to the intersection of actuarial science and artificial intelligence in emerging market contexts. Claim severity modelling is a core actuarial task underpinning product pricing, IBNR reserve estimation, and catastrophe loss projection. While neural network applications to insurance claim modelling have proliferated in European and North American literature, their performance and adaptation requirements in emerging market data environments with limited observations, high noise, and significant structural heterogeneity are poorly understood. This study assembles a unique dataset of 62,000 individual non-life insurance claims from motor, fire, and marine lines in Nigeria from 2016 to 2023, sourced from four cooperating insurers. Multilayer perceptron, convolutional, recurrent, and attention-based transformer neural network architectures are developed, trained, and compared against traditional actuarial benchmarks including lognormal GLMs, gamma GLMs, and Tweedie compound frequency-severity models. Transfer learning from South African insurance datasets is evaluated as a technique for improving model stability with limited Nigerian data. Findings reveal that attention-based models achieve the best predictive performance on the Nigerian holdout dataset but require transfer learning augmentation to avoid overfitting with smaller claim portfolios. Traditional GLMs remain competitive for frequency prediction but are significantly outperformed in severity tail estimation. The study contributes an original deep learning actuarial severity modelling methodology for data-scarce emerging markets and recommends NAICOM encourage industry claims data pooling to enable broader AI model development.

Keywords: deep learning, claim severity, neural networks, non-life insurance, actuarial science.