Rezervace pohledávek v éře strojového učení

Zelený, Ondřej

This thesis evaluates whether claim-level machine learning and deep learning methods improve non-life insurance claims reserving relative to classical actu- arial models, once point accuracy, bias, predictive uncertainty, computational cost and interpretability are jointly assessed. The empirical comparison uses a synthetic micro-level portfolio generated by the Individual Claims History Simulation Machine, in which the true ultimate claim amounts are observable and every method can be scored against the same reference. Three classi- cal methods, Chain-Ladder, Cape Cod and a generalised linear model with over-dispersed Poisson variance, are compared with an XGBoost claim-level pipeline, a trajectory-aware claim-level neural network and a triangle-based neural network. All models are evaluated in a rolling-origin design over five val- uation years, with point forecasts supplemented by bootstrap and seed-based predictive-uncertainty diagnostics. The claim-level neural network attains the lowest mean absolute error, but at the cost of substantially wider predictive distributions. XGBoost is the most balanced modern method, combining the lowest mean absolute bias with predictive variability comparable to the classi- cal benchmarks. Classical methods remain competitive once uncertainty, run- time and...

guest :: login Digital Repository
		Search		Submit		Help		About