The research addresses the fundamental challenge of leveraging generative models to create high-fidelity synthetic non-life insurance premium datasets.
I propose a comprehensive comparative study benchmarking traditional Conditional Gaussian Mixture Models against state-of-the-art deep learning architectures, including Conditional Variational Autoencoders, Conditional Variational Autoencoders with a Transformer-based Decoder, Conditional Diffusion Models, and GPT-5.1 Large Language Models.
The validation framework integrates multiple complementary approaches: visualization, statistical tests, and prediction.