Files

ISBA_DP_2023-25.pdf
  • Open Access
  • Adobe PDF
  • 1.33 MB
Download

Details

Authors
Abstract
This article explores the application of variational autoencoders (VAEs) to insurance data. Previous research has demonstrated the successful use of generative models, particularly VAEs, in various domains such as image recognition, text classification, and recommender systems. However, their application to insurance data, specifically heterogeneous insurance portfolios with mixed continuous and discrete attributes, remains unexplored. This study introduces novel insights into utilizing VAEs for unsupervised learning tasks in the actuarial field, including dimensionality reduction and synthetic data generation. We propose a VAE model with a quantile transformation of continuous data and a reconstruction loss that combines categorical cross-entropy and mean squared error, along with a KL divergence-based regularization term. The architecture of our VAE model eliminates the need for pre-training layers to fine-tune categorical features representations. We analyze our VAE's ability to reconstruct complex insurance data and generate synthetic insurance policies using a motor portfolio. Our experimental results and analysis highlight the potential of VAEs for addressing challenges related to privacy and anti-discriminatory regulations, bias correction, and data availability in the insurance industry.
Affiliations

Citations

  • APA
  • Chicago
  • FWB

Jamotton, C., & Hainaut, D. (2023). Variational autoencoder for synthetic insurance data (LIDAM Discussion Paper ISBA 2023/25). https://hdl.handle.net/2078.5/104110