A Novel Bayesian Pay-As-You-Drive Insurance Model with Risk Prediction and Causal Mapping

Abstract

The modern insurance industry is increasingly adopting Pay-As-You-Drive (PAYD) models to align premiums with actual driving behavior, promoting fairness and incentivizing safer practices. This presentation introduces a Bayesian approach to PAYD insurance, combining the strengths of Naive Bayes classifiers and Bayesian Networks to enhance risk assessment and causal inference.

Our Naive Bayes model achieved an impressive 87.5% accuracy in predicting risk partitions, demonstrating its effectiveness in handling the probabilistic co-occurrence of driving-related variables. Complementing this, the Bayesian Network provided a Directed Acyclic Graph (DAG) to visualize causal relationships among key factors such as annual miles driven, geographic location, and vehicle fuel economy. This DAG not only improves interpretability but also challenges traditional actuarial assumptions, such as the relevance of geographic grouping in pricing, and highlights actionable insights for insurers.

Practical applications of this research include optimized premium pricing, targeted fraud detection, and eco-friendly incentives for fuel-efficient vehicles. The DAG’s causal pathways also offer a framework for future observational studies, addressing ethical and practical constraints in demographic research.

Despite limitations like geographic specificity and reliance on 2010 data, the study underscores the potential of Bayesian methods to refine PAYD models. The presentation will conclude with discussions on integrating advanced techniques like NLP and computer vision for further innovation in usage-based insurance.

This work contributes to equitable and dynamic insurance pricing, benefiting both insurers and policyholders. We invite attendees to explore these insights and their implications for the future of personalized insurance.

About the Presenter

Dr. Zichao Li is currently a researcher at University of Waterloo. He holds PhD in management science. He is currently conducting research on machine learning optimization algorithms. I leverage on traditional optimization techniques used in transportation model to improve deep learning’s knowledge graph structures. The main application area of his research is in financial market fraud and sentiment detection. His research interests are Pattern recognition for medical engineering; Big data and deep learning; Graph Convolutional Network; Neural Networks; Optimization Algorithms for big data; Reinforced Learning and Adversarial Learning; Self Supervised and Unsupervised Learning; Bayesian Optimization; RNN, LSTM, GNN; Knowlege-Graph Embedding; Explainable AI; Distributed Statistical Model. He also work as chief scientist at a fintech firm