New Framework for Autonomous Vehicle Liability Pricing Under ODD Shift

The widespread deployment of Automated Driving Systems (ADS) presents a fundamental challenge for insurance ratemaking: how to accurately price liability given limited historical data, constantly evolving operational design domains (ODDs), and non-stationary risk profiles across different software releases. This research introduces a hierarchical Bayesian credibility framework designed to tackle these issues. The framework pools data from various sources, including different cities, software versions, and geographical territories, by employing a learned ODD-similarity kernel. This kernel helps to identify and leverage similarities between different operational contexts, effectively nesting the traditional Buhlmann-Straub method as a special case. The framework was tested using 648 verified Waymo crashes from the NHTSA database, matched against 116 million miles driven across four U.S. metropolitan areas. Results indicate that city-aggregate credibility weights are moderate, partial pooling significantly outperforms no pooling, and the learned kernel demonstrates a detectable advantage, particularly as the number of deployed cities increases.