Flexible hierarchical risk modeling for large insurance data via NumPyro

Data analysis and individual policy-level modeling for insurance requires working with data exhibiting strong spatiotemporal correlations, non-Gaussian data, and relatively large volume with interesting hierarchical structure. In this work, we show that by employing gradient-based Markov chain Monte Carlo (MCMC) accelerated by graphics processing units, previous tradeoffs between rich model structure and scalability for inference no longer exist at the million-record level. By writing our model in NumPyro, we are able to use its off-the-shelf MCMC capabilities to fit a model with several nontrivial components including latent conditional autoregression and spline-based exposure adjustment with a speedup of 88\% relative to a CPU-based implementation.. We employ this model in a case study of 2.6 million policy-level claim count records related to automobile insurance in Brazil from 2011. We highlight how this modeling workflow can substantially enhance ongoing efforts towards risk assessment for highly multivariate correlated outcomes. Code and data are available at https://github.com/ckrapu/bayes-at-scale.