Bayesian Smoothed Quantile Regression

Bayesian quantile regression based on the asymmetric Laplace distribution (ALD) likelihood suffers from two fundamental limitations: the non-differentiability of the check loss precludes gradient-based Markov chain Monte Carlo (MCMC) methods, and the posterior mean provides biased quantile estimates. We propose Bayesian smoothed quantile regression (BSQR), which replaces the check loss with a kernel-smoothed version, creating a continuously differentiable likelihood. This smoothing has two crucial consequences: it enables efficient Hamiltonian Monte Carlo sampling, and it yields a consistent posterior distribution, thereby resolving the inferential bias of the standard approach. We further establish conditions for posterior propriety under various priors (including improper and hierarchical) and characterize how kernel choice affects posterior concentration and computational efficiency. Extensive simulations validate our theoretical findings, demonstrating that BSQR achieves up to a 50% reduction in predictive check loss at extreme quantiles compared to ALD-based methods, while improving MCMC efficiency by 20-40% in effective sample size. An empirical application to financial risk measurement during the COVID-19 era illustrates BSQR's practical advantages in capturing dynamic systemic risk. The BSQR framework provides a theoretically-grounded and computationally-efficient solution to longstanding challenges in Bayesian quantile regression, with compact-support kernels like the uniform and triangular emerging as particularly effective choices.