Dynamic Forecasting of Regional Retail Sales and Identification of Influencing Factors Based on Bayesian Spatiotemporal Models

Regional retail sales forecasting is critical for policymaking and business strategy, yet existing methods typically neglect spatial dependence or fail to quantify parameter uncertainty. This paper proposes a Bayesian spatiotemporal hierarchical model (BSTHM) integrating province-level spatial random effects and a temporal random walk, applied to a panel of 31 Chinese provinces over 2005-2022. Bayesian inference via the No-U-Turn Sampler yields an RMSE of 0.056 and [[MATH_EQ_001]] of 0.998, outperforming OLS, fixed-effects, random-effects, and random forest benchmarks. Regional GDP (b = 0.380; 95% CI: [0.330, 0.430]), per capita income (b = 0.117), and urbanization rate (b = 0.094) are identified as significant drivers, while population size is statistically insignificant. Pronounced spatial heterogeneity is documented, with coastal provinces exhibiting higher retail potential than western regions. These findings offer guidance for differentiated regional consumption policies.