A1857
Title: Forecasting U.S. inflation with Bayesian state-space models: Predictive coverage and posterior diagnostics
Authors: Keongmuk Lim - Chungnam National University (Korea, South) [presenting]
Sangin Lee - Chungnam National University (Korea, South)
Abstract: Forecasting U.S. inflation is important for monetary policymaking, yet conventional time-series models such as AR and VAR exhibit miscalibration at long horizons, with empirical coverage falling below nominal levels. A Bayesian state-space model (BSSM) with a latent factor structure and policy inputs is applied to the inflation forecasting task, complemented by posterior diagnostics on the role of policy variables in the latent dynamics. The observation vector contains four price indices (CPI, core CPI, PCE, PPI), and the policy variables include short- and long-term interest rates; posterior inference is performed via Gibbs FFBS. In a 132-month rolling out-of-sample evaluation on FRED-MD (2015-2025), the BSSM achieves comparable point forecast accuracy to standard time-series benchmarks such as AR and BVAR, and provides higher long-horizon coverage than a comparable Bayesian VAR, reducing the gap between empirical and nominal coverage in those benchmarks.
