A BAYESIAN VAR MODEL FOR INFLATION: THE CASE OF AZERBAIJAN

This study examines the effectiveness of Bayesian Vector Autoregressive (BVAR) models in forecasting consumer price inflation in Azerbaijan. Given the country's limited and often low-frequency macroeconomic datasets, traditional forecasting models frequently yield poor forecasting accuracy. To address this, we construct BVAR models using three alternative priors, namely Minnesota, Normal-Wishart, and Sims-Zha Normal-Wishart (dummy observations) and compare their forecasting accuracy with benchmark models: a univariate random walk and a standard unrestricted Vector Autoregressive (VAR) model. Using quarterly data from 2003Q1 to 2024Q2, the study divides the sample into estimation and pseudo-out-of-sample forecasting periods. We evaluate forecast accuracy using relative root mean squared forecasting errors (RMSFEs), while the Diebold-Mariano (DM) test is employed to assess the statistical significance of forecast differences. The models incorporate key domestic and external drivers of inflation, including the M2 money supply, manufacturing producer prices, real non-oil GDP, nominal effective exchange rates, and foreign inflation. The results show that all BVAR models outperform the random walk model across nearly all forecast horizons, while the BVAR with dummy observations prior consistently yields the lowest RMSFEs. Although Minnesota has underperformed in short-term forecasts, it improves in accuracy over longer horizons. Compared to the VAR benchmark, the Sims-Zha Normal-Wishart prior demonstrates clear superiority, confirmed by statistically significant DM test results. The study contributes to the literature on macroeconomic forecasting in developing economies and provides practical implications for policymakers. Future research may focus on extending the framework to incorporate time-varying parameters, high-frequency indicators, or density forecasts for inflation uncertainty.