Graduate Institute of International and Development Studies Working Paper No. 07-2020
The purpose of this paper is to evaluate the performance of some leading univariate and multivariate models: ARIMA, the standard OLS VAR and Bayesian VAR models, in forecasting inflation in Bosnia and Herzegovina. Although the presented models are small and highly aggregated, they provide a convenient framework to illustrate practical forecast issues. Furthermore, they are a good starting point in the process of the forecast development. The empirical part of this paper estimates the domestic and international transmission effects on inflation and tries to find good predictors of the inflation. A variety of inflation indicators included in the VAR models are assessed as potential predictors of inflation. They have been suggested by economic theory and existing research. A pseudo out-of-sample forecast approach is employed to assess the models' performance at different horizons using a recursive strategy. The study then evaluates the relative forecast performance of univariate model and various alternative specifications of the VAR models and offers conclusions. The results confirm the significant improvement in forecasting performance at all forecast horizons when Bayesian techniques, which incorporate information from the likelihood function and some informative prior distributions, are used.
Bayesian VAR model selection inflation forecasting