Abstract: (coauthored with Vugar Ahmadov & Nijat Guliyev) In this study, we build factor augmented vector autoregressive (FAVAR) model for forecasting consumer price inflation and non-oil GDP of Azerbaijan for different time horizons. We use Principal Component analysis method to extract the unobserved factors from about 80 variables. The data covers the period from 2003Q1 to 2018Q4. The forecasting ability of FAVAR is compared mainly to univariate AR model and Benchmark VAR model and the criteria for comparison is root mean squared forecasting errors (RMFE).
      We first extract principal components from 77 series and use them to estimate the FAVAR. The results show that in this case forecasting ability of FAVAR is less than other models, especially, simple autoregressive (AR) model. In fact, AR performs better than all other models. In this step poor performance of FAVAR can be caused by the poor performance of the PCs so that first 5 PCs explain less than 30 percent of the variation. Next, we decrease the size of the group for PC extraction based on the correlation between the target variable and variables in the group for PC extraction. We end up with relatively small groups – consisting of about 30 variables - for both target variables.  Doing so, we manage to improve the performance of PCs significantly – 3 PCs explain more than 50 percent of variation. Unfortunately, the performance of FAVAR remains weaker compared to alternative models in shorter forecast periods. However, FAVAR outperforms VAR model in medium and long term forecasting horizon, though still underperforms compared to AR model.

The overall results show that FAVAR does not lead to significant improvement over the alternative – especially existing – models. What is more, our results confirm the results found in previous literature that AR models have better performance ability for Azerbaijan economy.