Analysis of the factors affecting the general budget deficit in Iraq using stepwise regression, hierarchical regression, and artificial intelligence models

Abstract

The public budget deficit is a recurring structural problem in rentier economies, but the factors affecting the public budget deficit have yet to be quantitatively confirmed in the application of literature in the case of Iraq. The aim of this study is to reveal the most influential factors affecting the Iraqi public budget deficit and to order them in descending order based on their impact based on an analytical framework that combines traditional statistical analysis with machine learning models, and monthly data with the number of observations equal to 194 and the number of economic and financial data obtained from official sources, including the Central Bank of Iraq, the Ministry of Finance, and the Ministry of Planning. The research was carried out in two phases. The first step was to conduct diagnostic tests to determine whether the linear regression assumptions held true; the results suggested that there was no significant issues of multicollinearity as all the Variance Inflation Factor (VIF) values were below the critical threshold. The second stage comprised the estimation of three linear models, including the full model, stepwise regression model based on the Akaike Information Criterion (AIC), and hierarchical regression model with an 80/20 training–testing data split, and the comparison of these models with two machine learning models (Random Forest and XGBoost). The results indicated that stepwise regression model was significant, with a smaller number of variables than the full model, but also showed superior out of sample predictive accuracy as compared to the other models. In addition, the relative importance of four variables was strongly agreed by the models, with oil revenue, actual spending, domestic investment and non-oil revenue being most important in explaining the differences among the public budgets. Others, however, such as public debt, showed some variations in the importance of the variable between the linear and the nonlinear models, indicating that there are more complex relationships that could be investigated. The study finds that combining traditional regression methods with measures of variable importance from machine learning algorithms gives a more powerful tool for determining the factors behind the budget deficit which increases the reliability of the results when cross-validating the findings in various ways from different analytical methods.