ABSTRACT

The process of selecting an appropriate econometric model represents a historical and enduring challenge within the field. This challenge stems from the inherent complexity of reality, characterized by dynamic social structures and laws that undergo continual change. The multitude of methods and criteria employed by economists and empirical researchers to discern the most fitting model from an array of candidates contributes to a lack of clarity regarding the relative efficacy of these approaches.

Commonly used methods for assessing statistical procedures involve Monte Carlo experiments, where data is generated based on predetermined processes. However, a notable limitation arises from the specific set of assumptions under which the data is generated in these experiments. This raises concerns about the applicability of the findings to real-world data, where the validity of assumptions may be questionable.

In response to these challenges, this study adopts a real data-based comparison approach to assess model selection procedures. The primary metric for evaluating performance is the forecast error, calculated as the difference between actual and predicted values. This method, referred to as real data-based comparison, offers a more practical and applicable means of assessing the performance of econometric procedures. The objective is to discern the most effective procedure for selecting models under real-world conditions.

In instances where a variable of interest is associated with a multitude of theoretical models, each characterized by distinct sets of independent variables, the formulation of a generalized unrestricted model (GUM) becomes progressively impractical and, in certain cases, unattainable. Notably, in the domain of Growth Econometrics, Darlauf’s compilation of growth models reveals an aggregate inclusion of over 150 independent variables. Introducing a single lag for all variables amplifies the total count of regressors to 300, rendering the estimation of a GUM unviable, particularly when dealing with annual data.

In light of these challenges, our present study strategically narrows its focus to models that boast a minimum of three substantiated studies within the existing literature. This judicious selection process aims to circumvent the methodological complexities inherent in attempting to accommodate the comprehensive array of models, thereby enhancing the feasibility and rigor of our empirical investigation.

The study evaluates a spectrum of model selection procedures, including those based on information criteria, shrinkage methodologies (such as LASSO, Adaptive LASSO, WALS, and elastic net), coefficient consistency procedures (exemplified by Leamer’s and Sala-i-Martin’s extreme bound analysis), and automatic model selection procedures (including encompassing and automatrix).

To validate the utility of these procedures, the study employs two real-life problems: selecting a model for the balance of trade and a model for economic growth. Given the paramount importance of these variables in macroeconomics, understanding their determinants is crucial. The plethora of theories leads to a variety of econometric models, necessitating model selection to guide policymakers. Consequently, the study seeks to identify the most suitable models for each variable and determine the bestperforming model selection procedure based on forecast performance. In essence, the research aims to offer a comprehensive solution to the dual challenges of model selection in the contexts of both the balance of trade and economic growth.