Conditional Granger Causality and Genetic Algorithms in VAR Model Selection

Vasile George Marica,Alexandra Horobet

doi:10.3390/sym11081004

Abstract

Overcoming symmetry in combinatorial evolutionary algorithms is a challenge for existing niching methods. This research presents a genetic algorithm designed for the shrinkage of the coefficient matrix in vector autoregression (VAR) models, constructed on two pillars: conditional Granger causality and Lasso regression. Departing from a recent information theory proof that Granger causality and transfer entropy are equivalent, we propose a heuristic method for the identification of true structural dependencies in multivariate economic time series. Through rigorous testing, both empirically and through simulations, the present paper proves that genetic algorithms initialized with classical solutions are able to easily break the symmetry of random search and progress towards specific modeling.

Highlights

The vector autoregression (VAR) is an alternative to complicated dynamic stochastic equilibrium models that require expert knowledge about economic theory, with explicit assumptions being made about business cycles, economic growth and economic agents
Model identification through variable and coefficient selection in vector autoregressions represents a continuous challenge for the scientific community
Estimation methods into an evolutionary framework defined by genetic algorithms. Both through simulation and empirical data sets, we show that genetic algorithms can partially overcome the “dimensionality curse” for those cases where data is scarce

Summary

Introduction

In the field of econometrics, the vector autoregression (VAR) model has been used extensively by researchers, finance professionals and policymakers for describing the linear dependencies between economic variables and their past realizations. This framework makes use of simultaneous and linear equations to identify the underlying causal connections in complex systems, benefiting recently from information theory proofs that causality in the Granger sense and transfer entropy are equivalent [1]. Since its proposal [2], the VAR has proven itself as a reliable tool for forecasting, being extensively used in macroeconomics with ramifications to neuroscience [3] and genetics [4] It is mainly data-driven, with intervention from the researcher being required mainly for the customization of the model to the dataset, on the estimation procedure itself. The VAR is an alternative to complicated dynamic stochastic equilibrium models that require expert knowledge about economic theory, with explicit assumptions being made about business cycles, economic growth and economic agents

Methods

Results

Conclusion