Greek - Turkish arms race: an approach using neural networks

Abstract

Greece and Turkey constitute a prime example of perennial rivalry. Relations between the two countries have been entangled in a series of unresolved disputes, which have brought Greece and Turkey on the brink of war on several occasions. Defence economics research and, more specifically, quantitative arms race research, has concerned itself with the rivalry between the two neighbouring countries. However, the approaches adopted, based mainly on traditional econometrics, have proved inconclusive as to whether a Greek - Turkish arms race exists. Meanwhile, researchers have expressed concerns over the statistical and methodological issues involved. This thesis revisits one of the approaches that have only scarcely been applied in related literature: the approach of neural networks. It aims to answer the question whether an arms race between Greece and Turkey exists, whilst unveiling the methodological issues involved. Three different models (A, B and C), based on neural networks, are developed for years 1963 - 2018, each one utilising a different set of input variables. Of them, only Model C achieves a performance considerably superior to that of the benchmark models used. An assessment of input significance through the use of SHAP values on Model C reveals that Turkey-related variables are not prime determinants of Greek defence spending. The analysis also highlights several technical issues: the ambiguity of what is termed an 'arms race', the intricacies involved in choosing input variables, data reliability issues, correlation significance testing and the impact of correlations between variables on input significance measures. These issues highlight the need for more rigorous research design and testing procedures, as well as for a careful interpretation of the results.