A Machine Learning Approach to Forecast International Trade: The Case of Croatia

Abstract

Background: This paper presents a machine learning approach to forecast Croatia's international bilateral trade. Objectives: The goal of this paper is to evaluate the performance of machine learning algorithms in predicting international bilateral trade flows related to imports and exports in the case of Croatia. Methods/Approach: The dataset on Croatian bilateral trade with over 180 countries worldwide from 2001 to 2019 is assembled using main variables from the gravity trade model. To forecast values of Croatian bilateral exports and imports for a horizon of one year (the year 2020), machine learning algorithms (Gaussian processes, Linear regression, and Multilayer perceptron) have been used. Each forecasting algorithm is evaluated by calculating mean absolute percentage errors (MAPE). Results: It was found that machine learning algorithms have a very good predicting ability in forecasting Croatian bilateral trade, with neural network Multilayer perceptron having the best performance among the other machine learning algorithms. Conclusions Main findings from this paper can be important for economic policymakers and other subjects in this field of research. Timely information about the changes in trends and projections of future trade flows can significantly affect decision-making related to international bilateral trade flows.