Effective multinational trade forecasting using LSTM recurrent neural network

Abstract

Changes in foreign trade (export and import) constitute a crucial topic in international economics, international business management, and economic development. Numerous academics and industry leaders have sought effective means of forecasting foreign trade. However, with the uncertain nature of trade trends, obtaining accurate forecasts is a challenge. To analyze ten countries’ trade data, this study developed an effective foreign trade forecasting method that relies on a neural network with long short-term memory (LSTM); the results validated the effectiveness of the proposed method. This study is based on the economic theory that the two-way causal relationships present in trade data can improve trade forecasting. A multivariate LSTM-based method is proposed and exploited to extract temporal changes from trade data and provide effective trade forecasting. A comparison was conducted to understand the performance of the proposed method against time-series and economic structural models. The empirical results indicate that the method can appropriately model temporal information regarding uncertainty trends in foreign trade data. The method achieved almost perfect forecasting performance for data previously difficult to predict; in most cases, it had smaller values of root mean square error (RMSE) and mean absolute percentage error (MAPE) than did time-series models and economic structural models. On the export forecast, RMSE improved by 17.048% and MAPE by 1.463%, and for imports, RMSE improved by 40.939% and MAPE by 1.806%. This paper demonstrates the feasibility of the theoretical synthesis and provides a theoretical basis for interdisciplinary research in foreign trade forecasting.