ESTIMATING AND FORECASTING TRADE FLOWS BY PANEL DATA ANALYSIS AND NEURAL NETWORKS

This paper aims to investigate bilateral trade flows among EU15countries from 1964 to 2003 with their determinants by using panel data analysis and neural network modeling. When we compare explanatory power of both models, it appears that neural networks can explain larger variation in bilateral exports compared to the panel data analysis. Moreover, in comparing out-of-sample forecasting performances of panel model and neural networks, it is seen that neural networks produce much lower MSE which makes them superior to the panel model. One of the main relative benefits of the neural network model is nonlinearity, as it uses sigmoid functions instead of linear functions as building blocks. This partly explains its success in our study. Another advantage of neural networks is that they make no a priori assumptions about the population distribution and the relationship between explanatory variables and the dependent variable.