A GNN-based approach for accurate trade balance forecasting and interpretable analysis

Area of Science:

• Economics
• Computer Science
• Data Science

Background:

• Accurate trade balance forecasting is crucial for economic policy and strategy.
• Traditional methods often struggle with the complexity and interconnectedness of global trade.

Purpose of the Study:

• To develop and evaluate a machine learning pipeline for predicting trade balances across 229 countries.
• To compare the performance of a Graph Neural Network (GNN) against various deep learning and regression models.

Main Methods:

• Data preprocessing included handling missing values and feature normalization.
• Feature selection utilized a Random Forest Regressor.
• Evaluated models included GNN, Deep Neural Network (DNN), Transformer, Random Forest, and ensemble methods using regression metrics.

Main Results:

• The GNN model achieved superior performance with an MSE of 0.06, RMSE of 0.26, MAE of 0.18, and R² of 0.91.
• GNN demonstrated higher accuracy, robustness, and consistency compared to all other evaluated models.
• Residual plots and ROC curves validated the reliability and performance of the GNN.

Conclusions:

• Graph Neural Networks represent a powerful advancement in trade balance forecasting.
• The GNN model provides policymakers and economists with a more adaptable and precise tool for global trade analysis.
• This study enhances economic forecasting methodologies through the application of advanced machine learning.