Abstract:
GDP forecasting is a major issue that requires accurate estimates on economic planning, policy formulation, trade policies and development objectives. Conventional statistical and machine learning models have delivered encouraging outcomes but have focused more on country level measures and have ignored the fact that countries are intricately linked through international trade. To overcome this shortcoming, this study introduces a GDP prediction method based on Graph Neural Network (GNN) which combines node-level economic signals and edge-level trade characteristics. The analysis makes use of the World Bank data (GDP, population, consumer price index, unemployment rate) and the BACI international trade database in order to create annual graphs, with the country being the node and the key trade relations the edges. Node features are based on economic indicators and edge features based on trade volumes of the ten most traded products worldwide. The Graph Attention Network v2 (GATv2) model incorporates multi-head attention, batch normalization and residual connections to learn country-level log-GDP representations and predictors. Through experimental analysis, the GATv2 model with edge features is shown to significantly outperform baseline machine learning models (linear regression, random forest) and previous graph-based models (GCN, GAT) across various evaluation metrics, including Mean Squared Error (MSE) and Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and R2 score. The findings underscore the need to integrate trade-based relational information in order to provide accurate macroeconomic forecasts. The paper will be relevant to the existing literature on graph-based economic modeling as it will illustrate the usefulness of attention-based GNNs in modeling dependencies in global trade. The given framework can be applied to other macroeconomic indicators and provide substantial information to policymakers, economists, and researchers in order to create data-driven economic policies.
