From Prediction to Interpretability of Artificial Neural Networks: Application to Senegal's GDP Per Capita

This article evaluates the predictive capability of ANNs (Artificial Neural Networks) and attempts to interpret their "black box." It provides a detailed analysis of their architecture and explores different interpretability techniques for predictions suited to opaque models—both model-specific and agnostic approaches. The performance analysis of various ANN architectures reveals that the model with 2 hidden layers and 8 nodes remains the most effective. It offers the best balance between accuracy and generalization, with a high-test coefficient of determination (R² = 0.95) and minimal errors (RMSE = 0.084, MAE = 0.058). The graphical analysis highlights the complex relationships between several economic variables and their impact on GDP per capita. This type of ANN embodies a synthesis of technical sophistication and economic pragmatism, making it ideal for predictive or decision-making analyses in uncertain environments, such as that of Senegal. In summary, the key findings indicate that economic policies should focus on controlling inflation, strengthening productive investments, and ensuring efficient management of public spending. These results thus provide a valuable foundation to guide economic decisions and optimize strategies for economic and social development.