Explaining the Dynamics of Key Macroeconomic Indicators through Deep Learning State-Space Models

To ensure economic stability, accurately forecasting the effects of domestic and external factors has become increasingly critical. This study aims to develop a novel model to predict Mongolia’s macroeconomic dynamics by integrating theoretical economic relationships with deep learning methods. Quarterly macroeconomic data from 2015 to 2024 are employed, focusing on key indicators such as inflation, unemployment, GDP growth, and the policy interest rate. The interdependence among these variables is dynamically estimated using Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks. For comparison, traditional ARIMA and VAR models are also applied to assess the predictive performance of deep learning approaches. The results reveal that deep learning models achieve higher accuracy in short- and medium-term forecasts (MAPE ranging from 3.7% to 5.2%) and exhibit greater sensitivity to business cycle fluctuations and policy shifts. Moreover, by incorporating a theory-guided deep learning framework, the model’s interpretability is enhanced, enabling a more realistic representation of the dynamic trade-off between inflation and unemployment. The primary contribution of this research is the development of a theoretically consistent deep learning state-space forecasting model that bridges economic theory and artificial intelligence. The proposed framework provides practical insights for macroeconomic policy analysis, fiscal planning, and monetary decision-making in Mongolia.