Journal of Basic and Applied Research International

This study develops and evaluates AI-based models for forecasting key macroeconomic indicators—inflation, GDP growth, and unemployment—with a focus on policy-relevant horizons. We use quarterly data for the United States and Vietnam from 2000Q1 to 2024Q4, combining official macroeconomic series with high-frequency financial and sentiment indicators to construct an enriched feature set. Traditional econometric benchmarks (ARIMA, VAR, Bayesian VAR) are compared with deep learning architectures (LSTM, GRU, TCN, Transformer) and a hybrid VAR–LSTM framework that models VAR residuals with a nonlinear recurrent network (Box et al., 2015).

Model performance is assessed using out-of-sample Root Mean Squared Error, Mean Absolute Error, Directional Accuracy, Theil’s U-statistic, and Diebold–Mariano tests of predictive accuracy. Across all three indicators, deep learning models significantly outperform the econometric benchmarks, with the hybrid VAR–LSTM delivering the best overall performance. For inflation and unemployment, the hybrid model improves directional accuracy by more than 15 percentage points relative to ARIMA and VAR and remains robust during the COVID-19 and post-pandemic periods. SHAP-based explainability analysis highlights oil prices, exchange rates, money supply, and financial sentiment as key drivers of the forecasts. The results indicate that AI-augmented forecasting frameworks can enhance short-term macroeconomic projections and provide actionable inputs for monetary policy in both developed and emerging economies.