Forecasting central bank policy rates using machine learning and deep learning approaches

Accurate forecasting of central bank policy rates is essential for effective monetary policy, stable market expectations, and overall macroeconomic stability. In emerging economies such as Mongolia, traditional econometric models, including the Taylor Rule, ARIMA, and SVAR, often fail to adequately capture nonlinear relationships, time dependencies, and structural changes in the economy. To address these limitations, this study develops and evaluates advanced forecasting approaches based on hybrid combinations of machine learning and deep learning models. The analysis uses a monthly dataset consisting of 26 macroeconomic variables from January 2008 to December 2024. Seven forecasting models are constructed and evaluated using RMSE, MAE, and R² performance measures. The results indicate that hybrid models, particularly XGBoost combined with Gradient Boosting and LSTM integrated with XGBoost, achieve the highest forecasting accuracy, with the best model attaining an R² value of 0.9355. Overall, the hybrid approaches outperform both conventional econometric models and individual machine learning or deep learning models in capturing complex macroeconomic dynamics and structural shifts. These findings offer a reliable data-driven framework to support monetary policy decisions in Mongolia and provide a methodology that can be applied to other emerging economies with similar economic conditions.