An expectile-based neural network approach for mixed-frequency economic forecasting

Timely and accurate forecasting of Gross Domestic Product (GDP) is critical for economic policymaking, yet it is complicated by the need to integrate high-frequency financial indicators with low-frequency macroeconomic data and to capture complex nonlinear dynamics. This study introduces a novel model, the Expectile Regression Neural Network for Mixed-Frequency Data Sampling (ERNN-MIDAS), designed specifically to address these challenges. Methodologically, our model advances upon existing frameworks by incorporating a fully differentiable expectile-based loss function, which enables more direct and stable parameter estimation than approximation-reliant methods. Empirically, we apply the ERNN-MIDAS model to forecast Indonesian GDP growth using quarterly data from 2001 to 2024, incorporating monthly high-frequency predictors like the Financial Stress Index. The results demonstrate that our proposed model consistently and significantly outperforms a range of competing models, including the state-of-the-art QRNN-MIDAS. Specifically, in out-of-sample tests, the ERNN-MIDAS reduces the Root Mean Square Error (RMSE) by up to 20% compared to its closest competitor under the year-on-year growth approach. This superior predictive accuracy, which is robust across both training and testing datasets, highlights the practical value of our methodological refinement and establishes the ERNN-MIDAS as a powerful and reliable tool for macroeconomic forecasting.