Multi-Layer Economic Forecasting Framework: A Hybrid Approach Combining Statistical Models and Neural Networks
DOI:
https://doi.org/10.56830/WRBA11202511Keywords:
Economic forecasting, ARIMA, LSTM, hybrid models, neural networks, ensemble learning, meta-learningAbstract
Economic forecasting has continued to be a thorn in the flesh because of the nonlinear, regime-dependent, and volatile character of macroeconomic time series.
This paper presents sequential enhanced predictive framework (EPF) combining three new modelling elements namely Adaptive Regime-Switching ARIMA with Dynamic Error Bands (RS-ARIMA-DEB), Dual-Channel Temporal Residual Encoder Network (DC-TREN), and Error-Topology Aware Meta Fusion Network (ETA-MFN) to enhance the forecasting of U.S. GDP growth using St. Louis Fed Economic News Index (ENI).
The model is based on a hybrid architecture that integrates linear statistical modelling, nonlinear residual learning as well as topology-based fusion.
Empirical analysis shows that there are significant improvements in accuracy, stability and directional forecasting performance, where the EPF has a directional accuracy of 98 and outperform classical econometric models, single neural networks and traditional hybrids in numerous economic regimes including financial crises and pandemics.
The findings indicate the strength and flexibility of the EPF, and it can be concluded that the tool is a strong instrument of macroeconomic forecasting and policy making in real-time.
