Macroeconomic Forecasting Using Transformer-Based Time Series Models
Dina Fallah Massod, Hameed H. Khalaf, Mohsen Aued Farhan · Digitalen Hochschulbibliothek Sachsen-Anhalt (Universitäts- und Landesbibliothek Sachsen-Anhalt) · 2026
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
Digitalen Hochschulbibliothek Sachsen-Anhalt (Universitäts- und Landesbibliothek Sachsen-Anhalt) · Universitäts Frauenklinik · other · 2026
Abstract
Macroeconomic forecasting is essential for policy formulation, risk evaluation, and investment strategy development. However, conventional econometric models like ARIMA and VAR frequently encounter difficulties with high-dimensional data, structural breaks, and long-term forecasting issues. In this research, we create and test transformer-based architectures-Autoformer, ETSformer, iTransformer, and Informer-for multi-horizon macroeconomic forecasting of important indicators such as GDP growth, CPI inflation, the unemployment rate, and policy interest rates.
The models are compared to classical and deep learning baselines like ARIMA, VAR, Prophet, LSTM, and GRU using publicly available datasets from FRED, OECD, and the World Bank. Results from rolling-origin evaluation show that transformer-based models always do better than baselines over both short- and long-term periods, lowering RMSE and MAE by as much as 20-35%. Also, using attention heatmaps for interpretability analysis shows how important policy rates and commodity prices are during times of inflation. These results show that transformers are good for making real-world macroeconomic decisions because they are both accurate and clear.
Open the resolved locator · 10.25673/123732
Metadata licence: OpenAlex metadata (CC0)