A multi-feature deep learning approach for sustainable exchange rate forecasting

Enhance the accuracy of exchange rate fluctuation forecasting. By establishing a data-driven spatiotemporal Transformer (DDST) exchange rate forecasting model, spatial correlation is modeled separately for each time period to capture dynamic spatial relationships. Simultaneously, multiple time-span branches are employed to complement spatial features across branches, thereby improving the accuracy of spatial correlation modeling. Finally, channel attention and spatial attention mechanisms are introduced during the feature fusion stage to fully integrate predictive features, enabling the model to simultaneously forecast both local and global spatial correlations. Experiments conducted using US/CN exchange rate fluctuation drivers from September 2014 to September 2024 demonstrate that under the RMSE evaluation metric, the DDST exchange rate prediction model achieves accuracy improvements of 16.91% (1-day horizon), 6.34% (3-day horizon), and 3.24% (6-day horizon) compared to the baseline ensemble model. The DDST model demonstrates superior performance as a sustainable and robust exchange rate forecasting framework. • We propose FCT-Stock, a mixed-frequency anomaly detection framework that integrates a parallel CNN–Transformer with bidirectional bridging, along with feature coupling modules enhanced by coordinate attention and lightweight ECA to capture complementary local and global information between stocks and concepts. • On the SSE 50 mixed-frequency dataset, FCT-Stock outperforms strong baselines across MSE, MAE, HMSE, and HMAE, and achieves statistically significant DM tests, demonstrating superior generalization, stability, and robustness.