Description:
With the aim of understanding the usefulness of cross-country interactions and long memory dependencies in exchange rate return forecast, we propose, compare, and evaluate various neural network architectures. These architectures include the proposed deep latent factor (DLF) and deep prediction (PRED-DP) models for cross-sectional description and time series prediction, respectively, as well as alternative models with or without specific modular structures. Drawing upon a comprehensive methodological comparison, we discover the distinct cross-country interactions and long-term memory effects. The architectures that incorporate both aspects exhibit a significant enhancement in out-of-sample forecast performance. Additionally, we observe that the cross-sectional explanatory performance improves with the complexity of the model, whereas simpler architectures yield better forecasting performance in time series. The interest rate-related and technical characteristics possess greater explanatory powers in both cross-section and time series models. The proposed DLF model generates statistically and economically superior profits in currencies portfolios compared to alternative models