A temporal fusion transformer network for enhanced international currency exchange prediction

The international currency exchange market, known for its speculative and volatile nature, operates within a complex financial ecosystem. Compared to other investment domains, it exhibits lower efficiency, highlighting the need for advanced predictive frameworks. This study introduces TFT-ICENet, a temporal fusion transformer network enhanced with temporal attention mechanisms, designed to improve accuracy in forecasting exchange rates. TFT-ICENet uses temporal attention mechanisms to extract data from various historical intervals and integrates long-term and short-term features through TFTs, significantly enhancing multi-step prediction precision. Extensive experiments were conducted using four comprehensive datasets capturing the volatility of major currencies, including the US dollar, euro, renminbi, yen, and Australian dollar. The model was compared against state-of-the-art algorithms to evaluate its generalizability and robustness. Results show TFT-ICENet’s superior predictive performance, with reduced mean squared error and high R² scores, reflecting its accuracy. It also outperforms baseline models in mean absolute error, confirming its reliability. TFT-ICENet not only surpasses existing methodologies but also sets a new standard for analyzing the currency exchange market, offering investors a powerful tool for decision-making and strategic planning.