Improving exchange rate prediction accuracy with 1D-CNN-BiGRU and walk forward analysis

Abstract

This study was conducted to improve the accuracy of forecasting the exchange rate of the Rupiah against the Japanese Yen, which was crucial for economic actors in managing risks, enhancing investment efficiency, and supporting adaptive financial policies. The research aimed to develop a forecasting method with the lowest possible error rate (MAPE) to address the limitations of conventional models and provide more accurate insights for decision- makers. The method used was a hybrid approach combining a 1D-Convolutional Neural Network (1D-CNN) and a Bidirectional Gated Recurrent Unit (BiGRU), along with a walk-forward analysis strategy to capture trend patterns and short-term volatility in exchange rates. The results showed that the Hybrid 1D-CNN-BiGRU model effectively predicted exchange rate fluctuations for the next five days, achieving a MAPE value below 1%, indicating higher accuracy compared to previous approaches. The conclusion of this study was that the Hybrid 1D-CNN-BiGRU model was an effective forecasting method for the Rupiah-to-Yen exchange rate, providing a stronger foundation for economic actors in making currency transaction decisions. The main contribution of this research was the implementation of a combined method that had not been previously applied in exchange rate forecasting, opening new opportunities for further research in exploring other hybrid models to improve prediction accuracy in various economic and financial fields.