A hybrid approach for accurate forecasting of exchange rate prices using VMD-CEEMDAN-GRU-ATCN model

Abstract

The foreign exchange (Forex) market has greatly influenced the global financial market. While Forex trading offers investors substantial yield prospects, some risks are also involved. It is challenging to accurately model financial time series due to their nonlinear, non-stationary and noisy properties with an uncertain and hidden relationship. Thus, developing extremely precise forecasting techniques is crucial for investors and decision-makers. This study introduces a novel hybrid forecasting model, VMD-CEEMDAN-GRU-ATCN, designed to improve Forex price prediction accuracy. To begin with, our proposed model utilizes the variational model decomposition (VMD) technique for breaking down raw prices into multiple sub-components and residual terms. The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) technique is utilized to extract features from the residual terms, which involves further decomposition and analysis of these complex information-containing terms. These sub-components are then predicted by the gated recurrent unit (GRU) model. To enhance the effectiveness of our hybrid model, we include the open, high, low, and close prices and seven Forex market technical indicators. Finally, an attention-based temporal convolutional network (ATCN) model is used to obtain the Forex price forecasts. For both one-step and multi-step ahead forecasting, our proposed VMD-CEEMDAN-GRU-ATCN model has demonstrated superior and consistent performance in predicting USD/PKR exchange rate price series.