Machine Learning-Based Forex Forecasting Using a Novel Signal Decomposition Method

Abstract

This paper proposes a novel signal decomposition (SD) method for ML-based forex forecasting. Traditional SD techniques like empirical mode decomposition (EMD) often face morphological instability in mode extraction, leading to pattern skewness and inaccurate ML pattern identification. Additionally, they generate modes with broad frequency bands, increasing computational costs and reducing accuracy. The proposed method addresses these issues by ensuring controlled decomposition levels with adjustable frequency bands, maintaining morphological continuity for more interpretable ML patterns. It employs a cascade of high-pass filters (HPFs), where each level extracts a mode and passes the residual to the next HPF until the desired SD level is reached. These modes are then used to train individual LSTM-based neural networks for joint price forecasting. Results on EURUSD rates demonstrate that the proposed method outperforms EMD in mode uniformity, extraction stability, and model training accuracy.