Applied & Interdisciplinary
Seyedeh Yasamin Hashemi; Omid Solaymani Fard
Abstract
As the Forex market becomes increasingly complex, accurate directional forecasting is essential for effective trading decisions. This study proposes a multi-timeframe framework to evaluate directional trend prediction for the EUR/USD exchange rate using Ichimoku Cloud-inspired feature engineering and ...
Read More
As the Forex market becomes increasingly complex, accurate directional forecasting is essential for effective trading decisions. This study proposes a multi-timeframe framework to evaluate directional trend prediction for the EUR/USD exchange rate using Ichimoku Cloud-inspired feature engineering and CNN-based hybrid architectures, including CNN, CNN-LSTM, and CNN-GRU. The study optimizes model hyperparameters using the Starfish Optimization Algorithm (SFOA). Specifically, the study investigates multi-step trend prediction for the H1 and H4 timeframes, and one-step-ahead prediction for the daily (D1) timeframe. Historical EUR/USD data are used to train continuous forecasting models, whose outputs are transformed into discrete trend classification labels. The framework ultimately employs a voting mechanism to determine the final trend decision at the daily (D1) timeframe. To assess the model's zero-shot classification capability and generalization, eight additional currency pairs are evaluated using the same evaluation criteria at the relevant stages to assess both individual model performance and overall framework effectiveness. Regression performance is assessed using MSE, MAE, and MAPE, while classification performance is evaluated using Accuracy and F1-score. Statistical robustness is further examined using the Diebold--Mariano test and bootstrap analysis. Experimental results demonstrate that incorporating GRU layers into CNN architectures consistently improves forecasting performance across multiple timeframes and currency pairs, with the proposed IICGS architecture achieving the most competitive overall performance. Selected results and implementation code are publicly available on GitHub.