Optimized BiLSTM and GRU Models Using QHBM for Forex Price Prediction

Abstract

The foreign exchange market is highly volatile and complex, making accurate price prediction challenging. This study aims to develop an optimized deep learning framework for predicting daily closing prices of seven major currency pairs (AUDUSD, EURUSD, GBPUSD, USDCAD, USDCHF, USDCNY, and USDJPY) by integrating Bidirectional Long Short-Term Memory (BiLSTM) and Gated
Recurrent Unit (GRU) models with optimization strategies. Historical data from the Federal Reserve Economic Data were evaluated using Fixed Date Split and Walk Forward Validation (WFV), where WFV consistently achieved better performance than the fixed date. To enhance model performance, hyperparameter optimization was conducted using the Queen Honey Bee Migration (QHBM) algorithm, a metaheuristic approach inspired by the migration behavior of queen bees, divided into two characteristics: high learning rate and low learning rate. The optimized models achieved performance improvements of approximately 10-70% in MAPE and RMSE compared to the baseline models, while maintaining high R2 values. The results indicate that optimal configurations are pair-specific, where
most currency pairs perform best with a high learning rate and high unit settings, while AUDUSD achieves superior performance with a low learning rate and low unit configuration. This study contributes a novel integration of WFV and QHBM-based optimization. Adaptive deep learning models with proper validation significantly improve forecasting accuracy, robustness, and generalization for
financial decision-making and algorithmic trading applications.