Executive Summary
This document presents a comprehensive, APA-formatted analysis of 24 ensemble learning studies for Forex forecasting conducted between 2021 and 2025. The analysis standardizes performance metrics across heterogeneous reporting formats and provides method-specific rankings, architecture analysis, and implementation recommendations.
- Analyzed 24 peer-reviewed studies from IEEE, arXiv, Springer, and other academic sources
- Time period: 2021-2025 (strictly post-2020 research)
- Focus: Most successful ensemble methods for currency exchange rate forecasting
- Six ensemble categories: Stacking, Boosting, Bagging, Hybrid Deep Learning, TCN, Transformer
- Key metrics: Directional Accuracy, Sharpe Ratio, RMSE, MAE, MAPE
- Best overall accuracy: 78.2% (Sadeghi et al., 2021) with multi-class SVM ensemble
- Emerging trend: Hybrid architectures (LSTM+Attention+Tree ensembles) outperform traditional methods
- Research significance: Evidence-based guide for selecting optimal ensemble configurations
Conclusion
This comprehensive analysis of 24 Forex ensemble studies from 2021-2025 reveals significant advances in ensemble methods for currency forecasting. Key findings include:
Major Trends:
- Performance Evolution: Directional accuracy improved from ~65-70% (2021) to 73-78% (2025)
- Dominant Methods: XGB/LightGBM ensembles and hybrid deep learning architectures lead performance
- Innovation Trajectory: From basic stacking (2021) → attention mechanisms (2024) → dynamic/transformer-based ensembles (2025)
Best Practices:
- For EUR/USD: AB-LSTM-GRU (RMSE 0.0012) or Multi-class SVM (78.2% accuracy)
- For multi-pair portfolios: XGBoost+LightGBM ensemble (76.1% accuracy)
- For HFT: TCN-based with dilated convolutions
- For emerging currencies: Ridge-regularized TCN+LSTM
Research Gaps:
- Reproducibility limited (avg 6.5/10)
- Few studies provide complete code/datasets
- Limited cross-currency generalization studies
- Insufficient real-time deployment benchmarks
Future Directions:
- Multi-modal fusion (price + sentiment + macro data)
- Adaptive/conditional ensembles based on market regime
- Lightweight deployment for mobile/edge devices
- Standardized evaluation frameworks