A Hybrid Machine Learning Framework for Systematic Trading in Cryptocurrency and FX Markets

This paper introduces a hybrid machine learning framework for algorithmic trading in highly volatile cryptocurrency and foreign exchange (FX) markets. The framework integrates high-frequency market data (OHLCV), order book microstructure signals, and sentiment analysis powered by Large Language Models (LLMs) to generate systematic trading strategies. The approach leverages a multi-model architecture: LightGBM for efficient feature learning, LSTM and Transformer neural networks for capturing temporal dependencies, and a regime-switching mechanism to adapt across different market conditions. Final trading signals are combined through a stacking ensemble method to maximize robustness and predictive accuracy. To ensure practical application, the framework incorporates walk-forward validation, realistic transaction cost and slippage modeling, and a risk management overlay based on ATR stops, volatility targeting, and maximum drawdown controls. Backtests using data from 2019–2024 show significant improvements in Sharpe ratio, drawdown reduction, and profit factor, outperforming individual models and benchmark strategies. Results highlight the importance of LLM-driven sentiment features, which provide measurable improvements in predictive power and trading performance. This work contributes to the growing literature on machine learning in finance, quantitative trading, and AI-driven systematic strategies, offering a scalable and adaptable solution for crypto trading, forex trading, and algorithmic portfolio management.