Novel Trading Algorithms augmented by Intrinsic Time and Machine Learning

The Directional Change (DC) Framework is an innovative event-driven approach designed to characterize price movements in micro financial markets using uneven time intervals as a system clock. This research investigates novel techniques for market price prediction and trading strategies based on the DC framework. We developed two distinct DC-based strategies: trend-following and counter-trend strategies. Empirical evidence highlights that the trend-following strategy offers higher potential profitability, albeit with relatively higher risk compared to the counter-trend strategy. To enhance the profitability of DC-based strategies, we integrated machine learning (ML) algorithms to predict the type and length of overshoot (OS) events. We employed five prominent ML models—Random Forest, Support Vector Machine (SVM), Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), and Convolutional Neural Networks (CNN). Additionally, we incorporated a reinforcement learning (RL) framework to further improve the performance of the DC-based trading strategies. Our DCRL strategy, combined with a deep neural network policy network (ResNet), demonstrated exceptional ability in selecting optimal strategies under natural market states. The study utilized minute-level data from the foreign exchange market, focusing on eight major currency pairs over a 15-year period from 2006 to 2020. The performance of the prediction models was evaluated using metrics such as Rate of Return (ROR), Maximum Drawdown (MDD), and Sharpe Ratio. However, the study has certain limitations: the dataset is confined to the foreign exchange market, which may affect the representativeness and generalizability of the findings. The time span may not be sufficient to support long-term trend analysis or the impact of specific events. Additionally, large-scale data analysis and complex models require high computational power, potentially posing technical and resource constraints. Future research directions could include developing specialized financial data network models, incorporating multi-agent reinforcement learning, expanding datasets and asset classes, and further exploring the DC framework. This study advances the integration of machine learning algorithms within the DC framework to improve trading strategy profitability and introduces multi-core processing to significantly enhance the learning efficiency of RL models, a substantial advancement that can greatly accelerate future research in this domain.