Deep reinforcement learning for trading strategy development on high-frequency currency data using directional changes sampling

High-frequency trading in the foreign exchange market presents unique challenges, requiring sophisticated techniques to address its complexities. This thesis investigates the application of deep reinforcement learning to algorithmically trade high-frequency currency data. Traditional trading algorithms typically use fixed interval sampling in both manual and automated trading strategies. While effective for less noisy price movements and longer trade durations, this approach can miss significant price shifts in high-frequency scenarios which limits profitability. To overcome these limitations, this thesis employs the directional change (DC) sampling paradigm, which captures significant price movements more effectively. By combining DC sampled data and deep reinforcement learning to train trading agents, the research explores whether this approach outperforms traditional fixed interval methods when trading at high frequencies. The initial investigation develops the Filtered Deep Reinforcement Learning (FDRL) trading framework, using deep reinforcement learning to create semi-autonomous trading agents. Results show that FDRL is effective at fixed transaction costs but requires rule-based interventions to manage trades. To enhance autonomy, the Positionally Aware Deep Reinforcement Learning (PADRL) framework is introduced, incorporating real-time positional awareness to eliminate the need for rule-based filters, further improving performance. The final contribution of the Spread Aware Deep Reinforcement Learning (SADRL) framework, refines the PADRL approach by using the bid-ask spread as opposed to fixed transaction costs, making the strategy more realistic and applicable to real trading environments. Each algorithm iteration demonstrates improved performance over traditional benchmarks like buy-and-hold and technical analysis. Financial metrics including Total Return, Maximum Drawdown and Calmar Ratio validate the superior performance of these deep reinforcement learning-based strategies, demonstrating their potential for advancing high-frequency trading in the foreign exchange market.