Automated Stock Trading by Reinforcement Learning

This work proposes a new application of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm to automated trading in stock markets across a variety of markets. We further enhance the effectiveness of the algorithm by innovating three aspects: the use of a reward clipping technique to penalize persistent suboptimal behavior, the incorporation of more than 50 technical indicators in state augmentation, and use of a turbulence index to measure dynamic market volatility. We demonstrate the proposed model against varied datasets covering cryptocurrency markets (Bitcoin and Ethereum) and traditional equity markets (NASDAQ, Dow Jones, S&P 500, and Chinese Shanghai Composite). The experimental findings showcase the model's capacity to yield high returns with up to $7,396,892 in profit with an initial investment of $100,000 on the NASDAQ dataset, with directional accuracy of up to 70%. Comparative assessments show that the proposed method consistently surpasses benchmark methods in profitability with varied risk-adjusted returns in all market conditions. These results emphasize how TD3-based reinforcement learning systems are capable of providing effective, adaptive financial trading system implementations in reality.