An Adaptive Box-Normalization Stock Index Trading Strategy Based on Reinforcement Learning.

Financial time series prediction and stock trading strategy have always been the focus of research due to the generous returns. Stock box theory is a classic investment strategy, which has been studied by investors and scholars for many years. In this paper, we propose an adaptive box-normalization (ABN) stock trading strategy based on reinforcement learning (RL), which improves the original box theory. In our ABN strategy, the stock market data is independently normalized inside each oscillation box. Given the data of each box, support vector regression (SVR) is applied to predict the maximum rise range and maximum fall range within a certain period in the future. Meanwhile, the genetic algorithm (GA) is employed to optimize the input features of SVR via the mean square error (MSE) of prediction. We construct the trading strategies by Q-learning for the trading of single-stock and two-stock portfolio. Finally, the trigger threshold of oscillation box is dynamically adjusted according to the volatility of the stock price. Extensive experiments support that our proposed strategy performs well on different stock indices and achieves promising results.