Developing a High-Frequency Trading system with Dynamic Portfolio Management using Reinforcement Learning in Iran Stock Market

Presence of the considerable gap between the time of receiving the buy/sell signals and the beginning of the price change trend provides an appropriate situation for implementation of algorithmic trading systems. Tehran stock exchange is one of these markets. A high-frequency trading system has some advantages (exploiting intraday stock market volatility) and disadvantages (high amounts of transaction cost due to the high transaction volume), thus we can augment the advantages and control the disadvantages by designing the system elaborately and modifying the trading regulations.
In this research, the “Local Traders” approach has been utilized to predict the future trend of stock and Reinforcement Learning has been used for dynamic portfolio management. According to the “Local Traders” approach, there is a local trader (an agent) for each stock that is expert at it. It predicts the future trend of its own stock based on stock’s intraday data and their technical indicators by determining how beneficial it is to buy, sell or hold. In this research, 2 models will be proposed based on Local Traders. Based on the first one, trades with fixed lot size were sought by exploiting the local traders’ recommendations. In the second model which is an extension of first model, one can dynamically manages the portfolio using reinforcement learning and local traders’ recommendations.
Results showed that, the proposed models outperformed the Buy and Hold strategy in Normal and Descending markets. Furthermore, in all kinds of markets, the second model outperformed the first one.
Generally, the Buy and Hold strategy works the best in an Ascending market, hence the proposed algorithms are not expected to outperform this strategy. However, the performance of the proposed approach along with Neural Network method to anticipate the future trend of stocks was considerable in Normal and Descending markets. In addition, the implementation of Reinforcement Learning model to dynamically manage the portfolio has improved the results.