Designing algorithmic trading strategy based on deep reinforcement learningCase study: Tehran Stock Exchange

Subject Areas : Financial Knowledge of Securities Analysis

saeed kazemian hoseinabadi ¹ , سید محمد رضا داودی ² , mohammad mashhadizadeh ³ , parsa jozi ⁴

1 - MA, Risk Management Financial Engineering, Dehaghan Branch, Islamic Azad University, Dehaghan,
2 - هیات علمی مدیر گروه
3 - Assistant Professor. Department of Management , Mobarakeh Branch, Islamic Azad University, Mobarakeh , Iran
4 - MA.RiskManagementFinancialEngineering,Dehaghan Branch,Islamic Azad University.Dehaghan ,Iran

Received: 2023-01-24 Accepted : 2024-06-18 Published : 2024-06-29

Keywords: neural networks, reinforcement learning, deep reinforcement learning, technical oscillator.,

Abstract :

Today, algorithmic trading is widely used in trading management. Algorithmic portfolio management is a new type of these system through which the portfolio manager helps to increase the quality of profit and reduce the risks of his portfolio using algorithmic tools. The purpose of this research is to design an algorithmic trading system based on deep reinforcement learning with the help of a neural network. In this approach, the agent or trader searches the search space to find more rewards, which is the same as more returns. The trader is faced with technical signals including relative strength index, stochastic oscillator, convergence-divergence indicator, and minimum, maximum, closing, and opening prices. Deep reinforcement learning replaces the Q value or quality function table with a neural network. Finally, upon receiving the state word, the mentioned neural network suggests one of the three actions of selling, buying, and holding. This proposal is in the form of three possibilities with a total of one, and the proposal with the maximum probability is implemented. The result of the implementation of the deep reinforcement learning trading system on the total index of Tehran Stock Exchange in the period of 2011 to 2014 shows that the research system was significantly different from the other three systems in the mean and convergence-divergence index. Also, the Sharpe ratio of the research system compared to the other three models showed growth of at least 1.4 times.