Using Machine Learning Techniques for Trading Agent Modeling

In the current thesis project the field of algorithmic trading is considered as an application area of machine learning algorithms. The goal of the project is creating a machine learning based model of a trading agent – a trading robot, which is able to gain profit by executing operations of buying and selling a financial instrument on a stock market. The following problems were being solved: study of the subject area and existing ways of applying machine learning algorithms to it; proposal of an algorithm with considering the revealed limitations of existing solutions; the algorithm implementation and testing on a real historical market data. During the study of papers devoted to trading agent modeling with machine learning techniques it was figured out that the main arising difficulty is overfitting. This problem was taken into account and an algorithm was propounded, which solves the task as a reinforcement learning problem and is applying genetic algorithm to building decision trees and then uniting them into ensembles, what is a method of reducing overfitting. Genetic algorithms are used to maximize the robot's trading performance on a training dataset. Input of decision trees are technical indicators calculated from the candlestick-formatted market data. Through performing experiments on real market data of different financial instruments it was demonstrated that the algorithm is able to successfully solve the task of profitable trading. Nearly in all experiments the robot showed much better results than benchmark “buy-and-hold” strategy and, moreover, it gained positive value of money. Also the efficiency of ensembling the decision trees method in reducing overfitting is shown.