Development of a Trading analysis and operation System for American stocks using Deep Reinforcement Learning

Víctor Clemente Teruel. (2020). Development of a Trading analysis and operation System for American stocks using Deep Reinforcement Learning. Trabajo Fin de Titulación (TFM). Universidad Politécnica de Madrid, ETSI Telecomunicación.

Trading is an activity in which specialized people have the ability to create profit from their actions. They are people that study a lot to reach a point in which they can predict or suspect what is going to happen at a specific point in the stock market. Most of the people think that take profit from stock markets is trivial, but the reality is that 98% of people who try to reach the same point lost all their money. The biggest problem that traders have is that they need to spend hours and hours in front of a computer to search which companies are going to have a big tendency to change, in a short or in a long period of time, depending on which type of trading they operate, and then join this position before the tendency change appears. To find these opportunities, they can analyze the fundamental characteristics (cash flow, balance sheets...) and/or the technical chart (tendency direction, statistics...) of a company. Nowadays, there is a growth in machine learning developments and utilities, because of the improvement of this technology. Most of the machine learning applications are implemented to make our day-to-day life easier or to automatize some activities that computers can prepare better than humans. The main objective of this project is to create a program that takes some of the knowledge that a specialized person can learn from courses, books or speeches, and train this program with different stocks datasets, to finally operate correctly in the American stocks market, all of it using deep learning and reinforcement learning. In addition, an analytic tool will be created to detect which tendency is the different stocks in a specific moment, comparing each other and with the main US index (S&P 500), to extract the datasets that will be provided to the program. This project is going to include a state of the art of each technology implemented, a description of which trading techniques are learned by the program, an explanation of which are the main premises to obtain train datasets, and the final results of how this intelligence operates in different market situations (bullish, bear or lateral trends