Data and AI for Trading

Abstract

In this thesis my objectives were to explore how data science and artificial intelligence can be applied to make price predictions in the financial market. Methods include sentimental, technical, and fundamental analysis, by feeding these types of data into machine learning and deep learning models like random forest, gradient boosting machine, support vector machine, gated recurrent unit and long-short term memory networks. It should be noted that successful methods within this application of artificial intelligence is extremely secretive for obvious reasons and is populated with a lot of fabricated information which has not been tested, verified, and used to earn actual money on the real market. Therefore, delving into this master thesis was like jumping into a jungle, not knowing what to look for, which data to use, which methods to use, what “good” looks like within this application and how to formulate the problem. It should also be noted that while most projects cannot be planned from start to finish without any changes, this project particularly emphasizes experimenting and making decisions based on the outcomes Price data was gathered from various financial markets like bonds, commodities, forex, and indices. Fundamental data was collected from forexfactory.com and sentimental data was collected from The Global Database of Events, Language, and Tone (GDELT), a project that focuses on documenting and analyzing news into an archive. At first, I had no idea where to gather data, how to gather data nor how far back I could get data. A lot of time was spent on literature review, planning and programming to retrieve data. My initial findings surprised me. I found that at least with the methods and data that I have tried so far, it is not sufficient to predict the financial markets, not even good enough to surpass a monkey guessing if a price will go up or down, getting 50% accuracy. I do not believe that doing more of the same (for example gathering data on more assets) or using more advanced models will help, because I have over time tried the approach of more data and more complex models without any luck. However, roughly six weeks before the deadline of this thesis, I was able to make some progress.