Class Year

2019

Document Type

Thesis

Honors Designation

High Honors

Department

Computer Science

Primary Advisor

Aaron Gember-Jacobson

Second Advisor (if necessary)

Darren Strash

Abstract

Algorithmic Trading (AT) is a financial sector that trades financial instruments, such as stocks, with algorithms and no human interaction. This allows the largest prop-trading firms in the world to conduct thousands of trades per second. In practice, the vast majority of strategies are implemented using mathematical formulas based on a variety of stock metrics, such as closing price or volume. However, the effectiveness of these algorithms isn’t publicly available due to the necessity of secrecy of implementation details. Part of the thesis aims to uncover and examine the effectiveness of existing metrics-based strategies. We find both pairs trading and a combined RSI and MACD momentum algorithm to be incredibly effective.

Moving beyond traditional AT strategies, this thesis further aims to investigate using news- or media-content-based strategies. By using Twitter data and Natural Language Processing (NLP), we create a unique trading strategy based on Twitter sentiment of publicly traded companies’ tweets using a bevy of machine learning algorithms and a deep learning algorithm. We use models which have simplistic or extended features and apply these features to a stacking model and a deep learning model. We find the simplistic model to be very ineffective while the extended model beats the baseline measure for 87.5% of stocks tested, generating profits of up to 529%. Our stacked model beats the baseline for 62.5% of stocks tested and proves to be very effective for specific stocks. Our deep learning model is far more risk averse than any of the other models explored.

Share

COinS