Date of Award

Summer 8-28-2023

Document Type

Thesis (Undergraduate)

Department

Computer Science

First Advisor

Dr. Soroush Vosoughi

Abstract

Artificial intelligence has powerful applications in virtually every field, and the financial world is no exception. Utilizing various elements of artificial intelligence, this research aims to predict the future value of the S&P 500 index using numerous models, and in doing so, identify relevant features. More specifically, models that include combinations of historical data, public sentiment, and technical indicators were employed to predict the stock price one day and three days forward. To account for public opinion, the sentiment of tweets and news headlines from the beginning of 2015 through the end of 2019 was calculated using FinBERT, a pre-trained version of BERT retrieved from the HuggingFace Model Hub and designed specifically for financial-related text. For each textual input, FinBERT provides three outputs: the probability that the text is positive, negative or neutral. These probability values were applied to approximate the number of positive, negative, and neutral tweets and news headlines each day. The following features were used in complex LSTM models: open, close, low and high prices; volume; the number of positive, negative, and neutral tweets and news headlines; relative strength index; and earnings per share. The highest performing predictive model for one day forward and three days forward utilized historical data, tweet sentiment, and the relative strength index. Coupled with other tools wielded by investors, this model can help anticipate market movements and inform decisions.

Download

Share

COinS