Publication Date

Fall 2021

Degree Type

Master's Project

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

Mike Wu

Second Advisor

Thomas Austin

Third Advisor

Wendy Lee

Keywords

Stock markets, generative adversarial networks, long short term memory, deep learning.

Abstract

Trading equities can be very lucrative for some and a gamble for others. Professional traders and retail traders are constantly amassing information to be a step ahead of the market to profit off the value of stocks on the market. Some of the tools in their arsenal include different types of calculations based on a variety of data collected on a stock. Technical analysis is a technique for traders to analyze the data of equities presented on charts. Often, the way the price changes over time can be used as an indicator for traders to predict how future prices will move. This practice can be done due to the investing psychology of the masses that indicates a certain sentiment towards a stock. As artificial intelligence and machine learning have developed, researchers have also studied how to utilize this technology to analyze the data and forecast how prices will change.

Most recently, neural networks in deep learning approaches have been shown to outperform traditional machine learning methods. From the deep learning approaches, utilizing Long Short-term Memory (LSTM), a recurrent neural network (RNN) architecture, has been able to use time-series stock data to forecast future stock prices. This project proposes to extend this prediction process by layering another deep learning approach called Generative Adversarial Network (GAN) which pairs two networks to compete and improve each other . The approach was used to train an LSTM network to further improve the performance. The performance of the stacked LSTM-GAN model was compared to the stacked LSTM that had not been improved by the GAN. Results showed that the proposed model was able to outperform the no GAN extension by 22-25%, based on the RMSE and MAE, on the test set of the data used in training. Additionally, on a random set of stock data, the model performed 4-5% better.

Download

Share

COinS