| dc.description.abstract | The digital technology age has created an unprecedented production and consumption of data. With this surge, data-analytics have become increasingly important in industries that can generate value based on predictive models: online retail, meteorology, transportation, social media, etc. This paper will focus on the finance industry because it contains years of available time-series data, a high correlation between prediction accuracy and generated value, and a high ceiling for creating a perfect prediction. Historically, time-series forecasting has been used to predict how different companies’ stock prices might behave in the future given how they have behaved in the past. However, focusing solely on sequential data for prediction is missing the potential of interdependency within the information. Companies do not exist in a vacuum; they exist in markets that can heavily influence multiple entities at once: government regulation, supply and demand, etc. By grouping these entities and examining their relationships in the form of graph based networks, interdependent data can be used in conjunction with sequential data to improve prediction accuracy. In this thesis, Long Short-Term Memory (LSTM) deep neural networks will act as the foundation for which Graph Convolutional Neural networks (GCN) will be built atop each other to combine both sequential and relational embeddings in formulating stock price prediction. To verify this hypothesis, closing stock price data from over a five year period was used. | |
