Inspiration
The inspiration for our project originated from Pavel's groundbreaking research on Graph Neural Networks (GNNs) for solving Partial Differential Equations (PDEs). Intrigued by this innovative approach, we decided to apply a similar analogy to financial time series, specifically the S&P 500 stocks. Our goal was to utilize the interrelationships among companies as a foundation to train our model, aiming to predict future stock prices with a higher degree of accuracy.
What It Does
Our model evaluates the intricate connections between different companies, predicting stock prices based on these relationships and other critical financial indicators such as open, close, high, and low prices. By leveraging the dynamic interplay between these variables, our approach offers a nuanced method for forecasting stock market trends.
How We Built It
The construction of our model involved an extensive process of data acquisition and organization. We assembled datasets to create a comprehensive graph structure, where each node represents a company. These nodes are linked based on various criteria, including partnerships, competition, and supplier relationships. Our work primarily utilized PyTorch Geometric, a powerful library designed for processing and training graph-structured data, to facilitate our model's development.
Challenges We Ran Into
Our journey was marked by several challenges, notably in data sourcing and structuring. The task of finding datasets that accurately represent the complex relationships among S&P 500 companies was particularly daunting. Adapting this data to fit the PyTorch Geometric training interface required significant effort, involving meticulous structuring and consistent data description. Additionally, selecting and designing an effective model posed a substantial challenge.
Accomplishments That We're Proud Of
Despite encountering various hurdles, we successfully established a running training loop for our model. However, as the training progressed, we faced issues with numerical stability, leading to NaN values in later iterations. Overcoming these early obstacles and reaching a functional stage of model training stands as a testament to our team's resilience and adaptability.
What We Learned
Throughout this project, we deepened our understanding of Graph Neural Networks and the complexities involved in sourcing, processing, and preparing datasets for such advanced algorithms. The experience also highlighted the value of teamwork and collaboration in navigating the challenges of a large-scale project.
What's Next for Predicting S&P 500 Stocks Using Graph Neural Networks
Looking ahead, our primary focus is on refining the model to enhance its reliability and predictive accuracy. This involves addressing the issues related to numerical stability, optimizing the model architecture, and exploring advanced data preprocessing and training techniques. Our ultimate aim is to develop a robust model capable of accurately predicting stock prices, thereby contributing valuable insights to the field of financial analysis through the innovative application of graph neural networks.
Log in or sign up for Devpost to join the conversation.