A Graph-Based Approach for Price Prediction and Portfolio Management

In quantitative trading, investors seek to maximize profits while minimizing risk. To forecast future prices and assess risks, it is crucial to employ price prediction models such as RNN, GRU, and even combinations of these with convolutional networks. However, the high computational cost of running these models on the entire stock market data has led some researchers to focus on specific market segments or random selections of stocks. Even when computational cost is not a concern, time remains a critical factor in financial markets. In this study, we aim to achieve optimal returns by selecting stocks and predictive models with the lowest time complexity. We construct a graph of the stock market to identify key nodes, train deep learning models on historical price data, and predict future prices. Based on these predictions, we then form a portfolio using the Markowitz model. The performance of this constructed portfolio is compared against the benchmark, S&P 500, and demonstrates superior results.