Temporal Graph Neural Networks for Stock Market Trend Prediction: A Comprehensive Benchmarking Analysis

The stock market has always been a subject of study, from large market makers to small investors or speculators. Small arbitrages and hints in the data can lead to great pro tability, but that is precisely the challenge: nding them. Numerous models, from the simplest statistical ones to the use of machine learning models, use historical data series to try to predict market trends. However, perhaps the most important aspect is often overlooked: the topology and relationships that each stock has with others, as each company is not a world unto itself but a dense network of relationships, events, and information exchange. This thesis studies the application of various models for predicting stock price trends, ranging from traditional RNNs to Transformers, focusing mainly on innovative temporal neural networks (Temporal Graph Neural Net- works) and integrating a benchmark for multivariate prediction of stock price trends. A data-driven graph con- struction strategy is proposed, motivated by other studies, which is based on information entropy and signal en- ergy, allowing the modeling of dynamic and asymmetric relationships between stocks without relying on static domain knowledge. Several experiments were conducted on di erent stock exchanges, such as NASDAQ, NYSE, and SSE, to analyze the robustness in di erent environments and with di erent scenarios. Comparative analysis tries to emphasize the strengths and limitations of the various models, which are underlying the importance of cor- rectly identifying and exploiting the topology of the stock market. The study also explores the reasons why many scienti c ndings appear to achieve remarkable results on paper, but then may fail to outperform the market in practice, analyzing the critical issues in market forecasting due to their short-term trends.