GraphMixer in Financial Networks: A Novel Approach to Lead/Lag Pattern Detection

This thesis aims to represent lead/lag pattern detection in stocks and commodities from an innovative perspective by modeling the dynamic changes of the phenomenon as a link prediction task in a temporal graph. In this graph, the nodes represent the assets (stocks and commodities), while the temporal links signify the presence of a lead/lag relationship between the respective nodes at specific times. Understanding price movements in the market is crucial for optimizing trading strategies and managing risk. Moreover, analyzing these movements from the prism of lead/lag patterns refines the approach and broadens its impact by identifying price dependencies. For the purposes of this thesis, a novel dataset was compiled, encompassing 37 assets along with corresponding attributes for the nodes and links, including pricing data, financial indicators, and sentiment data, thereby enhancing the contribution of this work even further. To facilitate temporal link prediction for lead/lag pattern detection, six distinct models were adapted for our use-case. GraphMixer, a relatively simple architecture comprising solely Multilayer Perceptrons (MLP) mixers, was compared across various experimental configurations and feature subsets for the nodes and links of the lead/lag network. This comparison included more complex architectural solutions, such as RNN-based methods like JODIE, self-attention mechanism-based models like DySAT, TGAT, and APAN, as well as hybrid models such as TGN. The results show that GraphMixer outperformed baseline models, achieving state-of-the-art outcomes in detecting lead/lag patterns, which is a task that is underexplored in the current literature. Overall, the findings indicate that this thesis presents a novel approach to lead/lag pattern detection, with results sufficiently robust for real-world financial applications.