Social media has become a significant force shaping financial markets, particularly in speculative assets such as meme coins and meme stocks where community-driven narratives often substitute for traditional fundamentals. Understanding who influences these markets and how remains a largely unsolved problem. This doctoral research addresses this challenge through two interconnected projects. The first project investigates what factors contribute to user influence in financial social networks, employing heterogeneous graph neural networks (GNNs) and multimodal content analysis, encompassing both textual posts and visual content such as memes and infographics, to characterize influence under varying market conditions. The second project builds upon these insights to tackle the problem of social-aware market event attribution: given an anomalous market event, how can we automatically trace it back to its social network origins and explain why it happened? We propose a novel framework that combines GNN-based network tracing with LLM-based content attribution through an iterative collaboration mechanism. Together, these two projects form a coherent research agenda advancing our understanding of social influence in financial markets, with implications for market surveillance and platform governance.

Social network platforms like Reddit are increasingly impacting real-world economics. Meme stocks are a recent phenomena where price movements are driven by retail investors organizing themselves via social networks. To study the impact of social networks on meme stocks, the first step is to analyze these networks. Going forward, predicting meme stocks' returns would require to predict dynamic interactions first. This is different from conventional link prediction, frequently applied in e.g. recommendation systems. For this task, it is essential to predict more complex interaction dynamics, such as the exact timing. These are crucial for linking the network to meme stock price movements. Dynamic graph embedding (DGE) has recently emerged as a promising approach for modeling dynamic graph-structured data. However, current negative sampling strategies, an important component of DGE, are designed for conventional dynamic link prediction and do not capture the specific patterns present in meme stock-related social networks. This limits the training and evaluation of DGE models in such social networks. To overcome this drawback, we propose novel negative sampling strategies based on the analysis of real meme stock-related social networks and financial knowledge. Our experiments show that the proposed negative sampling strategies can better evaluate and train DGE models targeted at meme stock-related social networks compared to existing baselines.