Multi-Modal Temporal Dynamic Graph Construction for Stock Rank Prediction

Stock rank prediction is an important and challenging task. Recently, graph-based prediction methods have emerged as a valuable approach for capturing the complex relationships between stocks. Existing works mainly construct static undirected relational graphs, leading to two main drawbacks: (1) overlooking the bidirectional asymmetric effects of stock data, i.e., financial messages affect each other differently when they occur at different nodes of the graph; and (2) failing to capture the dynamic relationships of stocks over time. In this paper, we propose a Multi-modal Temporal Dynamic Graph method (MTDGraph). MTDGraph comprehensively considers the bidirectional relationships from multi-modal stock data (price and texts) and models the time-varying relationships. In particular, we generate the textual relationship strength from the topic sensitivity and the text topic embeddings. Then, we inject a causality factor via the transfer entropy between the interrelated stock historical sequential embeddings as the historical relationship strength. Afterwards, we apply both the textual and historical relationship strengths to guide the multi-modal information propagation in the graph. The framework of the MTDGraph method consists of the stock-level sequential embedding layer, the inter-stock relation embedding layer based on temporal dynamic graph construction and the multi-model information fusion layer. Finally, the MTDGraph optimizes the point-wise regression loss and the ranking-aware loss to obtain the appropriate stock rank list. We empirically validate MTDGraph in the publicly available dataset, CMUN-US and compare it with state-of-the-art baselines. The proposed MTDGraph method outperforms the baseline methods in both accuracy and investment revenues.