Financial News Event Prediction Using Temporal Knowledge Graphs

Abstract

Predicting future financial events and trends is crucial for effective decision-making in financial markets. Traditional approaches utilizing Temporal Knowledge Graphs (TKGs) primarily rely on specific relationship forecasting within knowledge graphs, often falling short in capturing the broader, more dynamic shifts that characterize industry trends. This thesis explores the feasibility of using TKGs in concert with advanced machine learning to predict such trends, shifting the focus from specific relationship predictions to identifying broader patterns in entity categories and relationship types. We investigated multiple techniques, including adapting existing graph convolutional network models and implementing custom model architectures leveraging Transformers and Multi-Head Latent Attention. Performance was evaluated across several publicly available TKG datasets. Our findings suggest that while defining a clear and actionable trend within a TKG presents a significant challenge, a trend-based approach using carefully curated datasets and appropriate machinelearning architectures has the potential to improve prediction accuracy, particularly when combined with methods addressing model generalization and data scarcity. Future research directions include exploring more sophisticated trend definitions, incorporating entity-specific category information, and further optimizing model architectures for enhanced performance and reduced computational costs.