A survey on the application and research progress of large language models in financial forecasting

Large language models (LLMs) are reshaping the technical paradigms of financial forecasting through their robust representation learning and reasoning capabilities. This paper systematically reviews the application pathways of architectures such as transformers and graph neural networks in scenarios like stock prediction and risk management, highlighting key technologies for enhancing prediction accuracy through knowledge injection and temporal modeling improvements. The study reveals that LLMs demonstrate significant advantages in unstructured data processing and cross-market correlation analysis but face challenges related to economic logic interpretability and data non-stationarity. Future research should focus on advancing causal reasoning augmentation and federated learning collaboration to achieve secure and trustworthy evolution of financial forecasting systems.