Abstract: Driven by the evolution of Large Language Models (LLMs), financial forecasting is shifting from statistical arbitrage to semantic intelligence. To address the limitations of traditional paradigms in processing unstructured data and modeling complex non-Euclidean spatiotemporal dependencies, this study proposes a novel hybrid architecture integrating LLMs with Graph Spatiotemporal Networks. First, we construct a domain-specific financial LLM utilizing "Market-implied Return" as a weak supervision signal to extract precise asset pricing factors. Second, we introduce a GAT-TCN architecture that couples Graph Attention Networks with Temporal Convolutional Networks to capture high-frequency non-stationarity and spatial spillover effects. Furthermore, to ensure interpretability and compliance, we incorporate a "Summarize-Explain-Predict" (SEP) neuro-symbolic framework aligned via Proximal Policy Optimization (PPO). Empirical evidence from China's SSE 50 index over a thirteen-year period demonstrates that this architecture achieves superior goodness-of-fit in volatility prediction and robustness against tail risks compared to traditional benchmarks, providing a viable technical path for the industrial deployment of LLMs in quantitative investment.

Keywords: Large Language Models (LLMs), Graph Spatiotemporal Networks, Quantitative Investment, Market-implied Sentiment, Neuro-Symbolic Systems, Proximal Policy Optimization (PPO).

Pages: 263-271