A large language model-driven framework for adaptive stock forecasting under China’s policy-driven market dynamics

Stock price forecasting in policy-sensitive markets like China’s A-shares presents a formidable challenge, as frequent regime shifts driven by external information can render traditional time-series models ineffective. While recent studies have started integrating textual data using Large Language Models (LLMs), they typically treat LLM-derived insights as static features, failing to adapt their core prediction logic when market dynamics fundamentally change. This paper introduces the Dynamic Priors-driven Adaptive Decomposition Network (DPAD-Net), a novel framework that pioneers a “perceive-and-adapt” mechanism. We reposition the LLM from a feature extractor to a macro-regime strategist that interprets real-time, multi-source information streams to generate dynamic priors—high-level judgments about the current market state. These cognitive priors act as command signals for a Regime-Adaptive Non-linear Decomposition Network (RAND-Net), which dynamically alters its mathematical structure to decompose stock prices into a stable value anchor and a volatile sentiment component. This adaptive decomposition allows the model to capture the complex, state-dependent interactions between fundamentals and market emotion. Extensive experiments on China’s A-share market demonstrate substantial improvements over state-of-the-art baselines: 13.6% reduction in MSE, 27.6% improvement in Sharpe ratio, and superior robustness during volatile market conditions. The framework’s “perceive-and-adapt” mechanism proves particularly effective during policy-driven rallies and sentiment-induced crashes, where traditional models anchored to historical patterns fail catastrophically.