The paper introduces S3G (Stock State Space Graph), a novel framework for stock trend prediction that explicitly models the evolving interdependencies between stocks over time. Unlike prior methods that rely on static industry graphs or per-slice similarity measures, S3G employs a state space formulation over temporal graphs to capture both instantaneous and delayed interactions. The framework consists of two main components: a Wavelet Denoising Net (WDN) that uses discrete wavelet decomposition with learnable soft thresholding to denoise financial series and extract salient patterns, and a State Space Graph Learning (SSGL) module that constructs time-specific graphs via Gaussian kernel similarity on patch embeddings and models their evolution through parametric state space equations. The predicted graph structure is then aggregated using a graph neural network to produce cross-sectional return forecasts. The model is trained end-to-end with a dual objective combining MSE and ranking losses.

Extensive experiments on the Chinese CSI 500 index from 2015 to 2024 demonstrate that S3G achieves state-of-the-art performance across multiple metrics, including Information Coefficient (IC), RankIC, annualized return, and Sharpe ratio, outperforming 15 baselines ranging from traditional machine learning (XGBoost) to deep learning methods (LSTM, Transformer, Mamba) and specialized financial models (FactorVAE, MambaStock). Ablation studies confirm the critical importance of both the denoising and graph evolution components, as removing either leads to significant performance degradation. The results highlight the value of capturing fine-grained, time-varying stock relationships for enhanced predictive accuracy in financial markets.