The formidable challenge of accurate stock price forecasting stems from three intrinsic characteristics of financial markets: persistent noise contamination, inherent non-stationarity in temporal patterns, and complex latent interdependencies among market variables. Current deep learning and hybrid approaches frequently overlook the implicit relationships between market entities, thereby limiting their prediction accuracy. Although Graph Neural Networks (GNNs) have emerged to leverage hidden market relationships for improved forecasting, they often struggle to effectively capture global non-linear interactions, handle asymmetric relationships, and resist market noise. To address these issues, we introduce a novel Spatial-temporal Graph Auto-Encoder (SPGAE), integrating a newly designed graph convolution operator that dynamically assigns trainable weights to neighboring nodes, enhancing local feature representations. Additionally, SPGAE employs a Graph Auto-Encoder (GAE) to capture global latent variables, thus improving noise resistance and facilitating a deeper understanding of macroeconomic market states. Furthermore, unlike most previous research, our proposed end-to-end architecture simultaneously facilitates optimized portfolio construction alongside stock price prediction, thereby substantially enhancing investment decision-making efficiency. Experimental evaluations indicate that SPGAE significantly surpasses mainstream models in prediction accuracy and stability.