Predicting stock prices presents significant challenges due to the inherent complexity and chaotic behavior of financial time series data. Traditional deep learning models often struggle to capture long-term dependencies and manage the non-stationary nature of stock prices. To enhance prediction accuracy while effectively addressing these challenges, this paper introduces a framework that employs graph-based feature extraction and Diffusion Variational Autoencoders (Diffusion-VAEs). By transforming stock data into complex graph networks and utilizing feature embeddings alongside a diffusion process, our approach effectively captures both the underlying structure and temporal dependencies, enriching the charac teristics of stock data. Experimental results on real-world stock datasets, spanning both regression and classification tasks, demonstrate that our method significantly outperforms recent state-of-the-art models by substantially lowering MAPE and RMSE while boosting the F1-score by over 20%, proving its accuracy and stability.