This study proposes MST-Net (Multi-Scale Transformer Network), an advanced deep learning architecture that integrates Transformer mechanisms, a novel Price-based Sinusoidal Encoding (PSE), and multi-scale CNNs for stock price prediction. While Transformers capture long-range dependencies effectively, they exhibit limitations in recognizing local patterns across multiple time scales. MST-Net addresses this by incorporating PSE, which applies learnable sinusoidal transformations directly to price data to capture cyclical behaviors, along with multi-scale feature extraction. The model is evaluated on Vietnamese stock market data (2020–2025) across three stock categories: stable, volatile, and cyclical. Results demonstrate that MST-Net achieves RMSE of 1.65, MAE of 1.29, MAPE of 3.04%, and R2 of 0.60 for 1-day predictions, representing improvements of 26.0% in RMSE, 25.9% in MAE, and 21.6% in MAPE compared to the base Transformer, with R2 enhancement from 0.21 to 0.60. While Directional Accuracy (DA) shows modest results at 44.64%, MST-Net prioritizes magnitude over directional prediction. Ablation studies reveal that PSE and multi-scale CNNs independently reduce RMSE, with their combined effect significantly contributing to MST-Net’s superior performance across multi-horizon evaluations and stock categories.