<p>Stock price forecasting is an important issue for research in the field of finance, which is of great significance for investment decision-making, risk management and policy formulation. However, the inherent complexity and multi-scale volatility of stock prices pose challenges for existing methods, which often fail to effectively capture both short-term fluctuations and long-term trends simultaneously. To overcome these limitations, we propose the Market-Scale Adaptive Decoding Network (MSAD-Net), a novel framework that integrates two specialized modules: the Trend Dynamics Catcher (TDC) and the Volatility Pattern Probe (VPP). The TDC leverages a hybrid architecture of Long Short Term Memory Network Recurrent Neural Network (LSTM) and Transformers to model long-term trends, while the VPP employs the synergistic capabilities of Convolutional Neural Networks (CNNs) and Transformers to accurately detect short-term fluctuations. To enhance the robustness of predictions, MSAD-Net incorporates an adaptive Multidimensional Loss Optimization Function (MLOF), which balances prediction errors across varying time scales. Experimental results on the Dow Jones Industrial Average (DJIA) dataset demonstrate that MSAD-Net outperforms baseline models, achieving significant improvements across multiple evaluation metrics. Specifically, MSAD-Net shows superior performance on the majority of the 30 component datasets, achieving statistically significant improvements including an MSE reduction of up to 0.2779, MAE decrease reaching 0.2469, MAPE error reduction of 10.79%, and <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {R}^{2}\)</EquationSource> </InlineEquation> score improvement of 0.2991 compared to baseline approaches. These comprehensive metric enhancements substantiate MSAD-Net’s enhanced capability in modeling stock price dynamics and volatility patterns.</p>