<p>A novel framework, the market-contextual adaptive wavelet transformer (MAWT), is proposed for stock price prediction to address the challenges of nonstationarity, noise, and data sparsity inherent in financial time series. Within this framework, to incorporate a broader market context, stock index and individual stock data are integrated into the learning process, allowing market-guided representations to be learned. An adaptive lifting wavelet transform (ALWT) module is proposed, which decomposes input sequences into smoother low-frequency components by applying learnable convolutional kernels within a lifting scheme, thereby enabling effective noise suppression and robust feature extraction. In addition, a multi-scale temporal attention (MSTA) module is constructed to enable the modeling of temporal dependencies at multiple scales. Extensive experiments are conducted on real-world Chinese stock market data, and consistent outperformance is observed over state-of-the-art baselines with respect to predictive accuracy, robustness, and investment returns. These results demonstrate that integrating data smoothing, multi-scale temporal modeling, and market-aware context contributes significantly to improved financial forecasting performance.</p>