Stock price prediction in a multi-stock environment is challenging due to market volatility, non-stationarity, and complex nonlinear dependencies. Additionally, stock movements influence each other through subtle, time-varying relationships. While deep learning models, particularly convolutional networks and Transformers, have improved the ability to capture temporal patterns, many still treat stocks independently or use multi-level attention structures that add computational cost without fully leveraging cross-stock dependencies. To address this, we propose StockCT, A Transformer-Convolution Hybrid Framework for Multi-Stock Prediction in the China Stock Market. StockCT integrates a patch-based module for noise suppression and inter-stock attention within the Transformer to capture dynamic cross-stock dependencies. A lightweight prediction head outputs normalized returns, achieving both higher efficiency and improved interpretability. Extensive evaluations on the CSI300 and CSI800 datasets show that StockCT outperforms strong baselines, achieving 79.6% and 82% improvements in Information Coefficient (IC) and 89.7% and 696% gains in Information Ratio (IR), respectively. These results demonstrate the effectiveness of combining local and global dependencies for multi-stock forecasting.