The Limit Order Book (LOB) serves as a real-time record of buy and sell orders for a specific asset, providing valuable insights into market demand and supply dynamics. Leveraging the information within the LOB, this study introduces a novel hybrid deep neural network framework that integrates convolutional neural networks (CNNs) with transformers to forecast future price changes in high-frequency trading data. The CNNs excel in extracting spatial features from LOB data, while transformers capture long-range dependencies, enabling the model to identify complex patterns across different price levels and time sequences. To enhance predictive accuracy, the framework includes a carefully designed convolutional kernel size tailored to the specific structure of LOB data, reducing the need for manual feature engineering. Thorough evaluations were conducted on the FI-2010 dataset, which comprises LOB data for five instruments from the Nasdaq Nordic stock market over a ten-day period. The results demonstrate that our model well outperforms traditional methods, achieving higher precision, recall, and F1-scores. This innovative approach not only advances the predictive accuracy in financial market forecasting but also offers potential applications in developing real-time trading strategies, improving market stability, and assisting in regulatory surveillance.