Forecasting intraday high-frequency gold prices presents a significant challenge due to the data's complex and non-linear dynamics. While deep learning models are widely applied, rigorous comparisons under standardized conditions remain limited. This study addresses this point by conducting a rigorous comparative analysis of five key deep learning architectures: one-dimensional convolutional neural networks (1D-CNN), simple recurrent neural networks (RNN), gated recurrent units (GRU), long short-term memory (LSTM), and bidirectional LSTM (BiLSTM). Utilizing a sliding window preprocessing technique, all models are systematically benchmarked under an identical training pipeline and evaluated across three distinct temporal window configurations to assess performance robustness. Experimental results demonstrated that the LSTM architecture consistently achieves superior forecasting accuracy, recording the lowest error metrics (MAE, MSE, RMSE) and highest R² score across all tested scenarios. These findings establish a rigorous performance benchmark, identifying LSTM as an effective and robust architecture for this high-frequency forecasting task.