Abstract: Trading robots, meticulously crafted programs, are designed to execute trades automatically. However, stock trading presents a unique challenge. Unlike finite game tasks, stock markets operate perpetually, making it arduous for traders to design appropriate reward functions for training Reinforcement Learning models. For stock trading tasks that can easily determine the optimal decision trajectory from historical data, previous studies showed that Behavioral Cloning has much better learning efficiency than Reinforcement Learning. In this study, we propose a novel Behavior Cloning algorithm that leverages Long Short-Term Memory (LSTM) networks and self-attention mechanism as core components. Our approach effectively captures temporal dependencies and interrelations among elements at various positions, aiming to enhance learning efficiency. Additionally, a strategic approach known as the positive transaction expert strategy was devised to guide the model training process. In our comparative analysis, we evaluated the proposed algorithm against supervised learning, reinforcement learning, and traditional time series trading algorithms. The empirical results indicate that the Attention-Based Behavioral Cloning algorithm exhibits an 83.33% likelihood of achieving the highest return.