Stock data have a long memory, that is, changes in stock prices are closely related to historical transaction data. Also, Recurrent Neural Networks have good time series feature extraction capabilities. The paper proposed prediction models based on RNN/LSTM/GRU respectively. The attention mechanism has the ability to select and focus "key information”. Therefore, based on the conventional Recurrent Neural Network, this paper introduced the attention mechanism and proposed a prediction model based on AT-RNN/AT-LSTM/AT-GRU. And the paper modeled and experimented with it. The results showed that: (1) In the most basic comparison test of RNN-M, LSTM-M, and GRU-M prediction models, the GRU-M and LSTM -M was significantly better than the RNN-M and the GRU-M was slightly better than the LSTM-M; (2) The introduction of the attention mechanism layer was helpful to improve the accuracy of the stock fluctuation prediction model;(3) Deeper neural networks did not necessarily achieve better results.

股票数据具有很长的记忆力，即股票价格的变化与历史交易数据密切相关。而且，递归神经网络具有良好的时间序列特征提取功能。分别提出了基于RNN / LSTM / GRU的预测模型。注意机制具有选择和集中“关键信息”的能力，因此，在传统的递归神经网络的基础上，介绍了注意机制并提出了基于AT-RNN / AT-LSTM / AT-GRU的预测模型。结果表明：（1）在RNN-M，LSTM-M和GRU-M预测模型的最基本比较测试中，GRU-M和LSTM -M明显更好比RNN-M和GRU-M稍好于LSTM-M；