Abstract: Stock price volatility is a highly complex nonlinear system. The adjustment of stocks is not based on a uniform time process and has its own process of advancement. Combining the characteristics of LSTM（Long Short-Term Memory） recurrent neural network and the characteristics of stock market, and after preprocessing operations such as interpolation, wavelet noise reduction, and normalization of data, all of this data will be inputted into the LSTM network model of different LSTM layers and the number of different hidden neurons in the same layer for training and testing. Comparing the evaluation indicators with the prediction results, it finds the appropriate number of LSTM layers and hidden neurons, and improves the prediction accuracy by about 30%. The test results show that the computational complexity of this model is small and the prediction accuracy is improved. It not only provides a useful reference for predicting stock trend before stock investment, but also helps investors to build a suitable stock investment strategy after further understanding of the actual stock price.

Key words: wavelet noise reduction, number of Long Short-Term Memory（LSTM） layer, hidden neurons, stock price forecast

摘要： 股价波动是一个高度复杂的非线性系统，其股票的调整不是按照均匀的时间过程推进，具有自身的推进过程。结合LSTM（Long Short-Term Memory）递归神经网络的特性和股票市场的特点，对数据进行插值、小波降噪、归一化等预处理操作后，推送到搭建的不同LSTM层数与相同层数下不同隐藏神经元个数的LSTM网络模型中进行训练与测试。对比评价指标与预测效果找到适宜的LSTM层数与隐藏神经元个数，提高了预测准确率约30%。测试结果表明，该模型计算复杂度小，预测准确率有所提高，不仅能在股票投资前对预测股票走势提供有益的参考，还能帮助投资者在对实际股价有了进一步的认知后构建合适的股票投资策略。

关键词: 小波降噪, 长短期记忆网络（LSTM）层数, 隐藏神经元, 股价预测