GRU递归神经网络对股票收盘价的预测研究

股票市场是个多变且复杂的非线性动力学系统，股票价格是个具有时序性的数据，基于此选用具有时间记忆功能的GRU（Gated Recurrent Unit）递归神经网络模型来处理时间序列数据的预测问题。本文选取上证中18支证券行业股票的日收盘价数据，该数据截止日期为2017年12月29日，每支股票数据量为1000天。本文作了2个实证研究，一方面用GRU递归神经网络预测未来10天的股票日收盘价，实证结果表明，GRU递归神经网络的测试误差和验证误差都比其余2个模型得到的同种类型的误差要小，而GRU递归神经网络在预测未来10天日收盘价的精度达到了98.3%，体现了GRU强大的学习能力和泛化能力。另一方面，对比序列长度分别为240天、120天以及60天时，GRU递归神经网络的测试误差、预测收盘价的方差以及验证误差。结果表明面对不同序列长度的数据集，GRU预测精度都很高，序列长度为240天的GRU模型得到的测试结果的方差明显低于其他2个，说明其稳定性更好。

GRU Neural Network’s Prediction of Stock Closing Price

The stock market is a nonlinear dynamics system, which is changeable and complicated, and the price of stock is a kind of data with the character of time sequence. Given that, this thesis selected the model of Gated Recurrent Unit（GRU） with the function of time memory to deal with the problem of predicting time series data. The thesis selected daily closing price data of 18 securities industry stock in Shanghai, and the deadline of the data is December 29th, 2017. The data volume of the thesis is 1000 days per stock. To predict the closing price of the stock in the next 10 days, the thesis made empirical research. The testing and validation errors of GRU recurrent neural network is smaller than that of the other two models in the same type of error, while the accuracy of the prediction of closing price in next 10 day of Gated Recurrent Unit（GRU） has reached 98.3%, showed in the empirical results. This has embodied the strong learning ability and generalizing ability of Gated Recurrent Unit（GRU）. On the other hand, comparing the test error, variance in forecast closing price, and validation error of testing of Gated Recurrent Unit（GRU）on the sequence length of 240 days, 120 days, 60 days, and it suggests that the predicting accuracy are all in a high precision. However, the testing result on the sequence length of 240 days has a lower variance obviously, showing its better stability.