深度稀疏修正神经网络在股票预测中的应用

Application of Deep Sparse Modified Neural Network in Stock Forecasting

传统的股票预测基于统计学的方法,虽然在股票的趋势预测中有效,但是价格预测的准确度不够令人满意。因此,文中利用神经网络研究股票价格的预测问题。神经网络具有很强的非线性拟合能力,适用于股票等非线性系统建模。文中采集了股票交易的历史数据作为数据集,对数据使用了归一化和主成分分析方法进行预处理,降低了数据维度,有利于简化模型和降低计算复杂度。针对模型的构建,设计了一种深度稀疏修正神经网络模型(deep sparse rectifier neural netw orks,DSRNN)。DSRNN具有多层网络结构,基于带动量项的BP学习算法训练参数,利用了激活函数ReLU (rectified linear units)和提出的权值初始化方法。并将其与其他三种基于传统方法建立的模型进行了比较,结果表明DSRNN在健壮性、精确度方面都有更好的表现。

The traditional stock forecasting is based on statistics,which is effective in forecasting stock trends but not satisfactory in the ac-curacy of price forecasting. Therefore,we apply the neural network for study on the forecast of stock prices. Neural network has a strongnonlinear fitting,suitable for non-linear system modeling like stock. In this paper,we collect the historical data of stock trading as the da-ta sets,using the normalization and principal component analysis to preprocess them,which are advantageous to model simplification andcomputational complexity reduction. For the model construction,we design a deep sparse correction neural network model （DSRNN）which has a multi-layer network structure,training parameters based on the BP learning algorithm with momentum,using the ReLU （re-active function units） and weight initialization method proposed. Comparing with the other three models based on traditional methods,itshows that DSRNN performs better in terms of robustness and accuracy.