Aiming at the current problem of limited prediction accuracy of nonlinear financial time series in the shallow model, a novel prediction model of stacked autoencoder neural networks is proposed, consisting of stacked autoencoder model at the bottom and regression neurons at the top.First of all, the unsupervised learning mechanism of autoencoder is applied to identify and learn the features of time series, greedily learning the neural network layer by layer.Then the stacked autoencoder is extended to be the SAEP model with supervision mechanism, and the parameters learned by SAE (stacked autoencoder) are used to initialize the neural network.Finally, the weights are fine-tuned by supervised learning.The experimental design uses the exchange rate time series as the training and test samples, and the effectiveness of the proposed model in the application of exchange rate time series prediction is verified, compared with the more mature methods.%针对目前具有非线性特征的金融时间序列浅层模型预测精度有限的问题,提出一种由底层的栈式自编码器和顶层的回归神经元组成的栈式自编码神经网络预测模型.首先利用自编码器的无监督学习机制对时间序列进行特征识别与学习,逐层贪婪学习神经网络各层,之后将栈式自编码器扩展为有监督机制的SAEP模型,将SAE学习到的参数用于初始化神经网络,最后利用有监督学习对权值进行微调.实验设计利用汇率时间序列作为训练及测试样本,与目前较成熟的方法进行对比实验,验证了所提出的模型在汇率时序预测应用中的有效性.
展开▼
