中文摘要：

为提高多维人民币汇率预测的精度和降低网络的训练负担，建立了一种融合独立成分分析（ICA）与BP网络理论的多维时间序列预测模型．首先提取训练集的独立成分，在识别最优滞后期组合的基础上，分别对各独立成分训练得到稳定的BP神经预测网络，然后结合转换矩阵，对测试集进行预测测试．以2008年以来的多维人民币汇率数据为计算实例，研究结果表明IC—BP网络具有较好的预测精度；基于降维技术的IC—BP网络可降低模型整体的训练负担且具有良好的预测精度．该方法能够实现了ICA技术与BP神经网络预测模型的优势融合，在多维人民币汇率预测方面表现出较强的能力．

英文摘要：

In order to improve the forecast accuracy of multivariate RMB exchange rate and reduce the training burden, BP neural networks model combining with independent components analysis used for forecasting multivariate time series was presented in this paper. Firstly, it conducted the extraction of independent components. Then the BP neural networks which used for forecasting each specific independent component was established by identifying the optimal lag periods. Finally, with the help of the mixing matrix, it was applied in the testing set. Empirical forecasting results which adapt the RMB exchange rate data from 2008 show that IC-BP networks has a better prediction. IC-BP model based on dimension reduction can reduce the computational burden and has good prediction accuracy. This method compromises the merits of ICA and BP neural networks. IC-BP neural model presents great ability to forecast multivariate RMB exchange rate.