基于奇异谱分析的GRNN模型在金融时间序列中的应用

摘要

奇异谱分析(SSA)作为一类无参数、独立于模型的时间序列分析技术,适用于具有非线性、非平稳性、含噪声的金融时间序列数据的分析与研究.目前,基于SSA的预测通常采用线性递归、BP神经网络等模型,但其预测精度、训练速度并不理想.为此,该文提出基于SSA的广义回归神经网络(GRNN)预测模型,它以SSA所获取的主成份重构序列作为GRNN的输入进行预测.以同方股份收盘价格为测试数据,预测日收盘价.结果表明,基于SSA的GRNN模型预测效果不仅略优于GRNN预测方法,而且明显优于常规的SSA算法.%Singular spectrum analysis (SSA) is a kind of non-parameter and independent model time series analysis technique, which can be suitable for analyzing and studying nonlinear, non-stationary and noisy financial time series. Nowadays, the prediction based on SSA often adopts linear recursion, BP neural network and others as its models. However, the prediction accuracy and training speed is not perfect. Therefore, this paper proposes a new method called general regression neural network (GRNN) based on SSA that uses reconstructed series of components from SSA as its inputs and makes the closing price of tong fang as test data to forecast daily closing price. Experimental results show that the improved method is much better than original one and also slightly better than GRNN.