基于非线性特征提取和加权K最邻近元回归的预测模型  

Prediction Model Based on a Nonlinear Feature Extraction and Weighted K-Nearest Neighbor

作　　者：唐黎[1] 潘和平 姚一永[1] TANG Li;PAN Heping;YAO Yiyong(School of Inelligent Finance,Tianfu College of Southwesterm University of Finance and Economics,Chengdu 610052,China;Business School,Chengdu University,Chengdu 610106,China)

机构地区：[1]西南财经大学天府学院智能金融学院,四川成都610052 [2]成都大学商学院,四川成都610106

出　　处：《信息系统学报》2019年第2期109-118,共10页China Journal of Information Systems

基　　金：国家社会科学基金项目(17BGL231)。

摘　　要：本文提出了一种智能的金融时间序列预测模型。该模型采用前向滚动经验模态分解(forward rolling empirical mode decomposition,FEMD)对金融时间序列进行信号分解,采用主成分分析(principal component analysis,PCA)对分解后产生的高维向量组进行降维.整个过程是一个复杂的非线性特征提取过程。再将提取的特征输入一种新的利用PCA输出的加权K最邻近法(K-nearest neighbor,KNN)进行回I预测。该模型在特征提取过程的构造和整体结构上都是具有创新性的,并提出了比简单的KNN预测更有效的改进算法。实证结果证实了该模型对中国股票指数的预测效果。This paper proposes an intelligence financial prediction model consists of a forward rolling Empirical Mode Decomposition(FEMD)for financial time series signal decomposition,Principal Components Analysis(PCA)for dimension reduction,and a weighted K-Nearest Neighbor for prediction.Generally,the structure of this model is original.The feature extraction process integrating FEMD and PCA is an advanced special extraction method for financial time series signal analysis.It has the adaptability,comprehensiveness and orthogonality of feature extraction.Moreover,the weighted KNN with PCA loading as weights is more reasonable and has better efect on classifying than a simple KNN,thus it has better prediction performance.The empirical results on CSI 300 prediction has confirmed that the FEPK model performs better than others.

关 键 词：经验模态分解 PCA K最邻近法 特征提取 预测 

分 类 号：F832.5[经济管理—金融学]