摘要
高频面板数据在时间维度的频繁波动给聚类的准确性造成了很大干扰。综合考虑这一问题,从小波分解的角度提取了面板数据主成分降维后指标的综合得分序列,利用小波变换提取综合得分序列的"周期"特征、低频部分的"均值"特征与"趋势"特征、高频部分的"波动"特征,最后采用熵值法对这些特征进行赋权并利用赋权后的特征数据和系统聚类方法实现高频面板数据聚类。通过股票高频面板数据的实证分析表明,该方法的聚类效果良好。
The frequent fluctuation of high-frequency panel data in time dimension causes great interference to the accuracy of clustering.In this thesis the comprehensive scoring sequence of the index is extracted after the panel data indicators is reduced in dimensionality by principal component analysis from the angle of wavelet decomposition.Then the periodic features,mean features and trend features in the low frequency domain and features in the fluctuation of variance in the high frequency domain of the comprehensive scoring sequence are extracted through the application of wavelet transform.Finally,the entropy method is used to weight these features,and the high frequency panel data clustering is realized by using the weighted characteristic data and the hierarchical clustering method.The effect of clustering proves good in the empirical analysis of high frequency panel data in stock.
出处
《统计与信息论坛》
CSSCI
北大核心
2018年第2期46-51,共6页
Journal of Statistics and Information
基金
国家社会科学基金项目<空间定位抽样技术在民族地区经济调查中的理论及应用研究>(13BTJ009)
国家自然科学基金项目<复合聚类网络同步能力的尺度可变性及其粗粒化方法研究>(61563013)
关键词
高频面板数据
主成分分析
小波分解
特征提取
聚类分析
high-frequency panel data
principal component analysis
wavelet decomposition
feature extraction
clustering analysis