Clustering Analysis Based On Forecast Densities And Its Sensitivity Analysis For Different Distance Choice

A new clustering method for time series is proposed, it is based on the full probability density of the forecasts. First, a resampling method combined with a nonparametric kernel estimator provides estimates of the forecast densities. It is proposed Kullback-libler distance, L2 distance, symmetric Kullback-libler distance, and separately use each kind of distance to carry on clustering analysis, obtains results of clustering through computer simulation; by gathering result of clustering, it obtains that symmetric Kullback-libler distance can fit for this kind clustering method best. The estimation of distance uses Monte Carlo method.This article is divided into five parts: The first part introduces clustering analysis briefly; The second part mainly introduces the main theories, methods and steps of clustering analysis which based on forecast density; The third part arises three kinds of distance, establishes the models and calculates exactly value of each distance; The fourth part carries on the simulation study, and in simulation, it uses each kind of distance to carry on clustering separately, then compares the clustering results; The fifth part obtains the conclusion that symmetric Kullback-libler distance can fit for this kind of clustering method best, and analyses the reasons of our results.