Research On Time Series Prediction Based On Component Decomposition

With the advent of the big data era,the demand for time series data prediction in various fields is increasing.Time series data is a series of random variables indexed by time,which is essentially an implementation of a stochastic process.Time series prediction employs various time series analysis methods to explore the internal rules of the time series,and estimate the future data of the sequence.Prediction models with excellent performance can help researchers better analyze data and make decisions,and reduce the negative impact caused by various unstable factors.Therefore,time series prediction technology has extremely important research value.And it has become a research hotspot for decades.In recent years,the volume of time series data is dramatically increasing.How to extract features from time series and capture the latent patterns in data has become a new challenge.This thesis analyzes and studies the fluctuation characteristics of different components in time series data based on frequency and time domains.And then it further explores the latent patterns within the data.To address the shortcomings of previous time series prediction models,we propose two novel time series prediction models: a time series prediction model based on multi-scale kernel adaptive filtering and a time series prediction model based on robust trend filtering decomposition.The main contributions are as follows:(1)To fully explore the multi-frequency trading patterns in time series data,we propose a time series prediction model based on multi-scale kernel adaptive filtering(MS-KAF).The proposed MS-KAF models the multi-scale frequency domain information inherent of time series data.It mainly consists of three stages: firstly,the model decomposes the input data into sub-sequences with different frequency domain information by stationary wavelet transform.In view of this,the model applies kernel adaptive filtering to perform sequence learning on the sub-sequences,so that it can describe the fluctuation patterns of stock time series data with different scales and capture potential multi-scale trading patterns.Finally,the model predicts the future stock returns by the captured multi-scale trading patterns.Experimental results show that compared with existing major prediction models,MS-KAF can obtain higher return rates and better stock return prediction performance.(2)To capture the temporal fluctuations of time series data,a time series prediction model based on robust trend filtering is proposed(RTF-KAF).It applies robust trend filtering which is insensitive to noise and has high accuracy in capturing trend changes to decompose stock time series data.And it aiming to acquire trend component with slow-varying fluctuation and residual component with random fluctuation.Then the kernel adaptive filtering method is employs to model the fluctuation characteristics and predicting future returns of stocks based on the trend component and the residual component.Experimental results show that RTF-KAF can accurately capture the multi-scale temporal fluctuations of time series,and effectively improve the accuracy of time series prediction.The proposed MS-KAF and RTF-KAF can decompose time series data based on both frequency and time domains,respectively.The MS-KAF prediction model achieves higher returns by capturing the multi-frequency trading patterns of stock time series data.The RTFKAF model effectively improves the model’s prediction accuracy by capturing the temporal volatility features of stock time series data.In addition,this thesis also provides some research ideas for the future works in this field.