Study On Application Of Support Vector Regression In High Frequency Financial Time Series

The research on high frequency time series has always been a hotspot in the research of data scientists,along with the continuous development of machine learning algorithms,a new breakthrough has been found for its research,and in many types of high frequency time series,the prediction of high frequency financial time series data is one of the most difficult time series to predict.Although the deep learning represented by neural network can achieve high precision,it requires a large amount of data,slow operation speed,high computer hardware requirements,easy to fall into the local extremum and a series of inherent defects,so that it can not fully adapt to highfrequency financial time series data such as a limited number of samples and there is a large number of noise data types.A single machine learning model is difficult to be competent for the study of high-frequency financial time series,while the research focus at home and abroad is mainly focused on the improvement of a single model and parameter optimization,it is clear that the study of high-frequency financial time series has yet to be further strengthened and enriched.Aiming at the inherent characteristics of high frequency time series,such as nonlinearity,non-stationarity and low signal-to-noise ratio,in order to improve the prediction accuracy of high frequency time series data,based on the improved evolutionary algorithm parameter optimization,this paper constructs a K-means-based one-step prediction model and multi-step rolling prediction model,respectively,in view of the characteristics of high frequency time series data non-stationarity and low signalto-noise ratio,and the support vector regression model of clustering is used for onestep prediction,and the support vector regression model based on wavelet denoising and K-means clustering is used for multi-step rolling prediction.The empirical results show that the performance of the two models in one-step and multi-step rolling prediction is better than that of the support vector regression model.In the first part,the research status of high frequency time series at home and abroad.From traditional mathematical statistics model to machine learning,from neural network to support vector machine,this paper compares the advantages and disadvantages of support vector machine and neural network,and focuses on the inference process from support vector machine to support vector regression.In the second part,according to the characteristics of high frequency financial time series nonlinearity,the support vector regression model is selected for prediction.This paper introduces the influence of model parameters on the accuracy of the model,and makes an improvement on the basis of the evolutionary algorithm,and uses the improved evolutionary algorithm to predict the high frequency time series data with the model after parameter optimization.The empirical results show that the support vector regression model based on evolutionary algorithm has higher prediction accuracy than the support vector regression model based on the traditional parameter optimization algorithm,and its convergence speed is faster,and to a certain extent,it can enlarge the selection range of the initial value,and it is more efficient and friendly for parameter optimization without prior experience.In the third part,aiming at the problem of non-stationary data which can not be overcome by support vector regression model in high frequency financial time series,a support vector regression model based on K-means clustering is constructed on the basis of improved evolutionary algorithm parameter optimization.On the one hand,the improved evolutionary algorithm is used to find the optimal parameters for the model.On the other hand,the unsupervised learning algorithm is used to cluster the data,which can theoretically distinguish the abnormal and normal fluctuations of time series and make the data more stable.Empirical results show that the SVR model based on Kmeans clustering can provide better prediction accuracy than the traditional SVR model in one-step prediction under the same conditions of parameter optimization with improved evolutionary algorithm.In the fourth part,aiming at the characteristics of low signal-to-noise ratio of high frequency financial time series and rolling prediction mode,in order to meet the practical needs of rolling prediction,a wavelet denoising algorithm is added to support vector regression model based on improved evolutionary algorithm and K-means clustering.The model reduces the difficulty of prediction,but also has practical significance.Considering the practical application,the model can be used for multistep rolling prediction,and the performance of the model is tested by sliding time window.The empirical results show that the SVR model based on wavelet denoising and K-means clustering has better prediction accuracy and stability than the SVR model based on wavelet denoising under the same conditions of parameter optimization with improved evolutionary algorithm.Finally,the paper summarizes the above,points out the shortcomings of this paper,and analyses the possible research directions in the future.