| With the rapid development of China’s economy and the continuous improvement of securities investment fund market mechanism,securities investment fund has become one of the most popular investment tools for the majority of investors.With the expansion of fund scale and the increasing demand for fund investment,the income of fund has become the most concerned issue for investors.Both fund management companies and individual investors are eager to benefit from the fund market,so the correct prediction of fund trends has become an important research topic.In view of the characteristics of time series data of fund market,such as unstable,nonlinear and high noise,this thesis proposes a data pre-processing method of wavelet threshold denoising,and selects a strategy from the denoising performance of coif N,db N and sym N wavelet functions under different order,decomposition level and threshold processing methods.In variable selection,a total of ten variables,such as weekly net fund share,cumulative net share,net growth rate,weekly disclosure days,industry concentration,shareholding concentration,consumer price index,are selected,involving the performance of fund net value,asset allocation,macroeconomic and other aspects;Several kernel functions and multi-kernel methods based on linear weighted combination are used to select the optimal kernel function.In order to solve the hyperparameter optimization problem of support vector regression model,the traditional grid search method,genetic algorithm and particle swarm optimization algorithm are used to solve the optimal parameter combination.The main conclusions are as follows: First,for fund time series data,co IF8 wavelet coefficients processed by soft-hard compromise have the best denoising performance under two-level decomposition,and the model’s robustness is improved after being processed by wavelet threshold denoising method.Second,MKL multi-core learning can reduce model training efficiency,while single-core RBF kernel function has better performance both in accuracy and efficiency.Thirdly,compared with the grid search method,the optimization algorithm greatly shortens the searching time and improves the prediction accuracy.Meanwhile,the performance of genetic algorithm is superior to particle swarm optimization algorithm in all evaluation indexes.Finally,based on the above conclusions,the WD＿GA＿SVR prediction model is constructed and compared with the BP neural network and random forest model.At the same time,eight different types of fund products are used for empirical prediction.The results show that WD＿GA＿SVR model not only performs better than other models,but also can accurately predict the net value trend of different types of funds.The research shows that the prediction model constructed in this thesis has good general adaptability and can provide effective data reference for fund investors,management companies and regulatory authorities. |
