Prediction Of Stock Fund Rankings Based On LightGBM Model

In the current era of "great asset management" and " Houses are for living in,not for speculation ","relocation of residents’ wealth",and "Individual endowment Plan",the development prospects of equity products are tremendous,and market volume of the stock publicly offered funds is rapidly expanding;Whether for investors who invest directly in funds or institutional investors who set up FOF products,the research and evaluation of fund performance is an indispensable core process for investors to select the best and eliminate the bad when selecting funds,and the core manifestation of the investment research capabilities of portfolio investment managers and fund investment consultants.Fund research has theoretical research value and practical significance for guiding investors how to construct investment portfolios.In this context,this thesis tries to discuss how to better evaluate fund performance and fund selection.In the related researches on fund performance evaluation,most of the previous scholars tested and analyzed singly from the perspective of time series or crosssection,ignoring many qualitative features,such as the impact of changes in fund managers,styles’ drift in funds,etc.And the research samples are often limited to a few specific funds rather than a sample of the entire class of funds.Therefore,based on the review of previous studies,this thesis puts forward a kind of innovation based on stock fund position information of the fund manager investment decision-making factor,combined with the features of many theories based on previous fund research.In this thesis,with the help of nonlinear modeling ability,using LightGBM to model the task about performance evaluation and fund selection within chinese stock funds and aggressive hybrid funds,construction of fund portfolios,and compared with a variety of fund evaluation methods to illustrate the system of screening stock fund based on the model of this thesis has advantages,filling the gap and deficiencies of the previous fund study.This thesis conducts empirical analysis,the results of the generalized least squares linear regression method with the next period funds’ ranking in the same category demonstrate that,the model’s adjusted R-squared is improved after joining the variables proposed by this thesis.In addition,the significance level of these variables reachs 1%.Rank IC analysis also confirms the effectiveness of the fund investment decision-making capacity indicators proposed in this thesis.This thesis uses machine learning to predict fund performance,the model performance is at a very good level and well robust.The test results of the Fama-French five-factor model show that the simulated portfolios of long and short portfolios have differences in style exposure,but the intercept terms are all positive,which may indicate that Chinese public fund managers generally have the investment skills to exceed the market returns after stripping out the risk-free returns and profit and loss of style.The reason for the difference of return level between long and short portfolios is the difference of excess return ability and style exposure.The evaluation and prediction methods are implemented within the same fund classification.The above description shows that this thesis takes into account the differences in the inherent characteristics of the funds,and the research ideas and data processing are more reasonable.The innovation of this paper lies in the combination of machine learning with stock multi-factor model,fund performance attribution model,and fund performance evaluation model,and a systematic solution of how to screen out future high-performing funds is proposed.In the literature study of the author’s involvement,this thesis is the first research to utilize the funds’ holdings information of investment portfolios to build simulation investment transaction portfolios,form a measurement system of the fund manager’s decision-making capacity factors,and conduct fund performance prediction.This thesis is also the first proposal to make use of multi-type fund feature panel data to predict funds’ ranking in the same category rather than the funds’ yield,considering the concept that publicly offered stock funds is a relative return product.This thesis has made innovations in research perspectives,theoretical practice,model application and data utilization,which has profound practical significance and provides a new thinking for promoting the application of artificial intelligence method in fund investment in China.