Research On Portfolio Strategy Of Stock Fund Based On Learning To Rank

With the continuous development of the domestic economy,the proportion of national savings is increasing,and the willingness to increase the value of original asset investment is also increasing.Therefore,securities investment funds have also ushered in opportunities for vigorous development.In recent years,domestic securities investment market has shown a volatile trend,coupled with the complex and volatile international situation,leading to significant plate rotation in most funds,especially equity funds heavily invested in the stock market.In this context,investors’ willingness to avoid risks in financial investment has strengthened.The emergence of Fo F funds has introduced the investment concept of using funds as investment targets to construct investment portfolios,which uses funds as the target of portfolio investment to conduct secondary smoothing of risks,hoping to reduce volatility in volatile market conditions.With the continuous rise of quantitative investment,more and more researchers have begun to study the practice of quantitative investment in the field of funds.However,most of the relevant research mainly focuses on fund performance attribution and performance prediction direction,and there are not much research targeting on funds’ selecting.In addition,most of the asset allocation issues of investment portfolios are based on traditional asset allocation models such as risk parity and minimum variance,the non normality of fund returns is not included in the risk measurement of asset allocation.To sum up,this article will combine the method of learning-to-rank with the theory of multiple factor stock selection models to predict and screen the ranking problems of stock based funds,and use the improved asset allocation theory of higher order moments to allocate asset weights,hoping to build effective portfolio strategies.In the empirical section of this article,a study is conducted on 2007 equity funds from January 2016 to December 2022.The frequency of portfolio adjustment is set to monthly,and the selection of funds is based on the sorting prediction results.Because of the fund data is sequential,this article updates the training set through sliding window method.Based on previous research on fund performance,this paper constructs a factor library and conducts effectiveness analysis on these factors.Factors are filtered through Rank IC values and correlation coefficients within each sliding window and then modeled.Three different methods of learning-to-rank,RFRanker based on Pointwise,XGBRanker based on Pairwise,and Lambda MART based on Listwise,are used to solve the sorting problem of selecting funds,the improved risk parity model based on higher order moments is used for asset allocation,and the performance results are compared with those of traditional asset allocation models.This paper uses the data from January2019 to December 2022 to conduct a strategy backtesting,and evaluates the strategy in combination with the ranking evaluation index NDCG and fund performance indicators.The empirical results show that the results of using XGBRanker based on Pairwise and Lambda MART based on Listwise to base are better than traditional multifactor models in terms of annualized yield and Sharp ratio.Among them,Lambda MART based on Listwise has the best effect,achieving an annualized yield of 17.87% and a Sharp ratio of 0.8.The risk parity model combined with skewness successfully improves the Sharpe ratio of the portfolio to 0.87,and 29.32% of the maximum drawdown is significantly improved compared to the original risk parity model,further increasing investment returns.The empirical results of this paper verify the effectiveness of ranking learning in selecting funds,providing a new approach to quantitative fund selection and constructing fund portfolios.