本文关注机器学习方法在中国市场主动管理型权益类公募基金截面收益率预测问题的应用。本文通过构造和使用一个较为详尽的基金特征集合，并应用不同的机器学习模型对中国市场公募基金的截面收益率预测问题提供了实证证据，并利用非线性模型显著提升了投资组合表现。具体而言，基于随机森林、梯度提升回归树和神经网络模型的投资组合在2014年1月至2022年4月的测试集内，月度超额收益率为1.58%-1.88%，同期万得偏股混合型基金指数的表现则为0.98%。本文的研究结果为投资实践中构建基金投资策略提供了重要借鉴，对国内公募FoF和基金投顾业务的发展有一定的推动作用，具备一定的现实意义。其次，本文通过变量重要性分析发现某些类型的基金特征是目前相关文献所忽略的重要特征，然而这些重要特征的纳入对于改进模型预测能力和机器学习投资组合绩效至关重要。通过变量重要性分析可以发现，非线性模型捕捉到了线性模型所没有挖掘到的重要特征和非线性关系，从而能够部分解释非线性模型表现更优的原因。此外，本文同时也发现中国公募基金市场较为重要的三类基金特征，分别是管理能力类、风格类和基金持仓类特征，进一步丰富了学术和业界对中国公募基金市场的理解。最后，本文通过对中国市场基金投资者的资金流入流出数据进行分析发现了一些投资者的买卖行为特点，并发现基金市场存在一定程度的资金错配现象。总体来讲，无论是机构投资者还是个人投资者对计算复杂但包含预测基金收益率信息的基金特征重视程度不足。过去1年历史收益是一个能稳健预测未来基金流的指标，但是个人投资者的买卖行为受基金历史表现的影响相较机构投资者而言更大。此外，机构投资者的长期买卖行为受到基金持仓类信息的影响，说明机构投资者在进行长期投资决策时可能使用了基金持仓的信息从而为公募基金带来了长期的资金流入。本文的实证分析强调了本文所构造的基金特征在投研信息生产中的重要性，并指出在公募基金投资中存在的资金错配现象。
This study applies machine learning methods in predicting the cross-sectional return of actively managed equity mutual funds in the Chinese market. First, it provides empirical findings on the prediction of cross-sectional returns of mutual funds in the Chinese market by constructing and using a comprehensive set of fund features and employing various machine learning models. The use of nonlinear models significantly improves the portfolio performance. Specifically, the monthly excess return of the portfolio, as determined by the random forest, gradient boosting regression tree, and neural network models, during the test set period from January 2014 to April 2022, ranges from 1.58% to 1.88%, whereas the performance of the Wind equity hybrid fund index is 0.98% over the same period. The empirical results presented in this paper serve as a crucial point of reference for the formulation of practical fund investment strategies, positively affecting the development of domestic fund of funds (FoF) and fund investment advisory business, which holds significant practical significance.Furthermore, through variable importance analysis, this study determines crucial fund features that have been currently overlooked in the existing literature, despite the fact that the inclusion of these important features is imperative for improving model prediction ability and optimizing machine learning portfolio performance. By analyzing the importance of variables, it can be found that nonlinear models capture important features and nonlinear relationships that have not been discovered by linear models, demonstrating the nonlinear models’ superiority when compared with linear models. In addition, this study identifies three significant categories of fund characteristics in the Chinese mutual fund market, that is, managerial skills, managerial style, and fund holdings, which contributes to the existing academic and industry knowledge and understanding of the Chinese mutual fund market.Finally, this study analyzes the fund flow of fund investors in the Chinese market, discovers key characteristics of their buying and selling behavior, and highlights the presence of a certain degree of capital misallocation in the fund market. In general, it is important for fund investors to place greater emphasis on fund characteristics that involve complex calculations while providing insights into predicted fund returns. The historical returns observed over the previous year can serve as a dependable indicator for predicting future fund flows. However, in contrast to institutional investors, individual investors tend to be more influenced by the fund’s historical performance of a fund when it comes to their buying and selling behavior. On the other hand, the long-term buying and selling behavior exhibited by institutional investors is influenced by information about fund holdings, indicating that institutional investors may have taken into account information about fund holdings during their long-term investment decision-making process. The empirical analysis in this study emphasizes the importance of the fund characteristics in producing investment research information, and points out the occurrence of capital misallocation in mutual fund investments.
