The practice of fund intelligent counseling system in the context of financial inclusion

Abstract

Based on the current "product-centered" fund sales model, fund sales institutions sell fund products in a revenue-oriented, flow-centered manner, with insufficient exploration of customers’ real needs, failing to achieve an effective match between customers and products, resulting in a poor sense of accomplishment. Therefore, under the background of regulation to promote the fund investment consulting business, fund companies can learn from overseas experience in the intelligent investment consulting business, and develop intelligent recommendation system for pre-investment customer products matching. On the one hand, it can improve the recommendation efficiency for fund selling, increase the purchase rate, and reduce the sales costs. On the other hand, it can facilitate the sales model transformation to a customer-centered business, allowing the fund company to offer consulting service through direct selling platform, and cover more customer groups and realize financial inclusion. The research content of this thesis is mainly divided into two parts: at the very beginning, in order to confirm that there exists a serious phenomenon of poor investor holding experience in China's fund investment market, and to explore the reasons for it, this thesis analyzes the investors’ fund subscription behavior as well as the fund issuance situation from the empirical point of view, providing a series of empirical evidences. Based on the empirical studies for the fund market, this thesis’s further take the usage of one of the top fund company's direct sale platforms to build an intelligent recommendation system with machine learning algorithms, to improve the direct sales ability and strengthen the customer recognition, which can help the fund company to cover more customer groups. In this thesis, the empirical study of China's mutual fund market analyzes the long-existing effect of "funds make money, funders don't", which is rarely mentioned in the literature that focuses on the Western mutual fund market. In the previous research on the investment behavior of fund investors, most of the discussion is from the investor behavioral preferences, less discussion of the behavior of the fund institutions in which the impact. In the mainstream research on institutional behavior analysis, the research object mainly focuses on the asset management behavior of institutions and the principal-agent problem that exists therein. On the basis of this series of literature, this paper discusses the "principal-agent" problem between investors and fund companies, and combines it with the reality of China's fund market. This thesis shows significant innovative highlights at both academic and practical aspects. From the academic point of view, this thesis relies on the massive product data and user data of the head fund companies, and combines the experience of business development to transform the traditional user image and product image technology, to create a rich user and fund product labeling system, and to solve the challenge of data completeness in the intelligent recommendation system. This study not only makes in-depth optimization progresses for the specific challenges in the fund recommendation process, but also carries out innovative exploration at the algorithmic level. Compared with a single recommendation model in the industry, this thesis covers several generations of intelligent recommendation technologies, successively using Statistical Mining Type Model, Neighborhood and Lookalike Association Model and Recall-Ranking Model, demonstrating the process of model iteration, optimization and complementation, which significantly improves the accuracy and diversity of the recommendation system. At the level of business scenarios, the whole model can not only recommend products with historical data, i.e., user behavior, but also recommend new products that are not supported by user behavior data and need a cold start. From the perspective of business practice, by providing personalized and high-quality fund recommendation services, the intelligent recommendation system can meet customers' investment needs and preferences, enhancing customers' trust and loyalty to the platform. By stimulating customers' investment behavior and consumption willingness, the model can reduce the costs of marketing, and cover more customer segments, which is in line with the idea of financial inclusion. For the research methodology, this thesis realizes the combination of traditional financial and economic research methods and frontier of machining learning algorithms in the field of engineering, with the unification of research problems and research methods in different research fields. In the empirical financial studies, we utilize the Two-Way Fixed Effect Regression method and time series analysis to examine the research hypotheses. The landing of the intelligent recommendation system involves four aspects: customer image, product image, recommendation algorithm and model evaluation. The research methodology for intelligent recommendation systems includes the above four points. Specifically, this thesis focuses on customer and product profiling construction, with the underlying machine learning algorithm, as well as the principles of the recommendation algorithms. Lastly, this study uses the hyperparameter tuning method and the offline and online evaluation process for the intelligent recommendation system. Three different types of recommendation models are successfully constructed in this thesis, namely, statistical mining type model, nearest neighbor and lookalike association model, and recall-ranking class model. These models have better performance on various evaluation indexes in offline training, the key indexes’ precision rates are 0.94/0.84 and the AUC values are as high as 0.97/0.82 on the training/testing set, respectively. Those results are dominating the performance of the benchmark models (e.g., random recommendation, popular recommendation, etc.). Within twice independent online evaluations, our recommendation algorithm model made good performances, with customer coverage as well as push accuracy rate significantly better than traditional manual recommendation channel. The model group (experimental group) efficiently covers the screened customer groups for the manual recommendation group (control group) in dimensions of the push population and the purchase population. The click rate of the long-tail customers of the incremental push is significantly higher than that of the control group, which indicates the good performance of our model on accuracy, stability, generalization ability and adaptability. The intelligent recommendation system constructed in this thesis not only realizes the purpose of reducing marking cost and increasing efficiency in fund sales business, but also consistent with the financial inclusion regulation, which provides valuable experience for the development of domestic intelligent investment adviser business.