A Data-driven Financial Products Recommendation Model And Algorithm

This paper uses the data from a security company's telemarketing to optimize the recommendation model and algorithm,and studies how to recommend financial products to new customers more accurately.In order to reduce the impact of data loss on the performance of the model,we tried to solve the problem of data loss with three solutions,namely,univariate interpolation,multivariate interpolation and the introduction of new features.We demonstrated the rationality of the above three methods through experiments.In addition,we experimentally demonstrated that the addition of additional socio-economic indicators would significantly improve the accuracy of the model.We examine five classical machine learning models: logistic regression,random forest,support vector machines,neural networks,stacking ensemble algorithm.What's more,we evaluate the model from different perspectives.The advantage of logistic regression and random forest is that the fitting model is easy to understand and also provides good prediction in classification.Comparing with the previous two classical statistical models,support vector machines and neural networks are more flexible,while showing the ability to learn from linear mapping to complex nonlinear mapping.Because of this flexibility,support vector machines and neural networks can often provide accurate predictions.But they are difficult to understand intuitively.Stacking ensemble algorithm fills the gap.Stacking,however,is slow because it involves a two-tier learning process,which means four complex models need to be.Interestingly,the theoretically optimal neural network classification is just of average performance,which fully demonstrates the difference between theoretical model and empirical research.The most profound lesson of this paper is: Data drives production.When we make a decision,we cannot brainstorm a model,rather we should choose the right model according to the data,hence we can make a reasonable and appropriate decision.On one hand,we should consider whether the original data supports the model we build.On the other hand,we need to evaluate whether the output results matches the decision maker's requirement.