Constructing and Understanding Customer Spending Prediction Models

Prediction models are being used more and more widely in many sectors. FinTech (Financial Technology) is not an exception. Many problems in FinTech can be considered prediction problems. Some notable examples are predicting the probability that a transaction is fraudulent or predicting the most suitable company to invest in, given some constraints. In this research, the focus is on customer spending prediction. More specifically, we are interested in knowing how much a customer may spend in a period given her past purchases. Such information is crucial for the optimal planning and budgeting of businesses. As a first step in tackling this prediction problem, this research explores the feasibility of different statistical methods and machine learning algorithms in accurately predicting customer spending. The subjects we investigate in this research include Beta Geometric/Negative Binomial Distribution (BG/NBD), Gamma–Gamma, Linear Regression, Random Forest, and Light Gradient Boosting Machine (LightGBM). To make the prediction models and their results more accessible to the average users, we utilize information visualization as the primary communication with human users. We hope doing so can bridge the gap between prediction performance and users’ insight into the reasons behind the performance. With better insight, users can make more appropriate decisions in selecting a method/algorithm to build a prediction model under a specific circumstance. The result of this research can also serve as a foundation for more in-depth work on the same problem in the future.