Using Regression and Machine Learning to Forecast Customer Spending Patterns and Turnover

The study is based on a customer database with 440,832 valid data and uses the following methods: Matrix based Regression and Machine Learning. The first goal is to see if there is a correlation between the customer's age and the number of dollars spent, and if there is a correlation, is there any gender trend? The mathematical underpinning is introduced via the least squares residual minimization, matrix conditioning and the full-column-rank condition for uniqueness. Using the entire cleaned dataset, different linear and quadratic models are fit for female customers and male customers. Age trends are small and negative in both linear models (R-squared values less than 0.01); quadratic models do not help much. A separate computational-cost exercise checks the first 1, 000 observations and finds a balanced gender composition of 51.4% female and 48.6% male. The second will involve prediction of turnover based on age, gender, duration of engagement, purchases, number of interactions for support, and amount spent. The ROC-AUC for logistic regression is 0.906 and the ROC-AUC for random forest is 0.967. The best practical indicator is the turnover rate which rises sharply when support contacts become more numerous. Results show that the age spending regression has weak explanatory structure, whereas turnover is strongly modeled in a multivariate framework.