Research on Retailer Churn Prediction Based on Spatial-Temporal Features

Accurate predictions about retailer churn, or retailer attrition, enable cigarette companies to follow market developments in an increasingly competitive tobacco market, having immense significance to increase sales and boost their brand power. How retailers place an order, however, is affected by geographic locations, distribution strategies, and marketing events, among others, and it has been fluctuating widely judging from historical data. It is, therefore, rather challenging to screen defecting clients by sorting out piles of orders completed by over five million active retailers, win them back through targeted, in-person visits, and keep the number of retailers steady or growing at a lower cost in terminal management. In this paper, a GBDT-based method of predicting retailer churn was proposed, which merged temporal features with the characteristics of geospatial data through a sliding window over time and geographic rasters, and then balanced samples by taking an approach to adjust the weight of the CART leaf nodes, so as to achieve higher forecast accuracy. Results found that upon the learning of spatial-temporal features, the CatBoost algorithm excelled at prediction and was superior to conventional RFM models in the accuracy and recall rates in 14 or 30 days of retailer churn. To conclude, the proposed method based on spatial-temporal features could deliver desired results when used for predicting cigarette retailer attrition.