Using transactions data to improve consumer returns forecasting
提出一种利用交易级数据（如购买和退货时间戳）的两步预测方法，在电子和珠宝零售商数据上比现有模型降低10-20%的预测误差，帮助运营经理优化库存、人员安排和退货处理。
Abstract Although generous return policies have been shown to have marketing benefits, such as a higher willingness to pay and a higher purchase frequency, counterbalancing these benefits is an increased volume of consumer returns, which presents significant operational challenges for both retailers and original equipment manufacturers (OEMs). Since accurate return forecasts are inputs into strategic and tactic decision support tools for operations managers, advancements in better forecast accuracy can yield significant savings from the returns management practice. We propose a forecasting approach that incorporates transaction‐level data, such as purchase and return timestamps, and predicts future return quantities using a two‐step “predict‐aggregate” process. To enhance the generalizability of our framework, we test it on two distinct datasets provided by a bricks‐and‐mortar electronics retailer and an online jewelry retailer. We find that our approach demonstrates significant forecasting error reduction, in the range of 10–20%, over benchmark models constructed from common industry practices and the existing literature. As our approach leverages the same data inputs as existing models, it can be easily adapted by practitioners. We also consider a number of extensions to generalize our approach into contexts such as restricted return time windows, new product returns, and inflated same‐day returns. Last, we discuss broad implications of return forecast accuracy improvements in the areas such as inventory management, staffing level, reverse logistics, and return recovery decisions.