An Exponential Factorization Machine with Percentage Error Minimization to Retail Sales Forecasting

This article proposes a new approach to sales forecasting for products (stock-keeping units [SKUs]) with long lead time but short product life cycle. These SKUs are usually sold one season only, without any replenishments. An exponential factorization machine (EFM) forecast model is developed solve this problem which not only takes into account SKU attributes, also pairwise interactions. The EFM significantly different from the original Factorization Machines (FM) two fold: (1) attribute-level formulation explanatory/input variables; and (2) positive response/output/target variable. formation excludes infeasible intra-attribute interactions results in more efficient feature engineering comparing conventional one-hot encoding, while demonstrated effective than log-transformation skewed distributed responses. In order estimate parameters, percentage error squares (PES) (ES) minimized by proposed adaptive batch gradient descent method over training set. To overcome over-fitting problem, greedy forward stepwise selection select most useful attributes Real-world data provided footwear retailer Singapore used testing approach. performance terms of both mean absolute (MAPE) (MAE) compares favorably off-the-shelf models reported extant demand studies. effectiveness external public datasets. Moreover, we prove theoretical relationships between PES ES minimization, present an important property minimization regression models; that it trains underestimate data. fits situation where unit-holding cost much greater unit-shortage (e.g., perishable products).