Driver-Moderator Method for SKU Sales Forecasting

We develop a method to forecast stock keeping unit sales that is accurate, transparent and consistent in handling similar situations. We leverage the marketing literature to define features reflecting potential drivers of sales in the presence of promotions and potential moderators of these drivers’ effect on sales. We use an epsilon insensitive support vector regression with L1 norm regularization to simultaneously select the relevant ones among the 600+ features and to estimate the model parameters. We compare the performance of the Driver-Moderator Method to that of the Regression Tree with Extensive Features method which had superior accuracy among the 30 methods in a recent study. The Driver-Moderator Method produces superior out of sample accuracy in an evaluation involving black tea sales data from a grocery store chain in Turkey involving five stores and 155 stock keeping units, while supporting managerial decision making with a concise and transparent model.

• Type: Conference
• Year: 2009
• Where: INCDM
• Authors: Özden Gür Ali
• Keywords: Data Mining, Driver-Moderator Method, Extensive Features Method, INCDM 2009, Stock Keeping Units