Title: Probabilistic forecasting of retail sales: A quantile regression approach
Authors: Mikhail Zhelonkin - Erasmus University Rotterdam (Netherlands) [presenting]
Abstract: Accurate forecasting of sales is essential in retail for operations and management. The vast majority of research concentrates on point forecasts. In addition to point forecasting the forecasting of the entire distribution is helpful. It can be used for several purposes, including personnel scheduling, shelf replenishment, optimization of the shelf space, to mention a few. The data is characterized by strong seasonality, time varying volatility and skewness, presence of outliers (big orders) and natural contamination by out-of-stocks. Clear, that the use of robust methods in such circumstances is highly desirable. We propose to use the $L_1$ penalized quantile regression with harmonic functions as covariates. This allows us to solve two problems simultaneously: first, it provides the characterization of the distribution of sales, and second, under assumption of moderate number of out-of-stocks (less than 50\%), we obtain a forecast of the actual demand.