Multi-period-ahead forecasting with residual extrapolation and information sharing - utilizing a multitude of retail series

Abstract

Multi-period sales forecasts are important inputs for operations at retail chains with hundreds of stores, and many different formats, customer segments and categories. In addition to the effects of seasonality, holidays and marketing, correlated random disturbances also affect sales across stores that share common characteristics. We propose a novel method, Two-Stage Information Sharing that takes advantage of this challenging complexity. In this method, segment-specific panel regressions with seasonality and marketing variables pool the data, in order to provide better parameter estimates. The residuals are then extrapolated non-parametrically using features that are constructed from the last twelve months of observations from the focal and related category-store time series. The final forecast combines the extrapolated residuals with the forecasts from the first stage. Working with the extensive dataset of a leading Turkish retailer, we show that this method significantly outperforms both panel regression models (mixed model) with an AR(1) error structure and the autoregressive distributed lags (ADL) model, as well as the univariate exponential smoothing (Winters') method. The further out the prediction, the greater the improvement.