"Multi-Period Ahead Forecasting with Residual Extrapolation and Information Sharing – Leveraging Multitude of Retail Series " , Ö Gür Ali, E. Pinar, International Journal of Forecasting 32 (2), 502-517, 2016.
Multi-period sales forecasts are important inputs to operations at retail chains with hundreds of stores, many formats, customer segments and categories. Beyond seasonality, holidays and marketing, correlated random disturbances affect sales across stores that share common characteristics. We propose a novel method, 2 Stage Information Sharing, that leverages this challenging complexity: Segment-specific panel regressions with seasonality and marketing variables pool the data for better parameter estimates. The residuals are 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 first stage forecasts. Working with the extensive dataset of the leading Turkish retailer, we show that the method significantly outperforms panel regression models (mixed model) with AR (1) error structure and the Autoregressive Distributed Lags (ADL) model as well as the univariate exponential smoothing (Winter’s) forecasts. The farther out the prediction, the more the improvement.
Key words: Multivariate time series, Sales Forecasting, Panel Data, Data Mining, Regression, Retail, Multi-period ahead forecast.