Sales forecasting at the UPC level is important for retailers to manage inventory. In this paper, we propose more effective methods to forecast retail UPC sales by incorporating competitive information including prices and promotions. The impact of these competitive marketing activities on the sales of the focal product has been extensively documented. However, competitive information has been surprisingly overlooked by previous studies in forecasting UPC sales, probably because of the high-dimensionality problem associated with the selection of variables. That is, each FMCG product category typically contains a large number of UPCs and is consequently associated with a large number of competitive explanatory variables. Under such a circumstance, time series models can easily become over-fitted and thus generate poor forecasting results.

Our forecasting methods consist of two stages. At the first stage, we refine the competitive information. We identify the most relevant explanatory variables using variable selection methods, or alternatively, pool information across all variables using factor analysis to construct a small number of diffusion indexes. At the second stage, we specify the Autoregressive Distributed Lag (ADL) model following a general to specific modelling strategy with the identified most relevant competitive explanatory variables and the constructed diffusion indexes.

We compare the forecasting performance of our proposed methods with the industrial practice method (benchmark model) and the ADL model specified exclusively with the price and promotion information of the focal product. The results show that our proposed methods generate substantially more accurate forecasts across a range of product categories.