Neural Networks for Retail Sales Forecasting

Introduction

Forecasting of the future demand is central to the planning and operation of retail business at both macro and micro levels. At the organizational level, forecasts of sales are essential inputs to many decision activities in various functional areas such as marketing, sales, and production/purchasing, as well as finance and accounting (Mentzer & Bienstock, 1998). Sales forecasts also provide basis for regional and national distribution and replenishment plans. The importance of accurate sales forecasts for efficient inventory management has long been recognized. In addition, accurate forecasts of retail sales can help improve retail supply chain operations, especially for larger retailers who have a significant market share. For profitable retail operations, accurate demand forecasting is crucial in organizing and planning purchasing, production, transportation, and labor force, as well as after sales services.

Barksdale and Hilliard (1975) examined the relationship between retail stocks and sales at the aggregate level and found that successful inventory management depends to a large extent on the accurate forecasting of retail sales. Agrawal and Schorling (1996) and Thall (1992) also pointed out that accurate demand forecasting plays a critical role in profitable retail operations, and poor forecasts would result in too-much or too-little stocks that directly affect revenue and competitive position of the retail business. The importance of accurate demand forecasts in successful supply chain operations and coordination has been recognized by many researchers (Chopra & Meindl, 2007; Lee, Padmanabhan, & Whang, 1997).

Retail sales often exhibit both seasonal variations and trends. Historically, modeling and forecasting seasonal data is one of the major research efforts, and many theoretical and heuristic methods have been developed in the last several decades. Different approaches have been proposed, but none of them has reached consensus among researchers and practitioners. Until now, the debate is still not abated in terms of what the best approach to handle the seasonality is.

On the other hand, it is often not clear how to best model the trend pattern in a time series. In the popular Box-Jenkins approach to time series modeling, differencing is used to achieve stationarity in the mean. However, Nelson and Plosser (1982) and Pierce (1977) argued that differencing is not always an appropriate way to handle trend, and linear detrending may be more appropriate. Depending on the nature of the non-stationarity, a time series may be modeled in different ways. For example, a linear or polynomial time trend model can be used if the time series has a deterministic trend. On the other hand, if a time series exhibits a stochastic trend, the random walk model and its variations may be more appropriate.

In addition to controversial issues around the ways to model seasonal and trend time series, one of the major limitations of many traditional models is that they are essentially linear methods. In order to use them, users must specify the model form without the necessary genuine knowledge about the complex relationship in the data. This is the main reason for the mixed findings reported in the literature regarding the best way to model and forecast trend and seasonal time series.

One non-linear model that recently received extensive attention is the neural network (NN) model. The popularity of the neural network model can be attributed to their unique capability to simulate a wide variety of underlying non-linear behaviors. Indeed, research has provided theoretical underpinning of neural network’s universal approximation ability. In addition, few assumptions about the model form are needed in applying the NN technique. Rather, the model is adaptively formed with the real data. This flexible data-driven modeling property has made NNs an attractive tool for many forecasting tasks, as data are often abundant while the underlying data generating process is hardly known.

In this article, we provide an overview on how to effectively model and forecast consumer retail sales using neural network models. Although there are many studies on general neural network forecasting, few are specifically focused on trending or seasonal time series. In addition, controversial results have been reported in the literature. Therefore it is necessary to have a good summary of what has been done in this area and more importantly to give guidelines that can be useful for forecasting practitioners.