Doctoral Dissertations

Kent L. Wolfe

Date of Award

5-1994

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Agricultural Economics

Major Professor

David B. Eastwood

Committee Members

John R. Brooker, S. Darrell Mundy, Pratibha A. Dabholkar

Abstract

Scan data have provided a suitable data source for estimating consumer demand relationships at the retail level. These data capture actual quantity and price information which can be combined with additional explanatory variables, such as advertising, promotion and seasonality, to estimate the retail demand function for specific products. Despite the availability of scan data, relatively little research has focused on forecasting at the retail level.

The competitive nature of the supermarket industry, both the encroachment of warehouse food retailers and generic private label products, have lead to an increased interest in consumer demand analysis at the retail level. The increased competition from nontraditional retail outlets has eroded the traditional supermarket's market share. The nontraditional grocery outlets are perceived to be less expensive than their traditional grocery outlets. Thus, traditional grocery outlet managers have become increasingly interested in reducing operating costs. One method of reducing operating costs is to reduce inventory levels via implementation of an efficient consumer response (ECR) strategy, a version of just-in-time delivery. The ECR strategy has the potential to reduce inventory levels which can directly lower inventory costs. The rise of ECR has created a need for accurate product demand forecasts at the supermarket level to maintain adequate inventory levels. The ability to forecast weekly demand in response to changes in seasonality, holidays, and promotional and advertising campaigns is very important to retail managers for implementation of an ECR strategy.

The objectives of this study are to 1) develop alternative forecasting methods that are suitable for scan data, 2) estimate and compare the alternatives with respect to food groups and individual products in terms of their forecast accuracies using a scan data base, and 3) estimate and compare the alternatives with respect to food groups and individual products in terms of two week trial forecasts.

The theoretical forecasting model was developed utilizing economic theory and previous consumer demand research. The model described weekly product item movement as being a function of own- and cross-prices, own- and cross-advertising (television, radio, and newspaper), holidays, and seasonality. The theoretical model for the brand product also included point of purchase and the start of the Knox County, Tennessee, school year.

The second forecasting model specification was developed using the Box-Jenkins methodology. This technique does not incorporate structural explanatory variables, but rather, identifies and replicates underlying patterns in the data series utilizing past item movement and disturbances in the series.

The third forecasting method combines the structural variables contained in the theoretical model with the pattern identification and replication ability of the Box-Jenkins model to produce a composite model known as a transfer function.

This study utilized weekly scan and advertising data (television, newspaper, radio, and point of purchase) which was supplied by a multi-regional supermarket chain. The data consisted of weekly UPC-level prices, item movement, and chain-initiated television, radio, and newspaper advertising. The data were pooled across five stores that catered to average to above average income food shoppers.

The data were divided into two subgroups. The first subgroup of data was used to estimate the alternative forecasting models and generate product backcasts for technique evaluation and comparison. The second subgroup of data, the last 26 weeks for each product, was used to generate a two week trial forecast. Again, the models and their forecasting abilities were evaluated and compared across alternative methods.

The three alternative forecasting models were estimated using the historic subgroup data. The alternative forecasting models were evaluated individually by the evaluation criteria to choose the "best model" to represent each technique. These model estimates were then used to generate backcasts of the data series for each of the three food products, brand b, group g, and steak. The alternative techniques were then evaluated and compared.

The results of the backcast forecast evaluation and comparison suggested that the transfer function forecast was superior to the Box- Jenkins and theoretical forecast in predicting weekly item movement for brand b and steak. Group g's weekly item movement was best forecast utilizing the Box-Jenkins methodology.

The results of the two week trial forecast evaluation suggested that the transfer function technique was superior to the theoretical and Box- Jenkins techniques in accurately forecasting weekly item movement for each of the three products, a highly process brand, its associated group, and steak a variable weight perishable product. This study has found that the transfer function is the best of the three techniques for use in forecasting weekly retail item movement for both brand and category peanut butter and the steak category. However, the results also indicate that each of the forecasting models was relatively accurate in forecasting actual item movement but performed poorly in predicting directional change.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS