Supply chain forecasting for retailing and manufacturing

Both manufacturers and retailers rely on forecasts to support their supply chain decisions on production, distribution, purchasing and marketing. Research in the Centre has focused on the systems and processes by which companies produce their forecasts.

In particular, the role of expert managerial judgment in order to understand their value in terms of forecast accuracy. In some circumstances these adjustments lead to larger forecast error. The research has investigated whether changes in software design can overcome this serious problem.

Fildes in a study of supply chain forecasting practices (Fildes and Goodwin, 2007; Fildes et al., 2009) identified the importance of expert judgment in forecasting accuracy. This affects both the models that should be used, the organisational processes by which the forecasts are produced and the software design that has the best organisational fit to the forecasting tasks undertaken (Asimakopoulos et al., 2011).

A second strand of research has been concerned with developing new statistical models to improve accuracy. A key issue is whether the incorporation of downstream point-of-sales data helps the manufacturer. Second, whether promotional information can be effectively introduced.  

Kourentzes, working with Centre associate and visiting Marie Curie fellow Trapero (now at the U. Castilla - La Mancha) have developed new models that extend exponential smoothing in intuitive ways to capture additional supply chain information (Trapero et al., 2013). A key piece of information that can have dramatic effects on the supply chain is a forthcoming promotion. All forecasters in the supply chain have to take this into account. New models have been developed which show that exponential smoothing models can be extended to include multivariate promotion effects resulting in better forecasts than those made within the organization that incorporate management expertise. The improvements achieved in forecasting accuracy are substantial, up to 35% as measured by MAPE.

The latest work to come from the Centre is concerned with new methods of forecasting retail SKU sales to take into account competitive promotional effects. Because of the complexity of market information available, new statistical methods are required to overcome the problem of big over-parameterized models (Hyang et al., 2013).

The second type of information that can help the manufacturer is from the downstream retailer who can supply many pieces of information that might help a supplier. This includes the retailer’s own forecasts, the stock position and critically, EPOS sales data. Trapero et al., (2012) have examined the benefits of using collaborative data supplied from the retailers and found them to be substantial. Improvement in forecasting accuracy exceeded 25% as measured by MAPE.

Further work is being carried out by one of the Centre’s doctoral students, Weller, who has surveyed 200+ forecasters who are involved at different levels with using collaborative information. The aim of the research is to gain a rich understanding of the data that may be shared and to design optimal forecasting schemes to meet these different data conditions. Weller (2012) presents preliminary details describing the survey results.

Another area of research that the centre has been investigating is the modelling and forecasting of slow moving items, often appearing in spare parts management. Nikolopoulos et al. (2011) proposed ADIDA, an aggregation methodology for accurately predicting such intermittent demand time series, while Kourentzes (2013) explored the use of neural networks for forecasting such time series, finding superior inventory performance. 

A full list of publications can be found here.