Using the Grouping and Shrinkage Approaches to Forecasting Subaggregate Level Seasonal Demand

Forecasts are needed at different levels. At the SKU level, short-term demand forecasts for each item are needed as a part of production and inventory systems. In most situations, demand is stochastic and that is when accurate demand forecasting is crucial. Forecasts and forecasting errors are used as inputs to various inventory control systems to either attain a target service-level at minimum cost or to minimise overall costs including components such as stock holding costs, stock out costs and transportation costs. These forecasts can also provide useful information for aggregate-level forecasting and planning, and impact marketing and financial plans of the organisations.Seasonality is a pertinent issue at the SKU level. Traditionally, seasonality is estimated from an item's own data history. This approach is not always satisfactory if the data is noisy and the length of data history is short. Two alternatives have been proposed in the literature: grouping and shrinkage. Previous research indicated that the grouping and shrinkage approaches are promising in terms of improving forecasting accuracy. In research conducted with John Boylan we developed theoretical rules to choose the best method from the individual approach and the grouping approach. However, a very important issue is yet to be addressed: how to put items into seasonally homogenous groups. Previous researchers either relied on company definition or statistical clustering. These methods are reasonable but are not optimal. Miller and Williams (2003) compared the individual approach with two shrinkage methods. They also established a set of rules to choose the best out of the three. We compared the two sets of rules empirically and found they were competitive to each other. A further step would be to combine the strengths of the two approaches to achieve more improvements.Although a very important research area, this is still generally under researched due to its complexity. This proposed research aims to narrow the gap by developing theoretically coherent and applicable rules to combine the individual, grouping and shrinkage approaches. The need for such a research project can be justified from both an academic and practitioner perspective.In this research, we aim to establish a grouping mechanism to divide series into seasonally homogeneous groups, develop a theortical rule to compare the individual, grouping and shrinkage approaches together, and to analyse the implication of forecasting improvement on stock control.Our methodology is deductive in the sense that universally applicable rules are sought to be developed. However, due to the complex nature of the problem under investigation, the research strategy cannot be purely deductive. Established theory is to be applied to both theoretically generated and empirical data with the objective of identifying issues that will be subsequently incorporated and reflected back to the theory. This will go through several loops before the final conclusions/recommendations will be reached. In these respects, the research strategy can be best characterised as semi-deductive or a very well-framed simulation-intensive exploratory investigation.This research problem is complex by nature. However, a good theoretical understanding, linked with fully operational rules will have a significant academic contribution and commercial benefit. Two companies have agreed to collaborate on this project by providing data and relevant information. We expect to see such solutions improve forecasting accuracy. We also strive to examine the implications on inventory control policies, as in reality forecasting and inventory control are well interacted. Therefore, for both forecasting and inventory control purposes, this is a timely and novel project.