John E. Boylan - Intermittent Demand Forecasting

Make to stock . In an MTS environment, customers are willing to wait no more than the time it takes to deliver the particular item to them and so the item needs to be available in stock, ready to be dispatched, or, in the case of retailing, it needs to be available on the shelf. In this case, demand is not known and needs to be predicted. The alternative environment is known as make to order (MTO), where the products are not assumed to be in stock, and the customer must wait until the manufacturer assembles the product for them. In this case, customer demand is known and does not need to be predicted. This situation is common for some products (e.g. furniture) but not for others (e.g. automotive or aerospace spare parts). There is also a move to 3D printing of products in some industries, which is a form of MTO but with shorter delays ( Technical Note 1.1).

Single stock keeping unit (SKU) approaches . We are looking at forecasting the requirements (and managing the inventories) of single SKUs. Although some of the methods to be discussed in this book rely upon collective considerations (across a group of SKUs), the rest of the material considers single SKU problems. This is because higher levels of aggregation are, typically, not associated with intermittent demand. Consider, for example, 10 intermittent demand items, all of which are replenished from the same supplier. It makes sense to consider the aggregate demand of those items to facilitate efficient transportation arrangements. However, although demand at the individual SKU level may be intermittent, aggregate demand (across all 10 SKUs), most probably, will not be intermittent.

Single stocking location approaches . We focus on determining inventory replenishment requirements at each single location, without taking into account interactions between locations. As such, we do not consider the possibility of satisfying demand by lateral transshipments of stocks between stores. This is because these decisions relate explicitly to joint inventory‐transportation optimisation, which is beyond the scope of this book. Further, and as discussed above, aggregate demand (across different locations in this case) is typically not associated with intermittence.

We should also mention that, although the term ‘demand’ is being used in this book when referring to forecasting, demand will not always be known and, in this case, actual sales must be used as a proxy. The terms ‘demand’ and ‘sales’ are used interchangeably in this book although, strictly speaking, the latter is often used as an approximation for the former.

This book starts by contextualising intermittent demand forecasting in the wider scholarship and practice of inventory management. We begin in Chapter 2with a discussion of inventory management and some of its implications for forecasting. Then, in Chapter 3, we examine the service drivers of inventory performance. The focus shifts in Chapters 4and 5to the characterisation of intermittent demand patterns by demand distributions. This forms a natural foundation for the next two chapters, which focus on forecasting methods. Chapter 8takes us back to inventory replenishment and the linkage between forecasting and inventory control. In the next chapter, we move on to the measurement of forecasting accuracy and inventory performance. Forecasting accuracy assessment is a notoriously difficult problem for intermittent series, and the chapter highlights the traps for the unwary and gives some pointers to good practice.

Although the main emphasis of this book is on forecasting, classification methods are also important in practical applications. In Chapter 10, we lay some of the groundwork for classification methods, discussed in Chapter 11, which have been designed specifically to address intermittence. In the next chapter, we turn our attention to obsolescence and forecasting methods that are particularly suited to this stage of the life cycle. Chapter 13presents an alternative perspective on demand forecasting, concentrating on methods that do not assume any particular form of demand distribution. By contrast, Chapter 14delves more deeply into methods that are based on demand distributions. The book closes with Chapter 15, which contains a discussion of software solutions for intermittent demand forecasting.

Recent IT developments have greatly expanded the areas of application of intelligent intermittent demand forecasting methods. Data at a very low level of granularity have become available, which means that environments where traditionally intermittence would not be a problem now become natural candidates for further consideration. Take the retailing sector as an example: this is a traditionally fast demand environment where even the slower moving items sell in considerable volumes every day, making intermittent demand forecasting redundant. However, the current availability and utilisation of data for replenishment purposes, as often as three times per day, means that more items have intermittent demand. Although daily demand may not be intermittent, half‐daily demand could be, and demand over a third of a day most probably will be.

Another factor in retail, highlighted by Boylan (2018), is the broadening of product ranges in larger retail outlets, with grocery stores introducing more clothing lines, for example. These items will often be slower moving than staple food ranges, thereby increasing the proportion of intermittent items. Recent discussions with major supermarkets in the United Kingdom such as Sainsbury's and Tesco indicate that intermittent demand forecasting has become one of their major problems.

Intermittent series occur in many other settings. For example, the planning of inventories for emergency relief must address highly intermittent and lumpy demand. Indeed, the benefits of good forecasting and planning (for any type of series) apply just as much to charitable and not‐for‐profit organisations as they do for profit‐making wholesalers and retailers. Support for the promotion and realisation of these wider benefits is being offered at the time of writing by the ‘Forecasting for Social Good’ ( www.f4sg.org) and ‘Democratising Forecasting’ initiatives launched in 2018 by Dr Bahman Rostami‐Tabar.

Nikolopoulos (2021) made a strong case for the use of intermittent forecasting methods for series that are not intermittent but have sporadic peaks. These time series can be decomposed into two subseries: a baseline series and one containing more extreme values. Standard time series or causal methods can be used for the former. The latter include rare but expected events (‘grey swans’) and truly unexpected special events (‘black swans’) (Taleb 2007) and can be addressed using intermittent forecasting methods, at least as a benchmark against which other methods may be compared. Methods to address intermittence have their origins in inventory planning, but Nikolopoulos (2021) argued that these forecasting methods can be used more widely in business, finance, and economics or, indeed, in any other discipline. This line of enquiry will not be pursued in this book, although it seems a very promising direction for future research.

The material presented in this book reflects the authors' emphasis on robust solutions that may perform well under a wide range of differing conditions. We present the state of the art in intermittent demand forecasting, paying particular attention to the interface between forecasting and stock control. There is a very considerable market for the application of those methods, and this can only expand as more highly granular data become available.