Forecasting for remanufacturing: The effects of serialization

• Date: 01 July 2019
• Published date: 01 July 2019
• Author: Aris A. Syntetos,Thanos E. Goltsos,Erwin van der Laan
• DOI: http://doi.org/10.1002/joom.1031

RESEARCH ARTICLE

Forecasting for remanufacturing: The effects of serialization

Thanos E. Goltsos

1,2

| Aris A. Syntetos

1

| Erwin van der Laan

3

1

Panalpina Centre for Manufacturing and

Logistics Research, Cardiff Business School,

Cardiff University, Aberconway Building,

Colum Drive, Cardiff, UK

2

Advanced Sustainable Manufacturing

Technologies (ASTUTE2020), Cardiff

Business School, Cardiff University,

Aberconway Building, Colum Drive,

Cardiff, UK

3

Department of Technology and Operations

Management, Rotterdam School of

Management, Erasmus University,

Burgemeester Oudlaan 50, Rotterdam,

Netherlands

Correspondence

Aris A. Syntetos, Panalpina Centre for

Manufacturing and Logistics Research,

Cardiff Business School, Cardiff University,

Aberconway Building, Colum Drive, Cardiff

CF10 3EU, UK.

Email: syntetosa@cardiff.ac.uk

Funding information

European Regional Development Fund;

Innovate UK;Engineering and Physical

Sciences Research Council, Grant/Award

Number: KTP 10171

Handling Editor: Alan Mackelprang

ABSTRACT

Remanufacturing operations rely upon accurate forecasts of demand and returned

items. Return timing and quantity forecasts help estimate net demand (demand

minus returns) requirements. Based on a unique data set of serialized transactional

issues and returns from the Excelitas Group and one of their defense contractors,

Qioptiq, we assess the empirical performance of some key methods in the area of

returns forecasting. We extend their application (for net demand forecasting), by

considering that demand is also subject to uncertainty and thus needs to be forecast.

Information on remanufacturing costs allows for an evaluation of the inventory

implications of such forecasts under various settings. A foray into the literature on

information technologies enables a discussion on the interface between information

availability and forecast accuracy and utility. We find that serialization accounts

for considerable forecast accuracy benefits, and that the accuracy of demand fore-

casts is as important as that of returns. Further, we show how the combined returns

and demand forecast uncertainty affects the inventory performance. Finally, we

identify opportunities for further improvements for the operations of Qioptiq, and

for remanufacturing operations in general.

KEYWORDS

empirical data, forecasting, remanufacturing, simulation

1|INTRODUCTION AND

MOTIVATION

In the product recovery hierarchy, remanufacturing is placed

toward the top tier, as it encourages the reuse of both parts

and products with minimal additional input of raw materials,

thus playing a prominent role in the circular economy.

It may be described as “the transformation of used products

(referred as cores), consisting of components and parts, into

products that satisfy exactly the same quality and other stan-

dards as new products”(Guide Jr & Jayaraman, 2000,

p. 3780). Its value is estimated at $43 billion in the United

States (USITC, 2012) and EURO 30 billion in the European

Union (ERN, 2015). While many operations are similar to

those of traditional manufacturers, remanufacturers have to deal

with extra uncertainties stemming from the core acquisition

and transformation process (Goltsos et al., 2018; Guide Jr &

Jayaraman, 2000).

In traditional manufacturing, the emphasis is on forecast-

ing the (independent) demand timing and quantity for final

products, and then “exploding”such forecasts into (depen-

dent) demand requirements for materials and parts through a

predetermined/static bill of materials, but also on a set of

largely constant operations, through a similarly static bill of

operations (BOO).

In remanufacturing however the emphasis shifts to fore-

casting net demand, which is the difference between demand

and returns (i.e., that part of the demand that cannot be

Received: 4 January 2018 Revised: 18 April 2019 Accepted: 19 April 2019

DOI: 10.1002/joom.1031

J Oper Manag. 2019;65:447–467. wileyonlinelibrary.com/journal/joom © 2019 Association for Supply Chain Management, Inc. 447

satisfied through the remanufacturing of returned cores). Net

demand then drives replenishment for new items (from origi-

nal equipment manufacturers, OEMs), although it can also

be met by cores (from core brokers, that still need to be

remanufactured), that together with the remanufactured

returns ultimately service total (customer) demand. That is,

although demand forecasting remains equally relevant, the

timing and quantity

1

of the returns also needs to be forecast,

compounding to a dual-source uncertainty.

Unlike demand, the stream of returned items is not an

independent random variable, but rather correlates with vari-

ables such as past sales (Goh & Varaprasad, 1986), the

installed base (number of items in the market), lead time sales

and returns (Kelle & Silver, 1989a), and at times the stage in

a product's life cycle (van der Laan & Salomon, 1997). It is

this relationship of returns to past sales that is integral to

returns forecasting. This means that univariate approaches,

such as those traditionally used in time series demand fore-

casting, are susceptible to underperformance (Kiesmüller &

van der Laan, 2001). It also means that the performance of

the returns forecasting is subject to the level of detail and

accuracy of information available about this relationship.

In addition, products are of unknown condition (quality)

when eventually returned, which is an important issue in its

own right (Guide Jr & Jayaraman, 2000). This implies that the

probability of returns needs to be used in conjunction with the

probability of cores being remanufacturable, or else the assump-

tion that returned cores can be used toward satisfying demand

does not hold. It is the probability of a remanufacturable return

that is of interest to remanufacturing operations. Any quality

induced variability in the remanufacturing lead times, while

being problematic, opens up scheduling opportunities for dis-

tributing operations. For example, a planner would like to know

which of the returns over the (forthcoming) lead time can be

used (remanufactured in time) toward satisfying lead time

demand. A first step toward that is to move away from aggre-

gate and into per period lead time forecasts, as we later do.

In is evident then, that in the context of remanufacturing

operations forecasting entails the simultaneous challenges of

(see also Guide Jr & Jayaraman, 2000): (a) forecasting of timing

and quantity of demand, (b) forecasting of timing and quantity

of returns, and (c) forecasting of the quality status of the returns

(at the very minimum how many returns are useable, that is, the

number of remanufacturable returns). Some research, reviewed

in the next section, has been conducted to address the second

problem. In particular, four methods proposed by Kelle and Sil-

ver (1989a) have attracted attention both in theoretical and

empirical domains. The methods assume varying degrees of

available information regarding the relationship between

demand and returns: from no information to serialized matching

of a stock keeping unit's (SKU) individual issues and returns

(and thus the time each spent with the customer). It is this latter

level of information that has been missing in previous empirical

studies, and this constitutes the main point of departure of our

analysis from previously offered results. The availability of such

information can be achieved through a number of ways as we

will later see.

An integral part of the methods' application is the distribution

ofthetimetakenforanitemtoreturnafterbeingsoldtoacus-

tomer (time-to-return distribution). Such methods are not specific

to remanufacturing but rather they are relevant in any circular

economy or reverse supply chain context. Important as they are,

they rely on knowing with certainty the probability that an item

will eventually return, the time-to-return distribution, and the

demand rate, with quality taken into account only implicitly. That

is, the four methods of Kelle and Silver (1989a) ignore challenges

1 and (to an extent) 3 above, while challenge 2 is addressed

assuming some “perfect”information. The same is, up to a great

extent, true for follow-up work conducted on these methods. Fur-

ther, empirical evidence on their practical validity and utility is

lacking. Some notable exceptions and how these contributions

differ from our work are discussed in the next section of the arti-

cle. The focus of this article is on challenges 1 and 2. What is of

particular interest is the impact of serialization on challenge 2:

how can serialized data be conductive to characterizing the rela-

tionship between past sales and returns, and what benefits does

this translate to (in terms of accuracy and utility)?

1.1 |Contribution and organization of the

paper

We assess the empirical behavior of returns and characterize

the time-to-return distributions (never addressed before in the

literature) by means of using a unique data set of serialized

transactional issues and returns from the Excelitas Group and

one of their defense contractors, Qioptiq. The company is a

defense and aerospace integrated logistics supplier (managing

flows and repairs), operating within various statemilitary sup-

ply chains globally. The product line we are concerned with

consists of 15 electronic night vision, thermal (infrared, IR)

and image enhancing visual aid equipment for the dismounted

soldier (head/weapon-mounted, handheld).

The data set allows us to be the first to test the methods of

Kelle and Silver (1989a) in empirical terms, but also to explore

the effect of serialization. We extend the methods' application

(for net demand forecasting) by considering that demand is also

subject to uncertainty and thus needs to be forecast, and by

means of relaxing some further unrealistic assumptions of the

forecasting procedures. Information on remanufacturing costs

allows for an evaluation of the inventory implications of the

forecasts. A foray into the literature on information technolo-

gies enables a discussion on the interface between information

availability and forecast accuracy and utility. As we will see in

448 GOLTSOS ET AL.