Content uploaded by Peter Hines
Author content
All content in this area was uploaded by Peter Hines on Mar 28, 2019
Content may be subject to copyright.
Towards responsive vehicle supply: a simulation-based
investigation into automotive scheduling systems
Matthias Holweg
a,
*, Stephen M. Disney
b,1
, Peter Hines
c,2
, Mohamed M. Naim
b,3
a
Judge Institute of Management, University of Cambridge, Trumpington Street, Cambridge CB2 1AG, UK
b
Logistics Systems Dynamics Group, Cardiff Business School, Cardiff University, Colum Drive, Cardiff CF10 3EU, Wales, UK
c
Lean Enterprise Research Centre, Cardiff Business School, Cardiff University, Colum Drive, Cardiff CF10 3EU, Wales, UK
Available online 8 December 2004
Abstract
Vehicle supply has traditionally been based on forecast-driven production, and a large fraction of cars has been sold from
stock—a practice which incurs considerable cost in terms of stock holding and sales incentives. Derived from successes in other
industries, the benefits of responsive supply systems capable of providing customized vehicles in short lead-times have been
pointed out. While the theoretical discussion of such ‘build-to-order’ (BTO) strategies is well advanced, the dynamic feasibility
of implementing these concepts is far from understood. Using a simulation of a multi-tier supply chain-system, this paper
investigates the impact of altering key aspects of the scheduling activities with the objective of determining the scope for
potential improvements in responsiveness of the supply chain. The simulation results show that current vehicle supply systems
are not capable of supporting BTO due to insufficient feedback between supply and demand, as well as due to the strong reliance
on forecasting in the scheduling process. The paper concludes with a set of recommendations on how to improve current
scheduling systems towards increasing the content of vehicles built to customer order.
#2004 Elsevier B.V. All rights reserved.
Keywords: Build-to-order; Scheduling; Taguchi method; Systems modeling
1. The case for responsive vehicle supply
Much like in the days of Henry Ford, vehicle supply
systems across key markets are still predominantly
driven by forecasts, rather than customer orders, and
the majority of vehicles are sold from existing finished
goods inventory in the market place. Stock levels of
finished vehicles in the market place in the US and
Europe range from 2 to 3 months, which incurs the
obvious cost for stock holding, incentives used to sell
overproduced models and discounts needed to
persuade customers to accept compromises on their
specification (Holweg and Pil, 2001, 2004). Overall,
with nearly 3 million new vehicles stored at US
dealerships at any one time, the cost of not being able
to align vehicle production to customer demand has
www.elsevier.com/locate/dsw
Journal of Operations Management 23 (2005) 507–530
* Corresponding author. Tel.: +44 1223 760 583;
fax: +44 1223 339 701.
E-mail addresses: m.holweg@jims.cam.ac.uk (M. Holweg),
disneysm@cardiff.ac.uk (S.M. Disney), hinespa@cardiff.ac.uk
(P. Hines), naimmm@cardiff.ac.uk (M.M. Naim).
1
Tel.: +44 29 2087 6310.
2
Tel.: +44 29 2087 6005.
3
Tel.: +44 29 2087 4635.
0272-6963/$ – see front matter #2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.jom.2004.10.009
been estimated at $600–1500 per vehicle (Lapidus,
2000). In Europe, one-off savings through a reduction
of the finished vehicle levels are estimated to be
around s6.5 billion (ICDP, 2000).
For those customers who are willing to wait, the
average order-to-delivery (OTD) lead-times in Europe
are 48 days for a volume car to be built and delivered
to order. For most Japanese models built in Europe, the
time is even higher at 63 days, and for European built
specialist and luxury vehicles the lead-time is 43 days
(Williams, 1999). These long lead times is the major
factor in discouraging a higher percentage of vehicles
being built-to-order. Research in the UK has shown
that 61% of customers want their vehicle to be
delivered within 14 days or less (Elias, 2002). Even in
the US, where vehicles traditionally are sold from
dealer stock, 74% of consumers would rather wait and
order the vehicle instead of buying one from the
dealer’s lot that is incorrectly equipped, but the
majority of North American consumers would wait no
more than 3 weeks to receive their custom-built
vehicle (Holweg and Pil, 2004).
However, recent pronouncements by vehicle
manufacturers have made it clear that there is some
recognition of these failures. Manufacturers like
Volvo and Renault have already started ambitious
build-to-order projects in order to reduce stock
levels and costly incentives. The best known of these
is the ‘Project Nouvelle Distribution’by Renault,
aiming at producing a built to order vehicle in 14
days from order to delivery. Volkswagen and Ford
both have similar 14 or 15-day car programs, BMW
are even attempting a 10-day order-to-delivery
(OTD) lead-time. Volvo was probably the first
adopter of this model with a plan to reduce OTD
times from 6 weeks to 28 days in the early 1990s,
and down to 14 days in 1995—whilst targeting a
100% ‘customer-based’production (Hertz et al.,
2001). These lead times refer to products ‘built to
customer order’; that is where the customer order
initiates the build of the vehicle, and not a forecast.
These lead times should not be confused with ‘order
amendment’lead times, whereby peripheral options
of the vehicle can be changed up to few days prior to
production. As Holweg and Pil (2001) point out,
order amendment has the same structural problem as
make-to-forecast (MTF), because unsold finished
goods inventory still occurs if no customer can be
found in time to amend the existing orders in the
system.
The notion of demand-, rather than forecast-driven
production is not at all new. When lean production was
set out more than a decade ago, it had the premise ‘to
build cars at the rate the customers demand it (sic)’
(Monden, 1983). Monden stressed that the overall
objective of the Toyota production system, the
ancestor of the lean philosophy (Womack et al.,
1990; Womack and Jones, 1996), was to only build
cars to order and hence avoid what refers to as the
‘worst waste of all’: overproduction (Ohno, 1988). A
decade later however, despite continuing improve-
ments in productivity at factory level (Pil, 2002;
Holweg and Pil, 2004), few volume vehicle manu-
facturers are able to build to customer order, and only
one builds solely to customer order.
From a static point of view, building vehicles to
order as opposed to a forecast has an unmistak-
able logic for both manufacturers and customers:
customers get the exact product they asked for, and
manufacturers can operate without the costly
inventories and incentives. Such a drastic change
in supply chain strategy however would have wide
ramifications for all players in the supply chain, and
potentially incurs additional cost in the manufactur-
ing process. As Raturi et al. (1990) point out, the key
benefit of the make-to-forecast strategy is that
factories can be decoupled from demand volatility
in the market and use stability to produce more
economically. In a build-to-order system, the need to
respond to changes in demand becomes paramount,
which requires the ability to alter production
volumes according to market demand. Indeed, our
interviews with vehicle manufacturer staff revealed
serious concerns about the implications of demand
swings on the capacity utilization at the assembly
plant level, and possibly even at the various supplier
tiers.
In this paper, we thus set out to model the
dynamic aspects of changing vehicle supply from a
forecast- to an order-driven strategy. In particular, we
address the question whether current vehicle supply
systems ‘structurally’are capable of supporting more
responsive supply chain strategies. We further extend
our investigation into a sensitivity analysis of the
system, in order to identify the areas of ‘maxi-
mum leverage’where current system features pose
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530508
potential inhibitors to implementing a build-to-order
strategy.
2. Simulating the information flows in the auto
supply chain
2.1. Research environment and approach
We present a model of the information flows within
an automotive supply chain from customer order,
through the vehicle manufacturer, to the first
component supplier. The model is based on a synthesis
of the findings from two major automotive research
projects, the 3DayCar Programme
1
and the Supply
Chain 2001+ Project
2
, allowing the holistic modeling
of a total supply chain from customer order through to
the component supply chain. In addition to previous
studies into the automotive supply chain, which have
focused on subsystems of the supply chain, i.e.
performance of the assembly plant (Krafcik, 1988;
Womack et al., 1990; Pil and MacDuffie, 1996)or
improvements of the link between supplier and
manufacturer (Helper, 1991; Cusumano and Takeishi,
1991; Hines, 1998) or aftermarket distribution (cf.
Hammant et al., 1999), the 3DayCar Programme had
the objective of investigating the effectiveness of the
entire supply chain from customer to delivery of the
completed vehicle.
Within both the 3DayCar and the Supply Chain
2001+ projects, system analysis methods and techni-
ques were used to analyze and understand the
processes and flows that determine the structure of
the production systems under study. While the
3DayCar concentrated on the vehicle manufacturers’
echelon in the supply chain, Supply Chain 2001+
focused on first-tier suppliers. The outputs from both
programs have been exploited and the system
structures of the vehicle manufacturers and first-tier
supplier have been constructed in causal loop diagram
form. Causal loop diagrams are a common tool
utilized in system dynamics modeling to define flows,
feedback loops and system structures (e.g. see
Checkland and Scholes (1990);Sterman (2000)).
In our case we define the system structures at each
echelon in terms of information flows and the ordering
decision rules as a prerequisite for formulating
dynamic models of the supply chain under study.
This follows the procedure defined by Evans et al.
(1998) for difference equation simulation modeling,
by which the causal loop diagram is translated into
block diagram to aid in the development of difference
equations that represent supply chain dynamics. As
with Evans et al. (1998), we implement the difference
equations using a proprietary spreadsheet package.
We extend the method further by the incorporation
of the Taguchi method, a succinct description of which
may be found in Roy (1990). For the purpose of this
paper we do not use the Taguchi technique in its full in
optimizing the supply chain studied. Instead, we limit
its application to undertaking a sensitivity analysis of
the key performance indices that characterize the
dynamic performance of the supply chain. The
objective is to provide guidance for managers to
judge the implications of various strategies, such as
true build-to-order versus current ordering policies.
The objective of the 3DayCar Programme was to
define a responsive production control system to
enable short order-to-delivery (OTD) times. In this
context, extensive process mapping has been con-
ducted, using both high-level process maps (Rother
and Shook, 1998) and detailed value stream mapping
(Hines and Rich, 1997). Major conclusions from this
work are that current order fulfillment (OF) systems
are incapable of satisfying customers within their
waiting tolerance, hence vehicle manufacturers are
forced to supply vehicles from stock (Holweg and Pil,
2001, 2004). Furthermore it was found that current
vehicle scheduling systems are heavily constrained by
long order throughput times and a high degree of
variability for components that is passed onto the
suppliers (Holweg, 2003).
Fig. 1 shows the structure of the current OF pro-
cess in the auto industry. This process map is an
aggregate of the six European-based vehicle manu-
facturers that were analyzed—two European-owned,
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 509
1
The 3DayCar Programme was a 3-year research project, which
aimed at developing a framework in which vehicles can be custom-
built and delivered within a minimal lead-time. A consortium of
vehicle manufacturers, suppliers, distribution and logistics compa-
nies and IT service providers funded the project. See also
www.3daycar.com for more information.
2
The Supply Chain 2001+ project, developed a set of decision
support systems, tools and techniques (e.g. Hammant et al. (1999)
and Naim et al. (2002a)), coupled with diagnostic methods (Naim et
al., 2002b) for designing the right supply chain for a particular
product.
two US-owned, and two transplant operations of
Japanese manufacturers. For more detail on the
outcomes of the process mapping see also Holweg
(2003).
As can be seen, the current system is heavily geared
towards aggregating a sales forecast through dealers
and National Sales Companies (NSCs), which then
becomes a crucial input into the monthly production
program. The second key input into the production
program are production capacities and potential
constraints in component availability. This program
defines the vehicle build in terms of volumes, engines
and sometimes also on equipment level. The planning
and scheduling targets are commonly decided 2–3
months prior to the actual build period. The customer
orders, once received, are transferred from the dealer
to a central Order Bank, where they have to be fitted
around the rigid framework defined in the production
program. From the order bank, the orders are fitted
into production schedules for the various assembly
plants, a process that generally happens daily, or
several times per week. Once the orders have been
assigned to their estimated build dates (EBD), the
production schedule is sent out to all stakeholders as a
forecast, including the component suppliers. The
orders remain, on average, 15 days in this holding
position to allow for planning processes, before they
are rescheduled and assigned to a slot in the actual
production sequence. Only at this point are the orders
firmly assigned to their actual build date, and suppliers
receive their firm component requirement notifica-
tions (call-offs). The main delays in the scheduling
process are meant to give supplier and logistics
operations the ability to plan, but as the information
tends to be too unreliable, interviews with suppliers
have shown that this information is actually of little
operational use (Holweg and Pil, 2004). Further delays
are caused by legacy IT systems, which require over-
night updates to transfer information from one system
to the next and can account for up to 5 days of the
overall delay.
Overall, the current system leaves little flexibility
for responding to an actual demand signal, which is
less problematic in a forecast-driven system, as order-
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530510
Fig. 1. Generic order fulfillment process map (simplified). Adapted from Holweg (2003).
to-delivery lead-times are not an issue, since vehicles
are sold from stock. In a build-to-order system
however, the actual lead-time to fulfill an order is a
crucial enabler, as the majority of customers are
demanding delivery within 2–3 weeks (Williams,
1999; Elias, 2002).
The actual ability to respond to an order reflects this
inflexibility. The OTD benchmark, based on the
critical path methodology suggested by Lockyer and
Gordon (1991), and with specific reference to the
order fulfillment process by Shapiro et al. (1992) and
Stalk (1988), shows an average capability of 40.1 days
(Holweg and Pil, 2001). In other words, the average
current automotive supply chain cannot provide a
custom-built vehicle in less than 40 days. Any
capacity constraints or quality problems are likely
to add further delays, which relates to the 48 days
average OTD lead-time for volume cars in Europe.
In terms of time delays, 85% of this time is lost in
the information flow (see Fig. 2); a total of nearly 70%
is accounted for by the order scheduling and
sequencing subsystems alone. Considering also that
much of the attention in automotive research has been
focused on the manufacturing operation, which only
accounts for 4% of total order fulfillment time, it is
very obvious that a more holistic perspective is
needed.
Compared to successful build-to-order systems in
other sectors, such as the Dell approach of assembling
personal computers to order, for example, it is
apparent that the current system is not laid out to
support build-to-order. Dell is able to transfer a
customer order to the factory within a day (the auto
industry takes a month), to build the computer ordered
within a day, and to ship the product to the customer
within 5–7 days after the order has been placed. The
crucial difference is obviously that the Dell system has
been designed to support BTO, whereas the auto-
motive system has the long heritage of a forecast-
driven approach. Furthermore, a motor vehicle
contains an average of 2000–4000 components,
whereas a computer is assembled from 15–50
components only (Holweg and Pil, 2001). Thus,
although the benefits of a BTO system even in the
automotive industry might be obvious, the question is
‘‘what changes to the current system would enable a
higher degree of vehicles to be built to order?’’
According to Simon (1962), there are two elements
to complexity: complexity of the structure, and
complexity of the dynamics of a system. The
structural characteristics of the system are known,
and we also know that customers are not willing to
wait as long as it currently takes to build a car to order.
Yet although we have the structural parameters, we do
not know how the system behaves dynamically. For
example, will reducing the lead-times involved in the
scheduling process already be sufficient to convert
the system towards BTO? As we know from the
System Dynamics field, structure drives behavior
(Sterman, 2000), which raises the question whether
one needs to change the structure of the scheduling
systems also.
This paper hence uses a simulation model to
investigate the consequences of switching from a
make-to-forecast approach to a build-to-order system,
by subjecting the current scheduling system to a range
of different settings of its scheduling parameters.
2.2. Modeling a multi-tier supply chain system
Our model is based on generalized order fulfillment
process maps from the six major vehicle manufac-
turers and global component suppliers, which were
generated as part of the systems research within the
3DayCar Programme, and findings of the Supply
Chain 2001+ project. Based on the system structure, as
found in the process mapping research, we used the
systems modeling and simulation methodology
suggested by Evans et al. (1998), coupled with a
factor sensitivity analysis (see Disney and Naim,
1999) to understand its dynamic performance. Such a
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 511
Fig. 2. Time delays in the order fulfillment process. Source: Holweg
and Pil (2001).
simulation approach, coupled with optimization
algorithms, has been utilized previously to optimize
logistics control systems (Disney et al., 1996).
The model considers the information flow from
the initial customer orders for two vehicles, models
A and B, right down through the supply chain in
order to obtain a fundamental understanding of the
underlying structure of the production scheduling
and supplier ordering systems. From this generalized
structure, insights into its dynamic performance and
opportunities for improvement are presented. A
sensitivity analysis is conducted on the model to
highlight the areas of ‘maximum leverage’within
the supply chain to reduce dynamic variability in the
supply chain and inventory levels and improve
customer service levels in the current decision-
making structure.
The simulation model reflects the actual system
structure and time delays in the system and evaluates
the sensitivity of the different factors in the production
system on key performance indicators. The supply
chain system is characterized by a multiple product
(here simplified to product A and B) production
system with market place demand variation. The
demand signal for model A is normally distributed
(m= 350, s= 82); model B is also independently and
identically normally distributed yet shows a higher
demand (m= 700, s= 107). After 6 months, the
demand for model A augments by 20% (m= 420,
s= 82), so still well within the boundaries of the
available capacity (1300 units/day, which represents
the output of an average large-scale facility, see also
Pil (2002)). Such a step change in demand has been
used previously in analyzing supply chains in order to
show a system’s response to a change in the dynamics
of the input (demand) signal. The most famous
examples include Forrester’s distribution chain, and
the ‘MIT Beer Game’(Forrester, 1961; Sterman,
1989).
Fig. 3 shows a causal loop diagram of the
simulation model. The complexity of the scheduling
system is a result of the overall complexity of the
operation, which has to cope with up to 4000
components per vehicle and up to 100 markets into
which the vehicles are supplied, with up to several
million specification variations (Pil and Holweg,
2004). Unsurprisingly, the simulation model shown
cannot entirely reflect this complexity. Our focus was
to reproduce the structural subsystem of the schedul-
ing operations in order to analyze the dynamic
behavior derived from it. Nevertheless, we had to
accept simplifications in order to create a feasible
model for our analysis. So, despite validation of the
model in focus groups with our industrial sponsors, we
would argue that the results obtained should be
regarded as relative indices, rather than absolute
values. In other words, we argue for validity of the
relative behavior of the 18 scenarios analyzed, rather
than for accurate absolute outcome values of the
performance indicators (such as production adaptation
costs, for example). The cost data required to make
this final step was regarded as highly sensitive by both
the manufacturers and suppliers involved in our
research, and therefore not available to convert indices
to absolute values.
Fig. 3 is a simplification of a complex non-linear
model. Even without the Taguchi analysis the
simulation model has over 30 parameters and over
1600 lines of code. Further detailed description of the
model and its features is given in Appendix A.Fig. 4
below shows a sample output from the model for a
standard simulation run of 2 years (including a 3
months initiation phase). Unstable schedules were
limited in growth by overall factory capacity and by
prohibiting negative production rates.
2.3. Measuring system performance
The performance of the scheduling system is
quantified through six key performance indicators
(KPIs). These KPIs are further weighted to ensure
that their respective impact corresponds to the value-
added content in the supply chain. For example, the
component costs reflects 75% of the vehicle cost,
which is the equivalent to a 25% value-added
content of the manufacturer (cf. Wells and Nieu-
wenhuis, 2001). The vehicle cost at the finished
goods inventory stage is weighted at 100%. The
overall performance of the system thus is jointly
characterized by the six KPIs together to provide an
overall score for the performance of the scheduling
system in each scenario (for more detail, and
complete results see Appendix B). It was not
possible to assemble a single outcome measure, as
for this actual cost data would have been required.
Such data however is regarded as highly sensitive by
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530512
the manufacturers, and was therefore not available.
Instead, the relative behavior of the indices is
discussed as proxy for the system’s performance.
At the vehicle manufacturer level the key
performance indicators are:
InvA and InvB: the combined cost of the inventory at
the vehicle manufacturer (inventory of finished
vehicles for models A and B, as a measure of
alignment of production to demand) and the cost for
lost sales (backlog). This variable is defined as the sum
of 20% of the absolute inventory deviation of finished
cars, for both models A and B. This is to reflect equally
the cost of holding inventory and lost sales, which are
weighted equally and combined in this variable. The
cost of lost sales, whereby the customer cannot be
supplied with the desired vehicle is also considered in
these inventory indices, is directly derived from the
methodology applied by the International Car Dis-
tribution Programme, cf. Kiff, 1997; Williams, 2000).
The cost of being out of stock, or the ‘lost sales’, are
accounted for as an opportunity cost here, or as lost
revenue. In the model, this cost marks a 20% a portion
of the finished vehicle value, thus even in a constant
backlog situation the results show cost of inventory.
Combining the two features, stock holding cost and
lost sales, has the advantage of reflecting the trade-off
between stock holding and customer service without
any distortion.
PrAdaptA and PrAdaptB: Production adaptation cost
at the vehicle manufacturer occur for both models A
and B, since the system is subjected to highly
variable demands signal for both models. This
adaptation, i.e. altering capacity levels and the
production mix, represents the cost of flexibility of
the system, measured through changes in production
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 513
Fig. 3. Causal loop diagram of the system studied.
levels in order to fit the demand pattern. The
adaptation costs are measured through the variance
of production rate at the VM for each model. It is
true that car plants have been able to reduce change-
over lead-times for robots in the welding operation,
and hardly any penalty for a change-over occurs in
assembly. However, the changing of fixtures and jigs
in body shop is still an issue, as is paint change in
paint shop, and so are the increased complexities in
logistics and materials handling issues. Furthermore,
akeybenefit of the make-to-forecast approach is the
stability of production, which enables cost reduc-
tion as the production system is protected from
variability (Raturi et al., 1990). Thus, there is a
penalty for variety to pay, which is modeled in this
variable.
At the first-tier supplier level the key performance
indicators are:
InvSup: Cost of inventory at the supplier (finished
component and raw material stock, as measure of
alignment of component production to vehicle
production). The supplier inventory cost is defined
as the sum of 20% of the absolute inventory deviation
of raw material (see above).
PrAdaptSup: Production adaptation cost equally
occurs at the supplier, when component production
schedules need to be altered to accommodate
changing demand from the vehicle manufacturer.
The cost is defined as the standard deviation of the
supplier scheduling rate, a formulation that is finding
common currency as a measure of dynamic behavior
(Disney and Grubbstrom, 2004).
Furthermore, as an additional overall measure of
alignment of demand and supply, the correlation
coefficient rof the demand vehicle supply pattern is
measured for each of the two models, CorrA and
CorrB, respectively. The better the alignment, the
more capable of supporting BTO a system is. This
correlation factor hence marks a simple measure that
gives an indicator to what percentage the supply chain
system in question can respond to the demand pattern.
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530514
Fig. 4. Simulation output (example).
3. Scenario and sensitivity analyses
The sensitivity analysis undertaken uses part of the
Taguchi method. The Taguchi method, derived from the
field of design engineering, provide an efficient strategy
for dealing with multiple and interrelated problems
using orthogonal arrays (OA), while focusing on cost as
a key consideration (Roy, 1990). Whereas in traditional
experiments only one variable of the experiment was
changed at a time, the strength of the Taguchi technique
is that one can change many variables at the same time
and still retain control of the experiment. In this study,
the method isonly used to the point where the analysis of
variances (ANOVA) is conducted and re-tested with the
main contributing factors to determine their percentage
contribution to the design. The final step, where the
design is optimized, is omitted as this was not within
the remits of the underlying research question. In its
underlyingform, the procedure can be used to determine
the sensitivity of KPIs to realistic changes in the values
of system parameters. The step-by-step application of
the Taguchi method is outlined in more detail in Disney
and Naim (1999).
3.1. Determining the effect of factors on system
performance
The sensitivity analysis considers the effect of eight
factors (F
i
) on the key performance indicators
discussed above. The output of the sensitivity analysis
reveals the contribution of each factor to the KPI, thus
providing a clear picture of the areas of ‘maximum
leverage’in the system.
The factors analyzed at the vehicle manufacturer
level are:
F
1
: Representing the period (considering the timeline
of historic demand data) used to determine the
optimum order buffer size. This factor mirrors the
average fill in the order bank, i.e. how many days of
order are held as buffer in order to ensure capacity
utilization of the assembly plants. In our interviews
with planning staff at the vehicle manufacturers, a
minimum of 5 days order buffer was commonly
quoted.
F
2
: The decision time delay between the point where a
scheduling decision is taken to the point in time when it
takes effect. This factor relates to the delay incurred, as
scheduling decisions are not implemented straight away.
For example, in the current system, changes to the
production program are made up to 3 months in
advance, and hardly ever to the immediately following
month. Thus, drastic changes to the production program
hardly ever occur with less than a delay of 1 month.
F
3
: A parameter, which determines what fraction of
the order buffer error is corrected in a planning cycle,
i.e. 1/1 (full error is considered), 1/4 (25% error is
considered), or 1/12 (8% error is considered). This
‘error gain’refers to the strength of feedback of the
backlog, or in other words, how strongly a past
backlog is considered in the current planning. When
the full error is considered, the system tries to ‘catch
up’with all unfulfilled orders at once, which can lead
to drastic swings in the schedule. A low error gain on
the other hand means that the order backlog is only
slowly recovered, and more lost sales are incurred.
F
4
: The rescheduling frequency (daily, weekly, or
monthly), i.e. for how long are the schedules actually
kept. This is one of the key variables, as it describes
how often the assembly plants are scheduled. Dell
reschedules on an hourly basis, current car plants are
scheduled weekly at best, and some even receive
monthly orders only. A short rescheduling frequency
is a main instrument of aligning demand and supply in
a BTO system. Together with F
2
, these two variables
describe how often scheduling decisions are being
taken, and after they have been taken, how long it
takes for these changes to take effect. Jointly, these
two factors describe the responsiveness of the
scheduling system in terms of adapting to new
circumstances, such as changes in demand.
F
5
: The period (considering the timeline of historic
demand data) used to generate the volume forecast and
the mix ratio between models A and B. This factor
describes the smoothing used in the forecast: the
longer the horizon, the more stable the forecast
becomes, but the less reactive it is to changes in
demand. This forecast becomes the backbone of the
capacity decision taken at the manufacturer, and hence
marks another critical input for the system.
At the first-tier supplier, the following factors were
analyzed:
F
6
: The decision time delay from the point in time
when a decision is taken to when it actually takes
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 515
effect at the first-tier supplier. Similar to the
scheduling delay at the vehicle manufacturer, this
variable considers the responsiveness of the supplier.
F
7
: The period (considering the timeline of historic
demand data) used to determine the supplier’s raw
material requirements. As at the manufacturer, this
variable describes the responsiveness of the forecast
being used in the scheduling process at the manu-
facturer.
F
8
: The safety buffer level (above average demand)
used to set the reorder point. This factor sets the buffer
stock, or slack, in the component supplier’s production
process in order to respond to unforeseen changes not
covered by the long-term forecast. Since component
availability is a crucial enabler for the vehicles to be
built, this factor allows for the investigation of the
effects of additional slack in the supply chain.
To determine the core effects on the system of each
of these eight factors, we transferred these into ort-
hogonal arrays. The Taguchi methodology allows for
the contribution of each factor to be determined with
the least number of simulations, significantly reducing
the computer time required. This way, complex sys-
tems such as the one discussed here can actually be
efficiently examined. The use of the L18 OA effect-
ively allows the full factorial design of the 4374 ex-
periments to be investigated with only 18 experiments
(Roy, 1990). Details of the Taguchi method and the
simulation results obtained are given in Appendix B.
3.2. Scenario analysis
The following section shows the outcomes of three
example scenarios tested (out of the 4374 theoretically
possible experiments analyzed), displaying their
settings and performance levels (the graph shows
the stock of finished vehicles (in the positive) and the
order backlog (in the negative), hence overall it gives
an idea on how close the actual production matches the
demand for the vehicles A and B. We have chosen
three scenarios that in our view best reflect our
findings within the remits of this research paper: the
‘current state’scenario, mimicking the status quo in
the automotive supply chain, a ‘build-to-order (BTO)’
scenario, which implements true demand-driven
scheduling and responsive capacity adjustment, and
a‘balanced scenario’, which models responsive
day-to-day scheduling close to demand, while
providing a more stable capacity planning framework.
Essentially we subject all three scenarios to a
variable demand pattern in the market place, and
measure their dynamic reaction in terms of how well
these systems are able to respond. Thus, although the
majority of orders are stock orders, and not customer
orders, we simulate a 100% customer-driven produc-
tion here by measuring how well the system can
respond to the demand pattern. In reality, prioritization
of orders might be used to mitigate some of these
effects, yet adding different order statuses was beyond
the scope of our model.
‘The current state scenario’essentially models the
current system, which is characterized by long
demand leveling and decision delays before schedule
changes can take effect. This model is a direct
consequence of the ‘Fordist’mass production logic,
using the benefits of MTF in terms of stability of
production. This is shown here in long lead-times for
change of the production schedule, and the long
forecast horizon used. These system settings are
transferred directly from our process mapping
research with the six vehicle manufacturers. Here,
the vehicle manufacturer production schedule is kept
for 1 month, before it is changed. As can be seen, the
system is unable to adjust to the change in demand
(product A has an increase of 20% in demand in month
6), and the total system oscillates and subsequently
fails to provide sufficient vehicles against the orders
for variant A, whilst still overproducing variant B (see
Fig. 5). The reasons for this effect are the slow non-
linear feedback loops and the lack of direct feedback
loops monitoring inventory levels and correcting
discrepancies. This severely inhibits the tracking
ability of the system. System settings and simulation
outcomes are shown in Table 1.
The most striking result of this simulation run is the
inability of the system to adapt to the demand swing
for model A due to the long lead-times for changing
the schedule, and implementing these scheduling
decisions. Interestingly, in the correlation analysis of
the production schedules and demand patterns, model
A shows a better result than model B, which means
that the schedule matches demand, but the system is
incapable of recovering the backlog. This is a key
reason why, in the real world, vehicle supply systems
need the 60–90 days of finished vehicle inventory.
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530516
Without this buffer, the risk of not being able to match
swings in demand would result in severe penalties in
lost sales. In our simulation we do not permit any stock
orders to be raised, as is the case in the real world, in
order to analyze the system’s ability to support BTO,
without the interference of other order fulfillment
strategies compromising the results. The ‘current
state’scenario models the effects of converting the
current MTF system to a BTO system, without altering
a single parameter in the system, and thus is the
reference scenario.
In many ways the poor performance of this scenario
is the explanation why manufacturers are not doing
BTO, because their current systems cannot cope. On
the other hand, the maximum backlog only reaches
5000 units, thus one could argue that the system could
well build to order, as 5000 units equals 4 days, which
is well within the customer waiting tolerance.
However, the dynamic response of the system to the
demand change is important here. The system fails to
reacts to the increase, and is not capable of recovering
the backlog. Thus, although the backlog seems small
in terms of daily production, the system never recovers
the backlog for product A. In case of more drastic
changes the system will fail to react, and constant
misalignment of demand and supply is the result. In
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 517
Fig. 5. Current state scenario: vehicle stocks and order backlogs (units).
Table 1
Results of the ‘current state’scenario
System settings System performance indices
F
1
Average historic demand data used to
determine target order buffer (VM)
1 month InvA VM finished vehicle inventory cost A 101
F
2
Scheduling decision delay (VM) 2 months InvB VM finished vehicle inventory cost B 31
F
3
Order buffer error weighting (VM) 1/12 PrAdaptA VM production adaptation cost A 30
F
4
Period production schedule is kept
constant for (VM)
1 month PrAdaptB VM production adaptation cost B 20
F
5
Average historic demand data used to
determine volume forecast and
mix ratio (VM)
1 month PrAdaptSup Supplier production adaptation cost 310
F
6
Scheduling decision delay (supplier) 1 week InvSup Supplier inventory cost 91
F
7
Average historic demand data used to
determine raw material requirements
(supplier)
1 month CorrA Correlation of production and demand, model A 0.1035
F
8
Safety buffer level used to set reorder
point (supplier)
4 days CorrB Correlation of production and demand, model B 0.0746
this model, this effect relates to the overall production
volume only. In reality however, there are multiple
facets to this misalignment, as products come in
various bodystyles, powertrain configurations, and
options. Hence, even if overall production volumes are
aligned to demand for a model, it might well be the
case that these all are estate cars, or all have manual
transmission, whereas customer demand may be much
stronger for sedans with automatic transmission. Thus,
the misalignment can happen on many layers below
the overall production volume, driven by the product
variety of the vehicle, which makes the response to
changes and the rapid recovery of backlogs crucial.
The question then arises, ‘‘What system features need
to be changed in order to make the schedule more
capable?’’, which leads to the following scenarios.
In a second step, we presented these findings to the
focus groups at the annual 3DayCar sponsors
conference, in order to validate the findings. The
focus group consisted of 28 senior managers from
vehicle manufacturers (8), component suppliers (1),
logistics companies (5), IT consultants (6), profes-
sional associations (4), as well as representatives of
the UK’s Department for Trade and Industry (4). The
presentation of the simulation findings was made 12
months after the results of the structural analysis had
been presented to them, so the participants were
familiar with these results already. After presenting
the results of the simulation to the group, the question
was posed to them whether the model was an accurate
representation of the current system, and whether they
felt that the quality of the dynamic response of the
‘current state’scenario was an accurate reflection.
While individual company representatives mentioned
their own concepts for improving the current system to
enable build-to-order strategies, there was no general
disagreement with our findings with regards to the
dynamic capabilities and behavior of the current
scheduling systems. In fact, the problems of adjusting
to shifts in customer demand was seen as a key
constraint in the current system, and the shift towards
diesel engines for passenger cars in Europe was given
as an example, where such sudden demand shifts had
occurred. The focus group was then split into two
groups, in order to develop a Gantt Chart for a 3-year
implementation plan of BTO look like, the detailed
results of which cannot be presented here due to space
constraints.
The ‘build-to-order’scenario (see Fig. 6) retains the
system structure but tries to operate as a real demand
‘pull’, hence aims at a minimum delay and maximum
response to the incoming orders, which essentially are
triggering production (as opposed to a production
forecast). Hence the system simulated reschedules
every day and only works to real orders (no leveling or
averaging). As one would expect, the system does not
overproduce, since only an actual customer order can
initiate the production. Such approach avoids the most
serious waste: overproduction (Monden, 1998).
However, although the system shows a better response
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530518
Fig. 6. Build-to-order scenario: vehicle stocks and order backlogs (units).
to the demand pattern, it slightly drifts into order
backlogs for both A and B. The root cause for this is
the day-to-day variability in demand. If the daily
demand exceeds the maximum capacity, the system
will produce to maximum output, and if the demand is
less than maximum capacity, the system will only
produce what is demanded in the market. As a
consequence, despite overall sufficient average pro-
duction capacity, the system will produce on average
always slightly less than the original demand, hence
the tendency towards the order backlog (see Fig. 6).
An initial analysis of this effect suspected that the
feedback on backlogs was insufficient. However, as
the buffer error weighting equals 1, the entire backlog
is considered each scheduling period, and this effect is
therefore is genuinely caused by insufficient capacity
to respond to demand peaks.
This effect is even more strongly expressed for
model A due to the increase in demand after 6 months,
although the system otherwise proves well capable of
handling the demand swing—in particular compared
to the ‘current state’scenario. Also, due to the daily
changes of the production program and product mix,
the production adaptation indices are relatively high,
whereas the finished goods inventories are signifi-
cantly smaller than in the previous scenario (see
Table 2). The fact that a positive finished good
inventory is accounted for relates to the fact that the
cost lost of sales is also recorded in this variable (see
above), which explains the cost incurred here.
One would assume that orders are built in the order
they are received, but this is not the case in reality. An
order bank reshuffles the orders so as to match
individual plant and component constraints, thus the
order lead-time for an individual order is not standard.
For example, if the production program prescribes a
build of no more than 50% estate cars due to line
balancing constraints, once demand for estate cars
exceeds this barrier, the OTD lead times lengthen. The
question then becomes how often this is ‘rescheduled’,
i.e. how often the production schedules are updated:
monthly, weekly, or even daily? In this scenario, the
plant receives a new production schedule every day,
whereas in the ‘current state’scenario the schedule is
kept constant for 1 month.
The high inventory levels at the supplier are a result
of the non-synchronized raw material ordering, which
cannot cope with the frequent rescheduling needed to
adapt component production to the VM demand
signal. The production adaptation cost at the supplier
however is, despite the daily rescheduling, much lower
than in the ‘current state’scenario, which indicates
that the component production is better synchronized
with the vehicle build, i.e. the supplier is flexible
enough to support the build-to-order strategy at the
vehicle manufacturer.
Another interesting feature is that the correlation
between production and demand is less strong than in
the previous scenario although the system responds
very well to the changes in demand. This effect is
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 519
Table 2
Results of the ‘BTO’scenario
System settings System performance indices
F
1
Average historic demand data used to
determine target order buffer (VM)
1 day InvA VM finished vehicle inventory cost A 21
F
2
Scheduling decision delay (VM) 1 week InvB VM finished vehicle inventory cost B 19
F
3
Order buffer error weighting (VM) 1/1 PrAdaptA VM production adaptation cost A 60
F
4
Period production schedule is kept
constant for (VM)
1 day PrAdaptB VM production adaptation cost B 92
F
5
Average historic demand data used to
determine volume forecast and
mix ratio (VM)
1 day PrAdaptSup Supplier production adaptation cost 20
F
6
Scheduling decision delay (supplier) 1 day InvSup Supplier inventory cost 421
F
7
Average historic demand data used to
determine raw material requirements
(supplier)
1 week CorrA Correlation of production and supply, model A 0.0171
F
8
Safety buffer level used to set reorder
point (supplier)
1 day CorrB Correlation of production and supply, model B 0.0431
caused by the adjustments in production schedules
made due to backlogs and demand. This increased
volatility, compared to the stable schedules in the
previous scenario, is captured in the lower statistical
correlation.
A‘balanced’scenario’, again retaining the system
structure shown in Fig. 3, which attempts to
compromise between demand leveling, production
stability and responsiveness to ensure best possible
customer service (measured in finished goods inven-
tory levels) at minimal cost (production adaptation and
supplier cost). As can be seen in Fig. 7, the system
adapts well to both the day-to-day variation, but also
copes with the 20% increase in demand for product A
in month 6. In terms of the correlation of production
schedules and demand patterns, this scenario achieves
the best fit for model A, despite the increase in
demand. Also, as can be seen in Table 3, the combined
costs for inventory and lost sales are the lowest of all
scenarios. Also, the production adaptation costs at the
VM are almost half that in the BTO scenario, but still
56% higher than in the ‘current state’scenario. At the
supplier, the greater stability in vehicle manufacturer
schedule only results in slightly reduced inventory
levels and production adaptation cost, which again
indicates well-synchronized component production,
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530520
Fig. 7. Balanced scenario: vehicle stocks and order backlogs (units).
Table 3
Results of the ‘balanced’scenario
System settings System performance indices
F
1
Average historic demand data used to
determine target order buffer (VM)
1 day InvA VM finished vehicle inventory cost A 12
F
2
Scheduling decision delay (VM) 1 week InvB VM finished vehicle inventory cost B 10
F
3
Order buffer error weighting (VM) 1/1 PrAdaptA VM production adaptation cost A 38
F
4
Period production schedule is kept
constant for (VM)
1 week PrAdaptB VM production adaptation cost B 40
F
5
Average historic demand data used to
determine volume forecast and
mix ratio (VM)
1 week PrAdaptSup Supplier production adaptation cost 18
F
6
Scheduling decision delay (supplier) 1 day InvSup Supplier inventory cost 420
F
7
Average historic demand data used to
determine raw material requirements
(supplier)
1 week CorrA Correlation of production and demand, model A 0.1362
F
8
Safety buffer level used to set reorder
point (supplier)
1 day CorrB Correlation of production and demand, model B 0.0139
yet problems in tying the raw material supply to this
demand signal. This effect marks a common problem
in the auto industry, whereby first-tier suppliers are
‘squeezed’between VM customers demanding just-
in-time deliveries, and unresponsive raw material
producers that often operate large-scale batch produc-
tion systems (Holweg, 2003).
In conclusion, each of the three scenarios presented
shows different impacts on the different subsystems of
the supply chain, with the balanced scenario showing
an apparent compromise between responsiveness
and the need for stability in the production process.
The next section takes the analysis a step further,
investigating which of these eight factors has the most
impact on the system’s performance.
3.3. ANOVA results
The sensitivity analysis results are shown in Fig. 8.
The sensitivity analysis shows how sensitive the
performance of the overall system is to a change in any
of these factors analyzed. The procedure has identified
that the time period used to identify the volume
forecast and the mix ratios (i.e. looking back over the
last week, month, 2 months in order to determine the
average volume and mix forecast) accounting for 45%
of the overall sensitivity has the most effect on the
dynamic performance of the system. The moving
average (used to define the target order buffers in the
VM) also has a strong impact on the supply chain, with
14%, as does the way that the error in this buffer is
used (20%).
Surprisingly, neither the scheduling frequency at
the vehicle manufacturer (i.e. daily, weekly, monthly),
nor the decision delay (i.e. schedule alterations take
effect in 1 month, 1 week, next day) have much
significance. In fact, it is more important to consider
the time period used to generate averages for the
volume forecast and mix ratios, before they set the
framework for the production schedule. Hence
the research indicates that in the light of day-to-day
demand fluctuations, it is most important to provide an
environment flexible enough to cope with variability
in demand and mix changes, yet at the same time
provide a stable capacity and volume plan.
This finding initially seems a tautology, and has
been discussed in the light of the advantages and
disadvantages of MTF and BTO strategies (Raturi
et al., 1990; Holweg and Pil, 2001, 2004), but is an
important addition to the BTO knowledge. It is
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 521
Fig. 8. ANOVA results: factor sensitivity of the system.
obvious that BTO will incur a certain cost penalty in
terms of providing flexibility of production capacity,
but the findings shows that not religiously striving for
flexibility yields the best results, but in fact a certain
amount of flexibility is not only sufficient, but also
yields better results. It seems most beneficial to
provide a fairly stable background in terms of
production program, i.e. to reschedule on a weekly
rather than daily basis, but still to work to actual
demand, i.e. operate without any forecast. Thus, we
can show that the amount and type of flexibility needs
to be aligned to the expected outcomes, whereas going
beyond does not improve the results.
Since testing our findings in real scheduling
systems was not a realistic option, we verified the
results from the scenario and sensitivity analysis
through focus groups—presenting the findings back to
the project sponsors, and discussing the implications
of our findings within an open forum. The results did
spark a very lively discussion about the general
feasibility of the build-to-order with current systems,
in particular amongst the scheduling and planning
managers at the vehicle manufacturers. The key
problems highlighted in this discussion were the
strong organizational resistance to abandon the rigid
production program, and the issue of making assembly
plants more flexible, which would incur additional
cost. Also, the lead-time of altering or renewing
outdated scheduling systems was pointed out. The
discussion further revealed that logically the findings
of the simulation model seemed accurate, and were
helpful in taking the BTO agenda forward, yet without
realigning performance measures and financial sys-
tems an implementation would be difficult. Top-level
management buy-in and continuity of support were
seen as critical enablers.
3.4. Limitations of the findings
In their review of different techniques for modeling
supply chains, it is noted by Riddalls et al. (2000) that
no one approach is a panacea; their assessment of
continuous time, discrete time, discrete event and
operations research approaches concludes that all have
their own advantages and disadvantages. A particular
strength of modeling using computer simulation is
the ability to better view the whole rather than the
individual parts.
In this analysis we used a difference equation
simulation model to represent the automotive supply
chain from the customer demand, to the vehicle
manufacturer, through to the first-tier component
supplier. Since the shift towards BTO will have wide
ramifications for all tiers in the system, we felt that
considering less tiers would have been a serious
limitation to our study. The difference equation
approach has the advantage of representing both
continuous and discrete time approximations of
supply chains, where the non-linearities or time
invarying parameters make it difficult to analyze
mathematically. For example, Evans et al. (1998) use
difference equation modeling to support continuous
time mathematical modeling, using a Laplace notation
of a supply chain echelon, while Disney and Towill
(2002) use difference equations and Z-transforms to
model a vendor managed inventory supply chain in
discrete time.
However, continuous and discrete time models do
have some disadvantages (Pidd, 1999). Compared to
discrete-event models, for example, the features of
each entity or event in the system cannot be specified
in the same level of detail. With structural modeling it
is not easy to assign product variety attributes or a
customer waiting tolerance, which is more easily
accommodated with proprietary discrete event simu-
lations. Thus, in our simulation we could not link
product attributes to the supply chain, and for example
simulate the effect of different component types
(option versus standard parts, for example) on the
supplier subsystem. Nevertheless, starting with a high-
level model, that encompasses the structural complex-
ity we have observed in this case, is appropriate, as it
lays the foundation for future, more specialized
models, which invariably will be detailed but limited
in capability to understand the whole of the system
structure (Sterman, 2000).
4. Discussion
4.1. Simulation findings
The simulation model discussed in this paper is a
simplified representation of a multi-tier scheduling
system within one of the most complex high-volume
manufacturing settings. Despite necessary simplifica-
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530522
tion, we are confident that we have found an adequate
structural representation of this system, which we
analyzed and verified through a focus group of
industry experts.
The key finding is that the current scheduling
system is poorly designed from a dynamic perspective.
As could be seen in the ‘current state’scenario,
subjecting the scheduling system to a variable demand
pattern demonstrates the system’s inability to adjust
to variable demand, as found in the market place.
Therefore, it is clear that the current system
‘dynamically’cannot support BTO—which is in
addition to the fact that the current average OTD
lead-time of 40 days means that the current system
cannot support BTO ‘structurally’(Holweg and Pil,
2001). In other words, we can now state that not only is
the current system too slow to respond to customer
orders within their waiting tolerance, but also that the
system dynamically cannot respond to changes in
demand, and therefore would incur a great financial
risk of lost sales for the manufacturer if BTO was
adopted.
Upon further investigation, we could show that this
flaw is deeply embedded in the structure of the
feedback loops in the system. The BTO scenario,
which simulates the concept of having an ultimately
responsive and purely demand-driven system, cannot
match production and customer demand and shows
persistent order backlogs, production adaptation cost,
and supplier inventory. On the other hand, a frequent
rescheduling approach with a reasonable smoothing
and proportioned order buffer error feedback loop
seems to be a viable compromise. This was shown
in the balanced scenario, which is more capable to
match supply and demand, at a lower need to adapt
production levels.
The sensitivity analysis further showed that the
most important factor affecting the performance of the
automotive supply chain is the time period over which
the average demand is taken (i.e. the leveling period
used for defining volume forecasts and mix ratios at
the vehicle manufacturer). Inspection of the factor
sum of squares also shows that the longer the average
used the better the dynamic performance of the system
becomes. Interestingly this seems to suggest that
responsive scheduling (i.e. frequent rescheduling,
maybe even daily) and full buffer error weighting
(instant correction of misalignment) are indeed
feasible, so long as the underlying capacity and mix
are stable over a longer period (e.g. averaged over 1
month, although daily rescheduling). This argument is
very important, as current manufacturer scheduling
systems rely on both long scheduling cycles, as well as
long lead times for capacity and mix alteration. The
reason given for this rigidity is generally the fear of
instability, high production adaptation cost and low
capacity utilization. This system however, by focusing
merely on plant volume and efficiency, results in high
stocks of finished vehicles and puts strain on the
suppliers. The major learning points here are that—
within a stable volume setting—more frequent
rescheduling leads to only slightly higher production
adaptation cost, yet eliminates the majority of finished
goods inventory or backlog. With respect to the
transition to build-to-order strategy in the automotive
sector, our findings clearly suggest that more
responsive scheduling, linked to actual customer
demand, is indeed not only feasible, but also highly
beneficial for reducing the inventory of finished cars,
i.e. at the most expensive point in the supply chain.
4.2. Managerial implications
Our research sets clear guidelines what manufac-
turers aiming at reducing their OTD lead-time and
increasing their BTO content need to focus on. In
addition to reducing the throughput lead-times for the
scheduling system, which currently account for 85%
of the delay in the system (Holweg and Pil, 2001;
Holweg, 2003), two main areas for improvement
could be identified. First, the feedback loops in the
system need to be reengineered. In fact, we were
surprised to find out that there is no direct feedback of
information on the number of finished cars in the
current state system when setting new production
targets—only a ‘second guess’via the order buffer is
used as control. As a result, the inventory level of
finished cars is largely uncontrollable and unpredict-
able. In fact the current system inventory of cars is
almost like a ‘random walk’and there is no direct
attempt by the scheduling system to keep the number
of finished cars at a small positive level, providing
high customer service levels at least cost. Where
currently the feedback of a mismatch of demand and
supply is indirect at best (and further weakened
through the buffer error weighting), a direct link
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 523
between demand and supply needs to be established.
For example, imminent overproduction should be
halted by directly linking orders and production
schedules, without interference of the production
program, which marks the second key area for
improvement. The current system is heavily driven
by the production forecast, which was clearly shown
in the sensitivity analysis. A key reason for can be seen
in the historic fact that vehicle supply has traditionally
been forecast-driven with the widespread adoption of
Henry Ford’s mass production techniques. And as
Holweg and Pil (2004) argue, the fundamental strategy
has changed little since the early days of the industry,
so it is hardly surprising to find the forecast as the
factor with the strongest impact in the system. Overall,
our simulation could show that the ‘14-day car’cannot
be achieved by simply altering scheduling parameters
to reduce OTD lead-times, but that structurally the
feedback loops need to be improved, and that the
reliance on the forecast needs to be changed in order to
implement BTO.
Finally, our scenario analysis provides strong
evidence that the level of flexibility needed to support
BTO needs to be adjusted to the circumstances. The
BTO scenario, which analyses the most responsive
scheduling approach possible (daily scheduling and
planning, no forecast at all, strongest feedback on
buffer error) does not provide the best results. In the
BTO scenario, we still observe the dynamic problem
of the matching variable demand with finite capacity.
Only in the balanced scenario, we have a hybrid
system, which aims at a trade-off between BTO and
MTF. Here, a certain stability within an overall
responsive system that is not based on forecasts, but
still aims at keeping the schedules level, shows the
better performance. Thus, the level of flexibility needs
to be aligned to the system requirements; the
maximum flexibility might lead to system nervous-
ness, and sub-optimal dynamic performance. Further-
more, additional stability can be achieved by
managing demand. Dell for example suffers from
considerable swings in demand, and uncertainties in
supply due to capacity constraints. In order to control
the capacity utilization and avoid bottlenecks, Dell
adjusts prices to manage demand. For example, if an
80 GB hard-drive is in short supply, the relative price
for a 100 GB drive is lowered, in order to shift demand
and avoid backlogs and long OTD lead-times.
On a final note, the simulation further was able to
show that, depending on the settings used, cost can be
incurred or saved within the tiers in the system. Hence,
depending on the power in the supply chain, individual
players might be able to optimize their own cost
structure at the detriment of the other players, which
potentially might oppose any potential improvements
to the whole system. Thus, any optimization of the
supply system will also have to be considered within
the remits of existing power regimes (cf. Cox, 1999).
5. Conclusion
Changing current vehicle supply from a mainly
forecast-driven to an order-driven approach marks a
drastic shift in the supply chain, and one that brings
wide ramifications for all players. Previous contribu-
tions have shown that current vehicle supply systems
are ‘structurally’too slow to respond to customer
orders within their waiting tolerance (Holweg and Pil,
2001, 2004). However, we lacked the ‘dynamic’
perspective, as we did not know whether simply
altering scheduling frequencies would have enabled
current systems to build to order. In other words, even
if the scheduling parameters are adapted to support
OTD lead-times of 14 days, as envisaged by many
vehicle manufacturers, the current system is likely to
fail by not being able to respond to demand variability
in the market place.
The simulation analysis showed that, from this
dynamic perspective, current vehicle supply systems
are not capable of supporting BTO. The quality and
speed of the feedback loops between demand and
supply is inadequate, as well as the strong reliance on
forecasting in the scheduling process are the two key
inhibitors we could identify in this study. In many
ways, current scheduling systems operate like driving
a car by looking into the rear-view mirror—reliance on
retrospective data, and slow feedback that makes
counter steering almost impossible. In contrast to
scheduling systems that have been designed to support
BTO, such as at Dell for example, current vehicle
supply systems still show key features of their make-
to-forecast heritage, which makes an immediate BTO
implementation problematic. The outcomes of the
simulation model illustrate well why the current
systems find it so hard to build to order. It is simply
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530524
much less risk-prone for the manufacturers to produce
to forecast, and thus circumnavigate the danger of lost
sales, which—as we have shown—often cannot be
recovered.
Although an immediate implementation of BTO
might not be feasible, we could identify a clear set of
actions that need to be taken in order to move towards
BTO. As the comments from the focus group further
showed, changes to the system are only one piece
to the puzzle. Alignment of performance measures
and reward structures, as well as strong top-level
management support are further essential ingredi-
ents. Further research should therefore focus on the
organizational changes required to support BTO
implementations, and more detailed models on the
implications of demand variability and product
variety on the different subsystems, as well as
explore the link between to demand-driven schedul-
ing and proactive demand management, which marks
an area where the automotive industry lacks far
behind.
Acknowledgements
We gratefully acknowledge the support of the
3DayCar Programme and its sponsors, the Supply
Chain 2001+ project, TRWAutomotive, the Computer
Science Corporation, and the UK’s Engineering and
Physical Sciences Research Council (EPSRC).
Appendix A. Simulation model framework and
assumptions
The model is based on the following framework
and assumptions:
Each run simulates two calendar years, with a
3-month initiation phase.
The average demand is A: 350, B: 700 units per
period. Schedule A plus schedule B always add up
to the overall volume forecast (i.e. no other models
are produced or planned for). See Fig. A.1.
The volume forecast is subject to production
constraints, whereby a maximum 1300 units/day
can be produced.
VM calculates average daily demand of A (looking
back 2 months) and sets target buffer for A. The
target buffer is 5 days. The same procedure happens
for model B, which has an equal buffer of 5 days.
The buffer error compares the actual buffer with the
target buffer settings, and takes a fraction of the
error (e.g. 1/X).
The mix ratio is calculated as follows: for A, the
average daily demand is summed up for the last Y
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 525
Fig. A.1. Demand pattern for models A, B. Model A (lower curve) has a 20% increase in demand after 6 months.
periods, plus the buffer error, which is distributed
over 1 month. For A and B: sum up average daily
demand for last 2 months, add buffer error distributed
over 1 month. The ratio are, for A = A/(A + B), and
similarly for B = 1 ratio for A.
Production schedule is set as follows: the volume
forecast times the mix ratio sets production
schedule for A. The same for B. The production
volume is either the maximum capacity, or the
volume forecast if less than maximum capacity. The
schedule is set for Zdays.
Supplier operations: One component C
1
only goes
into model A. Predicted usage for this component is
the average demand over Xtime. Ydays are the
average decision delay for scheduling raw material
supply. The raw material schedule is set as a normal
distribution of part utilization times production
schedule for part A, to model mix. Assuming a
100% delivery performance, the value of the part is
set at 75% of vehicle’s value. Orders on raw
material supplier are based on ROP and a fixed
order quantity (equals whole steel coils). Raw
material supply lead-time is 5 days.
The following key assumptions are underlying the
model:
The cost for inventory and lost sales is set at 20%
each, and the cost of parts compared to the cost
vehicle equals 75%, as discussed above.
The utilization of the component that goes into A
follows a normal distribution, and thus is unpre-
dictable (i.e. an optional part that is not used every
time).
The buffer error is recovered over the whole month.
Appendix B. Outline of the Taguchi method used
The Taguchi method will be explained in the
following by a description of the analysis conducted
on the sensitivity of the total costs in the supply chain
using the L18 array. The first step is the design of
experiments (DOE). This has two objectives: to
determine the number of trials, and to specify each of
these trials.
The first step of the DOE is to identify the key
factors to be studied and to determine suitable levels
for each factor. This was achieved through a series of
interviews and the process maps of the companies
taking part in the study. The factors and levels are
shown in Table B.1.
The next step was to select an appropriate
orthogonal array to use in the DOE. This is achieved
by calculating the total number of degrees of freedom
(d.f.) for all the factors. The d.f. for each factor is equal
to the number of levels of each factor minus one. Next,
an OAwith at least as many d.f.s as the total d.f., and at
least as many columns as factors under study is
selected. Overall, the one with the smallest number of
trials to reduce the cost of the experiments is the most
favorable choice.
The next stage is to assign the factors to column
with reference to any interactions (using a triangular
table of interaction or a linear graph) and the number
of levels in a column. This completes the DOE stage
and the appropriate simulation studies are completed
to determine the results of each experiment. The next
stage is to conduct an ANOVA study on the results to
determine the percentage contribution of each factor
to the simulation outputs. This is done as shown in
Table B.2.
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530526
Table B.1
Factors, levels and degrees of freedom
Factor Level 1 Level 2 Level 3 d.f.
F
1
Decision delay at suppliers schedule 3 days 3 weeks 3 months 2
F
2
VM level scheduling period 1 day 1 week 2 months 2
F
3
VM decision delay for schedule alterations 1 day 1 week 1 month 2
F
4
VM buffer error weighting 12.5% 25% 100% 2
F
5
Gain in supplier (days of orders) 12 4 1 2
F
6
VM keeps schedule for days 1 day 1 week 1 month 2
F
7
Average for volume forecast and mix ratio 2 days 2 weeks 2 months 2
F
8
Supplier average predicted use 1 week 1 month NA 1
Total d.f. 15
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 527
Table B.2
Experimental results
Supplier:
average
predicted
use
Target
buffer
retro
period
Decision
delay for
schedule
alterations
Buffer error
weighting
Gain in
supplier
(days of
orders)
Keep
schedules
for days
Average
of volume
forecast and
mix ratio
Decision
delay at
supplier’s
schedule
A
production
adaptation
Inventory
cost A
B
production
adaptation
Inventory
cost B
Supplier
production
adaptation
cost
Experiment
1 1 1 1 1 1 1 1 1 44 7 53 16 610.4
2 1 1 2 2 2 2 2 2 34 7 41 10.55 313.4
3 1 1 3 3 3 3 3 3 24 22 25 25.69 42.7
4 1 2 1 1 2 2 3 3 20 28 22 17.06 297.5
5 1 2 2 2 3 3 1 1 49 42 56 169 33.77
6 1 2 3 3 1 1 2 2 52 24 55 24 616.8
7 1 3 1 2 1 3 2 3 35 32 42 30.91 731.3
8 1 3 2 3 2 1 3 1 23 15 26 17 311.1
9 1 3 3 1 3 2 1 2 46 29 55 66.05 31.04
10 2 1 1 3 3 2 2 1 37 9 42 11.77 17.95
11 2 1 2 1 1 3 3 2 20 46 22 23.28 600.2
12 2 1 3 2 2 1 1 3 56 18 65 22.44 297.8
13 2 2 1 2 3 1 3 2 21 15 24 16.08 18.37
14 2 2 2 3 1 2 1 3 58 45 61 85.68 711.9
15 2 2 3 1 2 3 2 1 33 39 41 74.79 290
16 2 3 1 3 2 3 1 2 56 68 62 117.9 291.5
17 2 3 2 1 3 1 2 3 33 9 41 13.05 33.25
18 2 3 3 2 1 2 3 1 21 28 22 23.87 581.7
Level 1 1 Week 1 Day 1 Day 12.50% 12 1 Day 1 Day 1 Day
Level 2 1 Month 1 Week 1 Week 25% 4 1 Week 1 Week 1 Week
Level 3 na 2 Months 1 Month Unity 1 1 Month 1 Month 1 Month
The ANOVA procedure:
1. Total all the results, t.
2. Calculate the correction factor which is equal to t
2
/
18.
3. Square each of the experimental results.
4. Total the sum of squares: This is the total of the
square of the experimental results. Calculate the
total variation as the total sum of squares minus the
correction factor.
5. Determine the factor sum of squares: This is done
by determining the total effect of each level in each
factor, by summing the results of each experiment
with the factor at the appropriate level. This is then
squared and divided by the number of experiments
that used the factor at that level. The factor sum of
squares is then the sum of each level for each factor
minus the correction factor. See Table B.3.
6. Determine the total and factor degrees of freedom,
and fill the table as shown in Tabl e B .4, using the
appropriate d.f. assigned to each factor as outlined in
the DOE earlier. The error d.f. is the d.f. of the OA
minus the total d.f. The column marked Sis the factor
sum of squares determined in the previous step.
7. Calculate the mean square (variance, V): This is the
factor sum of squares divided by the d.f. in each
factor.
8. Calculate the percentage contribution: This is
taken as the factor sum of squares multiplied by
100 and divided by the total variation as defined in
step 4.
9. The next stage is to pool those factors that have a
small contribution to the error by the following
procedure:
Add the sum of squares (S) of non-contributing
factors into the error.
Calculate the new error d.f. This is defined as the
OA d.f. minus the total d.f. of the remaining
factors.
The variance of error can now be calculated, as it
is now deterministic.
Factor Fratios for significant factors can be
calculated as the variance of the factor divided
by variance of the error.
The pure sum of squares (S0) for signifi-
cant factors can now be determined as the
sum of squares of the factor minus the product of
the d.f. for the factor and the variance of the
error.
Finally, the new percentage contributions of
significant factors can be determined. It is
calculated as the pure sum of squares multiplied
by 100 divided by the total variation (See
Table B.5).
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530528
Table B.3
Determination of the factor sum of squares
Supplier:
average
predicted
use
Target
buffer
retro
period
Decision
delay for
schedule
alterations
Buffer
error
weighting
Gain in
supplier
(days of
orders)
Keep
schedules
for days
Average
of volume
forecast and
mix ratio
Decision
delay at
supplier’s
schedule
Average
effect of
level 1
3522 2856 3429 3236 3995 2888 5172 3533
Average
effect of
level 2
3706 4274 3849 3617 3461 3412 3269 3695
Average
effect of
level 3
3712 3565 3990 3386 4543 2402 3615
A1 31,700 17,137 20,576 19,415 23,973 17,327 31,033 21,197
A2 33,358 25,646 23,091 21,703 20,768 20,473 19,614 22,172
A3 0 22,275 21,391 23,940 20,318 27,258 14,411 21,689
111,655,853 48,947,603 70,562,270 62,824,330 95,780,302 3,309,578 160,508,067 74,885,793
123,637,972 109,617,365 88,868,574 78,502,265 71,884,198 69,857,638 64,116,455 81,933,003
0 82,695,801 76,259,396 95,520,917 68,800,632 123,829,194 34,613,898 78,401,600
152,644 6,119,588 549,058 1,706,330 1,323,951 8,585,229 24,097,239 79,215
References
Checkland, P., Scholes, J., 1990. Soft Systems Methodology in
Action. Wiley, Chichester.
Cox, A., 1999. Power, value and supply chain management. Supply
Chain Management 4 (4), 167–175.
Cusumano, M.A., Takeishi, A., 1991. Supplier relations and
management: a survey of Japanese, Japanese-transplant, and
US auto plants. Strategic Management Journal 12, 563–
588.
Disney, S.M., Grubbstrom, R.W., 2004. Economic consequences of
production and inventory control policy. International Journal of
Production Research 42 (17), 3419–3431.
Disney, S.M., Naim, M.M., Towill, D.R., 1996. Dynamic simu-
lation modelling for lean logistics. International Journal of
Physical Distribution and Logistics Management 27 (3),
174–196.
Disney, S.M., Naim, M.M., 1999. Improving the effectiveness of
supply chains. In: Proceedings of the 15th International
Conference on Production Research, Limerick, August.
pp. 637–640.
Disney, S.M., Towill, D.R., 2002. A discrete transfer function model
to determine the dynamic stability of a vendor managed inven-
tory supply chain. International Journal of Production Research
40 (1), 179–204.
Elias, S., 2002. New Car Buyer Behaviour. 3DayCar Research
Report, Cardiff Business School.
Evans, G.N., Naim, M.M., Towill, D.R., 1998. Application of a
simulation methodology to the redesign of a logistical control
system. International Journal of Production Economics 56–57,
157–168.
Forrester, J.W., 1961. Industrial Dynamics. MIT Press, Cambridge.
Hammant, J., Disney, S.M., Childerhouse, P., Naim, M.M., 1999.
Modelling the consequences of a strategic supply chain initiative
of an automotive aftermarket operation. International Journal
of Physical Distribution and Logistics Management 29 (9),
535–551.
Helper, S., 1991. How much has really changed between US auto-
makers and their suppliers? Sloan Management Review 32 (4),
15–29.
Hertz, S., Johannsson, J.K., de Jager, F., 2001. Customer-oriented
cost cutting: process management at Volvo. Supply Chain
Management 6 (3), 128–141.
Hines, P., 1998. Benchmarking Toyota’s supply chain: Japan vs.
U.K. Long Range Planning, February 31 (6), 911–918.
Hines, P., Rich, N., 1997. The seven value stream mapping tools.
International Journal of Operations and Production Management
17 (1), 46–64.
Holweg, M., 2003. The three-day car challenge—investigating the
inhibitors of responsive order fulfilment in new vehicle supply
systems. International Journal of Logistics: Research and Appli-
cations 6 (3), 165–183.
Holweg, M., Pil, F., 2001. Successful build-to-order strategies start
with the customer. Sloan Management Review 43 (1), 74–83.
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530 529
Table B.5
Pooled ANOVA score
d.f. SVFSV p(%) p0(%)
Target buffer retro period 2 132,474,662.79 66,237,331.39 16.425556 124,409,507.32 13.60394699 12.69
Buffer error weighting 2 190,726,107.96 95,363,053.98 23.648163 182,660,952.50 19.58584236 18.67
Gain in supplier (days of orders) 2 68,891,823.99 34,445,911.99 8.541909 60,826,668.52 7.07456582 6.16
Average for vol. forecast and mix ratio 2 445,828,850.39 222,914,425.19 55.278395 437,763,694.92 45.7825815 44.87
Error 7 28,228,044.12 4,032,577.73 1 168,134,964.91 2.898764244 17.61
Table B.4
ANOVA of total score
d.f. SVp(%)
Supplier average predicted use 1 16,961,225.56 16,961,225.56 1.741764112
Target buffer retro period 2 132,474,662.79 66,237,331.39 13.60394699
Decision delay for schedule alterations 2 39,470,362.02 19,735,181.01 4.053248381
Buffer error weighting 2 190,726,107.96 95,363,053.98 19.58584236
Gain in supplier (days of orders) 2 68,891,823.99 34,445,911.99 7.07456582
Keep schedules for days 2 2,957,647.49 1,478,823.75 0.303723587
Average for vol. forecast and mix ratio 2 445,828,850.39 222,914,425.19 45.7825815
Decision delay at suppliers schedule 2 48,257,063.86 24,128,531.93 4.955563009
Error 0 28,228,044 #DIV/0! 2.898764244
100
Holweg, M., Pil, F.K., 2004. The Second Century: Reconnecting
Customer and Value Chain through Build-to-Order. MIT Press,
Cambridge.
ICDP, 2000. Fulfilling the promise—is there a future for franchised
car distribution? Research report, International Car Distribution
Programme, Solihull.
Kiff, J., 1997. Supply and stocking systems in the UK car market.
International Journal of Physical Distribution and Logistics
Management 27 (3–4), 226–243.
Krafcik, J., 1988. Triumph of the lean production system. Sloan
Management Review Fall issue, 41–52.
Lapidus, G., 2000. E-automotive: gentlemen, start your search
engines. Research report, Goldman Sachs, New York.
Lockyer, K., Gordon, J., 1991. Critical Path Analysis and Other
Project Network Techniques. Pitman Publishing, London.
Monden, Y., 1983. The Toyota Production System. Productivity
Press, Portland.
Monden, Y., 1998. Toyota Production System: An Integrated
Approach to Just-in-Time, third ed. Engineering and Manage-
ment Press, Norcross.
Naim, M.M., Childerhouse, P., Disney, S.M., Towill, D.R., 2002a. A
supply chain diagnostic methodology: determining the vector of
change. Computers and Industrial Engineering: An International
Journal 43 (1–2), 135–157.
Naim, M.M., Disney, S.M., Evans, G.N., 2002b. Minimum reason-
able inventory and the bullwhip effect in an automotive enter-
prise: a ‘foresight vehicle demonstrator, SAE 2002’. Tran-
sactions Journal of Materials and Manufacturing 196–203.
Ohno, T., 1988. The Toyota Production System: Beyond Large-
Scale Production. Productivity Press, Portland.
Pidd, M., 1999. Just modeling through—a rough guideline to
modeling. Interfaces 29 (2), 118–132.
Pil, F.K., 2002. Sustaining manufacturing excellence: results from
round three of the international assembly plant study. Sponsors
meeting, MIT International Motor Vehicle Program, Dedham,
MA, September.
Pil, F.K., MacDuffie, J.P., 1996. The adoption of high-involvement
work practices. Industrial Relations 35 (3), 423–455.
Pil, F.K., Holweg, M., 2004. Linking product variety to order
fulfilment strategies. Interfaces 34 (5), 394–403.
Raturi, A., Meredith, J., McCutheon, D., Camm, J., 1990. Coping
with the build-to-forecast environment. Journal of Operations
Management 9 (2), 230–249.
Riddalls, C.E., Bennett, S., Tipi, N.S., 2000. Modelling the
dynamics of supply chains. International Journal of Systems
Science 31 (8), 969–976.
Rother, M., Shook, J., 1998. Learning to See: Value Stream Mapping
to Create Value and Eliminate Muda. The Lean Enterprise
Institute, Boston.
Roy, K.R., 1990. A Primer on the Taguchi Method. Van Nostrand
Reinhold, New York.
Shapiro, B.P., Rangan, V.K., Sviokla, J.J., 1992. Staple yourself to an
order. Harvard Business Review 70 (4), 113–123.
Stalk, G., 1988. Time—the next source of competitive advantage.
Harvard Business Review 66 (4), 41–51.
Sterman, J.D., 1989. Modeling managerial behavior: misperceptions
of feedback in a dynamic decision making experiment. Manage-
ment Science 35 (3), 321–339.
Sterman, J.D., 2000. Business Dynamics—Systems Thinking and
Modeling for a Complex World. McGraw Hill, Boston.
Simon, H.A., 1962. The architecture of complexity. In: Williamson,
O.E. (Ed.), Industrial Organization. Edward Elgar Publishing,
Aldershot (1990).
Wells, P.E., Nieuwenhuis, P., 2001. The automotive industry—a
guide. BT executive guide, British Telecommunications, Lon-
don.
Williams, G., 1999. European new vehicle supply—the long
road to customer pull systems. ICDP Journal 1 (1), 13–
21.
Williams, G., 2000. Progress towards customer pull distribution.
Research paper 4/2000, The International Car Distribution
Programme, Solihull.
Womack, J., Jones, D.T., Roos, D., 1990. The Machine That
Changed the World. Rawson Associates, New York.
Womack, J., Jones, D.T., 1996. Lean Thinking. Simon and Schuster,
New York.
M. Holweg et al. / Journal of Operations Management 23 (2005) 507–530530
