Professional Documents
Culture Documents
Dr Stephen Disney
Abstract
There are various measures of the bullwhip effect proposed in the literature [2]. The
most common measure is the ratio of the variance of the order rate to the variance
of the demand rate, see equation (1). It is the measure that we will focus on herein,
and works best for stationary, stochastic, discrete time demand processes. However
there are other measures. Standard deviations could be used instead. Indeed, as we
will reveal later, this is more natural when the economics of the bullwhip (and
inventory) are considered. Another (practical) measure is ratios of the co-efficient of
variation ( COV=variance/mean) of order and demand rates. This is a useful
measure when there are multiple products going through multiple routes to market
and some comparison is needed across different products, businesses or routes to
market.
(1)
The biggest decision to make is whether to study the bullwhip problem in discrete
or continuous time. In discrete time, system states (demand rates, inventory and
WIP levels) and replenishment orders are made at the equally spaced moments of
time. In between these moments of time, nothing is known about the system. In
continuous time the systems states are monitored at all moments of time and the
order rate is continuously adjusted.
The Laplace transform was originally developed by Laplace and Euler in the 17 th
century for studying the orbits of planets. However, electronic engineers have
developed a whole range of tools, loosely termed control theory, for studying
continuous time systems based on Laplace transformed transfer functions. These
transfer function techniques work very well if the system is linear, time invariant
(LTI - a common assumption) and the system has no initial conditions (IC). Simon
(1952) [7] seems to have been the first to apply the Laplace transform to a
production and inventory control problem. Transform approaches work well in
Single Input and Single Output (SISO) scenarios as then only a single transfer
function is required. Transforms also contain complete information about frequency
response of the system. Interestingly, the transforms that describe cash flows are
directly related to the Net Present Value of that cash flow [8].
If the two assumptions of LTI and zero IC do not hold, then the analytical
approaches have to resort back to the (non-linear) differential equation forms.
Unfortunately there is no “standard approach” for analysis of such systems. Indeed
many systems have no known solution, and even when we can obtain a solution
there is often an infinite number of them, one for each set of IC and non-linearity.
Linear differential equations are also readily handled by state-space techniques.
These are essentially matrix representations of systems of equations. State-space
methods are especially good at handling Multiple Input, Multiple Output (MIMO)
systems and can be easily extended to include non-zero IC’s.
The discrete time analogue of the Laplace transform is the z-transform. It was
developed independently by scholars from the UK [11] and Russia [12] during the
Second World War for controlling such things radar and gun targeting systems and
other applications that involved the newly available digital computers. The first
book that brought together all of the developments of the z-transform was by Jury
(1964) [13], but the first person to apply the z-transform to a production and
inventory control problem appears to have been Vassian (1955) [14]. The
advantages of using the z-transform over the time domain difference equations are
the same as for the continuous case; convolution in the time domain is
multiplication in the frequency domain. However, the disadvantages are that it has
to be LTI and possess zero IC. However, problems with the pure-time delay are
completely avoided in discrete time as it forms the kernel of the z-transform. State
space methods (with the same advantages) are also available in discrete time.
In discrete time a lot can be done with stochastic techniques using the expectation
operator. However, the calculation of the co-variances can become very tedious
when complex systems are studied. Interestingly this difficultly is completely
avoided with transform approaches. Martingales have also been used to study
inventory problems, for example, see Graves (1999) [15]. Martingales are useful
tools as they can yield insights into magnitude of infinite variances that occur in
non-stationary time series.
(2)
Other approaches
Any time series, continuous or discrete, can be analysed using variations of the
Fourier transform. This is a frequency response method, where a time series is
broken up into a series of harmonics. Harmonics are sine waves of different
frequencies, amplitudes and phase lags. Understanding how replenishment rules
respond to the complete spectrum of individual harmonic frequencies allow us to
understand how they react any demand signal, thus the tool is particularly powerful
[17].
The beer game, a table top management game, is also very good demonstrator of
the bullwhip effect. It may also be used to generate knowledge and insights into
how actual people manage supply chains, Sterman (1989) [18]. The beer game has
been used extensively in business schools world-wide and many forms of it exist.
Some variants of the beer game are simply electronic or internet based [19]
versions of the original game, others have been adapted to represent particular
industries [20] or to investigate particular supply chain strategies such as
information sharing or VMI [21].
The bullwhip effect is greatly influenced by the structure of the material and
information flow in a supply chain and the how the replenishment decisions are
made. In this section we will briefly review the impact of the supply chain structure
and how it influences the bullwhip effect. We will do this with a “water tank model”
as it very easily illustrates the main principles [24]. In the water tank model, water
represents inventory, tanks represent companies, flow of water out of a tank
represents sales, flow in, deliveries. The replenishment or ordering decision is
conceptualised with a “ball cock” that regulates the flow.
In a traditional supply chain each echelon makes its own replenishment decision,
based on it own local information. This local information often includes sales (flow
out of its tank), inventory (water in the tank) and orders placed, but not yet
received or WIP (water in the pipe flowing towards the tank), see Figure 1. Most
supply chains operate in this mode. The problem is bullwhip increases geometrically
in traditional supply chains. Every time the order passes through an echelon of the
supply chain, the variance of the order rates “multiples up”, and this causes
inefficiencies.
Information sharing
Some supply chains have the ability to share point of sale data to the end consumer
with other members of the supply chain. Consider, for example, the supermarket
supply chains in the UK. As the barcodes are scanned at checkouts an electronic file
is populated from which sales patterns for particular products can be determined.
These sales patterns are then transferred (sold even) to suppliers. Often suppliers
use this in capacity planning activities, but the real benefit comes from using it in
their replenishment / ordering decisions (see Figure 2). Using the POS data solely in
the capacity decisions has no effect on the dynamics of the supply chain; the
bullwhip effect still exists, and it still increases geometrically. Thus suppliers are left
dismayed, wondering why the retailer’s orders fluctuate widely despite fairly steady
sales.
The gold standard in supply chains structures however comes from completely
eliminating replenishment decisions. This can be done via Vendor Managed
Inventory (VMI) arrangements. In VMI supply chains suppliers have complete
visibility of the downstream supply chain; they can see end consumer demand, their
customer’s inventory levels and the contents of the pipeline (WIP). They can then
base their replenishment rules on the state of the complete downstream supply
chain. It is then possible to get a multi-echelon supply chain to act dynamically as a
single echelon. This effectively removes the bullwhip problem from the supply
chain.
It is often assumed that the supplier is managing the retailer’s inventory. This is
simply unacceptable for many retailers, as they consider their inventory
management skill to be a core competency of their business. However, the supplier
does not need to be in direct control of the retailer’s inventory. Consider the second
water tank model of Figure 4. It does not matter who’s “hand” is controlling the
“cup” (the cup is an analogy for a truckload of product) to move the inventory from
one company to the other. It could be either the retailer or the supplier. What is
important is not who owns inventory or who makes the decisions. What is important
is how decisions are formed and what information is used within them.
The OUT policy makes replenishment decisions based on three factors; a forecast of
future demand, an inventory level discrepancy term and a WIP level discrepancy
term, see equation (3).
(3)
Notice in (3) we have defined the Target Net Stock as a time invariant constant.
This need not always be the case, it maybe appropriate to assume that it is a
function of time, t. The forecast of demand can be done in a variety of ways, but
here it is convenient for demonstration purposes to use simple exponential
smoothing, see equation (4).
(4)
To complete the definition of the OUT policy we will also need the inventory and WIP
balance equations. These are
, (5)
. (6)
In equations (5) and (6) there is a new term, T p, which is the replenishment lead-
time. This lead-time is the time between placing an order and receiving the
products ordered into stock. This order can be placed either on a supplier or on a
production system.
It is also obvious from (5) that there is a link between the bullwhip effect and “net
stock variance amplification” or NSAmp. NSAmp is the analogy of bullwhip in the
inventory levels. Clearly, when designing a replenishment rule, we want to avoid
both bullwhip (capacity related) costs and NSAmp (inventory related) costs.
(7)
Better forecasting
An important skill for companies to master in order to tame the bullwhip effect is to
create good forecasts. By “good” we mean forecasts that minimise mean squared
error (MMSE) between the forecast of demand over the lead-time (and review
period in discrete time systems) and the actual, realised demand over the lead-time
(and review period). Assuming an i.i.d. demand and simple exponential smoothing
forecasting is used inside the OUT policy, then the bullwhip is given by,
. (8)
It is also interesting to see what happens to the inventory bullwhip effect, NSAmp,
in our stylised example. It is given by equation (9),
(9)
Here we can see that the inventory variance is always greater than unity. Hence
inventory levels will always vary more than the demand signal and it is not possible
to achieve zero inventory policy without making the customer wait for the product
with this ordering decision. The last part of equation (9) tends to zero as Ta
approaches infinity. Thus, more accurate forecasting also leads to lower inventory
levels.
Reduce lead-times
Reducing lead-times will also reduce bullwhip. For example, see equation (8). Here
T p, the replenishment lead-time, occurs in the numerator of the bullwhip
expression and thus smaller T p’s will mean less bullwhip is produced. Furthermore,
for i.i.d. demand and MMSE forecasting, the lead-time drops out of the order
variance ratio expression. Unfortunately this is not a general conclusion. For other
types of demand processes (such as ARIMA demand) and other forecasting
mechanisms, lead-time does play a role. What is interesting though is the fact that
bullwhip will never increase when lead-time decreases.
Reducing lead-time will also have a positive impact on the inventory costs in a
supply chain. Inspection of equation (9) reveals that the inventory variance is made
up of three components. The first unit of inventory variance is due to the order of
events delay. That is, because demand for this planning period is not known until
the end of the planning period, inventory has got to vary at least as much as the
demand levels (in a continuous system this order-of-events delay does not occur
and thus this statement does not hold anymore). Inventory variance increases as
the production lead-time increases. In fact, we can see that inventory levels will also
vary by at least ((1+ Tp)*(Variance of demand)), no matter how well the forecasts
are generated. The third component of the inventory variance (equation (9)) is to
cover the forecasts errors introduced by the non-optimal forecasts.
A control engineer may look at equation (3) and not be at all surprised by the fact
that the OUT policy produces bullwhip effects. Scrutiny of equation (3) shows that it
has two feedback loops, one for the net stock and one for the WIP. Both of these
feedback loops have a gain of unity. That is, all of the discrepancy between actual
and target inventory (and WIP) is incorporated into the order rate.
(10)
(11)
In (11) we notice that bullwhip is now independent of the lead-time. Under the same
conditions the NSAmp expression becomes
(12)
where can now see that both T i>1 (smoothing) and T i<1 (bullwhip) increases
inventory requirements. However, the influence of the lead-time on the inventory
has remained the same; longer lead-times mean that the variance of the inventory
levels increases.
An interesting trade-off situation now occurs. If we assume costs were equal to the
order and inventory variances we may wish to minimise the sum of the bullwhip and
NSAmp expressions. Adding together equations (11) and (12), differentiating them
with respect to T i and solving for zero gradient shows that the optimal value of T i
is the golden ratio, 1.618034. This is illustrated graphically in Figure 5.
The use of proportional controllers need not be limited to feedback loops either. By
exploiting optimal control theory and state-space approaches it is possible to derive
policies that take full account of all the states in the system. These types of systems
are known as a full-state feedback systems [29] and can account for the structure
of complex demand process in a superior manner when compared to the traditional
(and the proportional) order-up-to policy.
Figure 5. Variance ratios with the golden feedback controller
When we consider multi-echelon supply chains then there a whole new range of
options available to us. For example, we have already spoke of VMI supply chains.
Here downstream supply chain states are communicated to suppliers and this
information is used in their replenishment decisions. This type of arrangement is
closely related to what is known as the echelon-stock policy or the echelon order-
up-to policy, Hoberg, Bradley and Thonemann (2007) [30].
Finally, there is a very innovative and novel method currently being developed by
Gaalman and Disney (2007) [33] for co-ordinating a multi-echelon supply chain. This
policy has been derived using optimal control theory and, in a sense, is like a VMI
supply chain in reverse. The core of the idea is that because the retailer’s lead-time
is shorter than the manufacturer’s lead-time, he can correct some of the
manufacturer’s forecast errors. In this way, the retailer is accounting for the state of
the “upstream” supply chain. The analysis of this policy is not yet complete but
initial findings are very interesting.
Economics of the bullwhip problem
) and N be the cost per unit per period of not producing to the available production
capacity. Then there will be piece-wise linear, convex lost capacity costs in each
period of
. (13)
Let P be the cost per unit per period of producing in over-time, so there are piece-
wise linear, convex over-time costs in each period of
. (14)
When this optimal amount of capacity exists then the bullwhip related costs are
. (16)
Equation (16) shows us that the bullwhip costs are linearly related to the standard
deviation of the order rates.
Let’s turn our focus now to a distribution network. Assume that we have multiple
retailers being served by a number of distribution centres. All transportation lead-
times are the unity, regardless of the number of distribution centres that exist. Each
retailer faces i.i.d. demand with the same mean and variance. Furthermore each
retailer and distribution centre employs a traditional OUT policy with MMSE
forecasting and unit feedback gains. For such a scenario Maister (1976) [34]
introduced the “Square Root Law for Inventory” when consolidation occurs in the
distribution network. Quoting directly from Maister,
“If the inventories of a single product (or stock keeping unit) are originally
maintained at a number ( n) of field locations (refereed to as the decentralised
system) but are then consolidated into one central inventory then the ratio
exists”, Maister (1976).
Amazingly, the square root law also exists for bullwhip costs. Consider the capacity
related costs at the DC echelon. Equation (16) shows us that bullwhip or capacity
costs are given by
. (17)
Y is a constant determined by the lost capacity and overtime costs. It is easy too
prove the square root law for bullwhip exists by considering that in the
decentralised supply chain the standard deviation of the orders at each of the n
distribution centres is
, and in the centralised supply chain (with only one distribution centre) the standard
deviation of the orders is
. Thus,
, (18)
which is the “Square Root Law for Bullwhip” [35]. This result surprised us, as
intuitively, we expected it to have the opposite impact. This result also suggests
that reasons to consolidate distribution networks are actually a lot stronger than
previously thought (as this is often based solely on inventory costs). The likely
impact of this is to force companies to consolidate even further than they have in
the past, increasing the amount of traffic on the road. Thus, internalising the
external costs transportation causes is now even more important.
Joint replenishment polices (JRP) which control more than one product in the same
inventory replenishment decision is a very promising but complex issue. It is closely
related to the inventory routing problem (IRP) and there are some initial results [36]
on both the JRP and IRP that have some interesting properties and show improved
economic performance. This is especially true when order set-up or transportation
costs are considered.
Recently we have discovered that there is a link between smoothing and lead-times.
We have been using queuing theory to model a manufacturer. The manufacturer
works on a make-to-order, first come, first served basis. When a retailer smoothes
his replenishment order that is placed on a manufacturer, the lead-time that the
manufacturer needs to produce and deliver the products is reduced. Thus, there is a
link between bullwhip and lead-time variability [37]. The smoother the retailer’s
order is, the quicker the manufacturer can replenish his order, on the average. Thus
there is actually a mechanism to break the bullwhip, inventory trade-off we spoke of
earlier. Other research that has recently considered similar problems was
conducted by Chatfield, Kim, Harrison and Hayya (2004) [38] who investigated the
impact of stochastic lead-times, information quality and information sharing in the
OUT policy via a simulation experiment. This approach is further clarified with
analytical insights in Kim et al (2006) [39].
Multi-echelon policies offer a very promising route for future bullwhip analysis.
However, they will require a significant industrial engineering effort to implement in
practice. Collaboration and co-ordination mechanisms are also needed. However,
the biggest problem with multi-echelon research is properly capturing the effect of
lost sales and capacity constraints. This is very difficult to achieve as these systems
are non-linear. However, Markov Chains do offer a means of analysis and will even
cope with quantised systems where only integer amount of products can be ordered
from a supplier or production system.
Finally, if you would like to explore more about the bullwhip effect, please go to
www.bullwhip.co.uk. There you will find a collection of simulations, java explorers,
table top games and reference lists associated with the bullwhip effect.
References