Professional Documents
Culture Documents
Efecto Latigo Wright2008
Efecto Latigo Wright2008
Abstract
The ‘‘bullwhip’’ effect, in which order variability increases as one moves up the supply chain, has been observed in a
range of industries, modeled by several authors and various remedies suggested. This paper provides a simulation of the
effect of improved forecasting methods, and finds that Holt’s and Brown’s methods substantially mitigate the bullwhip
effect across a range of performance metrics. The end result is to identify ordering policies that perform particularly well in
combination with these forecasting methods and indicate how they can be implemented in practice.
r 2008 Elsevier B.V. All rights reserved.
Keywords: Supply chain management; Bullwhip effect; Forecasting; Ordering policy; Simulation model.
0925-5273/$ - see front matter r 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.ijpe.2007.10.022
ARTICLE IN PRESS
588 D. Wright, X. Yuan / Int. J. Production Economics 113 (2008) 587–597
In terms of management science techniques, Yao forecasting procedures. Section 4 describes how
and Dong-Qing (2001) indicates that demand these results can be used in practice to alleviate the
forecasting and ordering policies are two key harmful effects of the bullwhip effect in supply
methods of controlling the bullwhip effect and Paik chain management.
and Seung-Kuk (2003), in a statistical study,
identified demand forecasting as one of the sig- 2. Methods and models
nificant variables for bullwhip control. Miyaoka
and Hausman (2004) also found that improved Our model is based on the four-stage supply
forecasting could reduce fluctuations in manufac- chain, shown in Fig. 1. Orders for goods propagate
turing production levels. upstream from left to right, and goods are shipped,
Based on this research, we focus on two factors: downstream, in an opposite direction. We use a
demand forecasting and ordering policies. Moving Basic Model (BM) based on Sterman’s model, with
average and single exponential smoothing methods a customer and four trading partners: retailer,
have been used by Graves (1999) and Chen et al. wholesaler, distributor, and factory. Each trading
(2000a, b). The first aim of the current paper is to partner has its own stock management system and
use more sophisticated forecasting methods, includ- ordering decision system. Extending the BM, we
ing Holt’s method and Brown’s double-exponential build three other models: BM using moving average
smoothing, DES, method. forecasting (BM+MA), BM using Holt’s forecast-
Most authors use ‘‘order up-to’’ policy, including ing (BM+Holt’s), and BM using DES or Brown’s
Chen et al. (2000a, b) and Lee et al. (2004). Order forecasting (BM+Brown’s).
rate is decided by requisitions received from the Fig. 2 shows the flow relationships between two
upstream trading partner, the gap between desired trading partners. The order decision system comes
and actual inventory, and delay in inventory. By from expected demand (EDt), adjustment of
changing the definition of desired and actual stock (ASt), and adjustment of supply line (ASLt).
inventory, Kohli (2005) gave a new equation for Orders (Ot) increase the supply line while shipment
ordering policy including lead time, safety stock, received (SRt) decreases the supply line. Goods flow
and order frequency, and Sterman (1989) applied out of stock to decrease inventory while shipment
generic stock acquisition and an ordering heuristic received (SRt) flows into the stock to increase
in his model. inventory.
The present paper is based on modifications to
Sterman’s model to study how different ordering 2.1. Bullwhip model for retailer, wholesaler, and
policies and forecasting techniques, either separately distributor
or in combination, can control the bullwhip effect.
We identify a range of ordering policies for which Since the business process model for the factory is
the bullwhip effect can be alleviated by using either different from other trading partners, the model for
Holt’s or Brown’s forecasting method. retailer, wholesaler, and distributor is described
The remainder of the paper is organized as first. In Section 2.1.1, a basic conceptual model is
follows. Section 2 describes the basic model and defined and in Section 2.1.2 we modify it to give a
how we have extended it to incorporate Holt’s and computationally efficient formulation.
Brown’s forecasting methods. Section 3 describes
the simulation methodology and the metrics used to 2.1.1. Basic conceptual model definition
measure the extent of the bullwhip effect. It also The equations in Tables 1 and 2 formalize
presents the results, identifying which ordering heuristics based on the Sterman’s model (Sterman,
policies can benefit from Holt’s and Brown’s 1989).
Wholesaler
SL t St
SRt
Supply Goods
Stock
Line
Ot
ASLt
SL t St
SR t
Supply Goods
Stock
Line
Ot
ASLt AS t
ED t Order Ot
OR t EDt
Table 1 Table 2
Original conceptual model Variables definition
Ot ¼ MAXð0; IOt Þ (1)
Variables Specification Initial
condition
IOt ¼ EDt þ ASt þ ASLt (2)
Ot Order quantity that trading O0 ¼ 0
ASt ¼ aS ðS ! & St&1 Þ, (3) partners place, should be non-
negative, which means trading
ASLt ¼ aSL ðSL! & SLt&1 Þ, (4) partners cannot cancel or reverse
order.
St ¼ St&1 þ SRt&1 & Lt&1 (5) IOt The indicated order rate is the
order the trading partners will
SLt ¼ SLt&1 þ IOt & SRt (6) place, which is based on the
anchoring and adjustment
EDt ¼ y ORt&1 þ ð1 & yÞEDt&1 ; 0pyp1 (7) heuristic (Tversky and Kahneman,
1974, pp. 1124–1131). In the
model, the anchor is a forecasted
demand at each time step, and we
2.1.2. Computationally efficient formulation adjust it based on the status of
stock and supply line.
The desired stock S ! and supply line SL! are
ORt Order received at time t. The
assumed to be constant. We define b ¼ aSL/aS and
quantity the downstream trading
S 0 ¼ S ! þ b $ SL! , yielding partner requests.
EDt EDt is the expected demand rate at ED0 ¼ 0
IOt ¼ EDt þ aS ðS 0 & St & b $ SLt Þ (8)
time t. We forecast the expected
Since S ! X0, SL! X0, aSLX0 and aSX0, therefore demand rate from the historical
0 demand from the downstream
S X0. Trading partners generally pay more atten-
trading partner. In (7), the expected
tion to the inventory than the supply line, so demand comes from a single-
that,aSLpaS, implying 0pbp1. exponential smoothing method.
The simulation is initialized in equilibrium. Each ASt Adjustment to correct
inventory contains 12 units S0 ¼ 12. discrepancies between the desired
ARTICLE IN PRESS
590 D. Wright, X. Yuan / Int. J. Production Economics 113 (2008) 587–597
and actual stock. The adjustment IOt ¼ EDt þ aS ðS0 & St & b $ SLt Þ (8)
for the stock creates a negative
feedback loop, which regulates the b ¼ aSL =aS (9)
stock. For simplicity the
adjustment is linear. S0 ¼ S! þ b $ SL! (10)
!
S Desired stock.
St Actual stock or inventory at time S0 ¼ 12 St ¼ St&1 þ SRt&1 & Lt&1 (5)
t. Inventory at time t is the
inventory at time t&1 plus goods SLt ¼ SLt&1 þ IOt & SRt (6)
received at time t&1 from
upstream trading partner minus EDt ¼ y $ ORt&1 þ ð1 & yÞEDt&1 ; 0pyp1 (7)
goods sent to downstream trading
partner at time t&1.
aS The fraction of the discrepancy 0paSp1
ordered in each period.
line is the volume of products ordered but not yet
ASLt Adjustment to correct
discrepancies between the desired produced. We assume it takes 3 time steps to
and actual supply line. As with produce goods, so that Eq. (5) in Table 3 becomes
ASt, the adjustment is linear and
creates a negative feedback loop. St ¼ S t&1 þ IOt&3 & Lt&1 (11)
SL! The desired supply line.
SLt The actual supply line records the SL0 ¼ 8
volume of goods that has been 2.3. BM+MA: basic model using moving average
ordered but not received from its forecasting
supplier. The supply line at time t
is the supply line at time t&1 plus
new orders placed by the trading Our first modification to Sterman’s model is to
partner itself minus the orders use the moving average forecasting method so that
fulfilled by supplier at time t&1. (7) becomes Eq. (12). We define n ¼ 6 in our
The initial condition for SL is set simulation model.
equal to twice the standard
deviation in customer demand. EDt ¼ ½ORt&1 þ ) ) ) þ ORt&n *=n (12)
aSL The fractional adjustment rate for 0paSLp1
the supply line.
SRt Shipment received from upstream 2.4. BM+Holt’s: BM using Holt’s forecasting
trading partners to increase stock method
or inventory at time t.
Lt Loss of the inventory, i.e. the When Holt’s forecasting method (Hanke and
amount of goods shipped to the
downstream trading partner. Lt of Reitsch, 1991, p. 128) is used, (7) is replaced by
a given trading partner is equal to EDt ¼ At&1 þ Bt&1 (13)
the SRt+1 of its downstream
trading partner. where
At ¼ l $ ORt þ ð1 & lÞðAt&1 þ Bt&1 Þ (14)
Using b the model can be efficiently simulated Bt ¼ gðAt & At&1 Þ þ ð1 & gÞBt&1 (15)
using the equations in Table 3.
At: smoothed value,
2.2. Bullwhip model for the factory l: smoothing constant for the data (0plp1),
g: smoothing constant for trend estimate
As the last node of the supply chain, the factory (0pgp1), and
does not place orders, but produces goods itself, and Bt: trend estimate.
delivers them directly to its own warehouse. In this
case the orders represent the volume of goods While forecasting the future demand using Holt’s
ordered from the production facility. The supply method, we minimize the mean-square error (MSE)
ARTICLE IN PRESS
D. Wright, X. Yuan / Int. J. Production Economics 113 (2008) 587–597 591
in order to determine the parameters l and g, based Customer Order Pattern One
50
on historical data. 45 Customer Order
40
2.5. BM+Brown’s: BM using Brown’s DES 35
forecasting method 30
25
When Brown’s DES method (Hanke and Reitsch, 20
15
1991, p. 128), is used, (7) is replaced by
10
EDt ¼ at&1 þ bt&1 (16) 5
0
where 1 10 19 28 37 46 55 64 73 82 91 100
At ¼ ab $ ORt þ ð1 & ab ÞðAt&1 Þ (17) Fig. 3. Retail customer demand pattern.
Table 4
While forecasting the future demand using Brown’s Cost definition
method, we minimize the MSE in order to determine
Cost name Variable Definition
the parameter, ab, based on historical data.
Inventory CostI Holding cost in units of dollars
3. Simulation experiments and results cost per unit item per unit time
period.
Shortage cost CostS Penalty cost for lacking goods
3.1. Simulation experiments procedure in units of dollars per unit item
per unit time period.
The purpose of our research is to investigate the Fix cost for CostF Fixed cost for each order in
impact of ordering policies and forecasting methods order units of dollars per order.
Purchase cost CostP Transaction based cost for
on the bullwhip effect. Therefore, we conduct
goods in units of dollars per
simulations in ordering policy space, i.e. (a,b) space, unit.
from (8), combined with the four forecasting
techniques, described above. (For simplicity we
use a ¼ aS.) Fig. 3 shows the retail customer
demand pattern generated using local trends mod- computed from the current stock minus the order
ified by random disturbances taken from a normal backlog; and the cost. The cost for each trading
distribution with mean 0 and variance 4. Fig. 4 partner is defined by its inventory level and
shows an example of the bullwhip effect that this transaction volume according to (21) and (22).
data creates with a ¼ 0.4, b ¼ 0.3. Table 4 lists related cost definitions. In (21), IsOrder
We collect three data series from each simulation is 1 if there is a current order from its customer;
run: the order volume; the inventory level, which is otherwise, IsOrder equals 0.
ARTICLE IN PRESS
592 D. Wright, X. Yuan / Int. J. Production Economics 113 (2008) 587–597
costt ¼ costI $ S t þ costS $ backLogOrdert + The variance of order rate, which has also been
þ costF $ IsOrder þ costP $ IOt (21) used by Chen et al. (2000, pp. 436–443) and Lee
et al. (2004, pp. 1875–1886).
where + The root mean square of the discrepancies
backLogOrdert ¼ backLogOrdert&1 þ ORt & St between the trading partners order and the retail
customers original demand.
backLogOrder0 ¼ 0 + The total cost according to (22).
X
100
Fig. 5 illustrates the procedure in the simulation
Total Cost ¼ cos tt (22) experiments, where ordering policy represented by a
t¼1
and b uses increments of 0.1 between 0.1 and 1.0.
For each pair of (a,b), we have four forecasting
3.1.1. Simulation in ordering policy space options: Sterman’s adaptive forecasting method, the
In order to evaluate simulation results system- moving average forecasting method, Holt’s method,
atically, we identify three performance measures of and Brown’s DES method. Fig. 5 summarizes
the model, to measure the extent of the bullwhip whether the bullwhip effect can be alleviated by
effect as follows. using Holt’s and DES compared with moving
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
X-Axis --Alpha--
using variables defined in Tables A1, A2, and A3 in which the ordering policy can be improved
the Appendix A. by using Holt’s or DES forecasting compared
It is also clear from Figs. 7 to 9, that when one of with Sterman’s adaptive forecasting. High
our metrics is improved, the other two are also values of a and b give higher values of variance,
improved. This was observed in general for other RMSE and cost, and are therefore inappropriate
ordering policies. We are therefore able to define an in practice, independent of the forecasting method
overall performance metric used.
Fig. 7 shows that the variance of order rate at the
varP þ rmseP þ costP
totalP ¼ (23) factory is alleviated by about 65% if Holt’s or DES
3 forecasting is used, for one specific ordering policy:
From the functions above, the smaller the costP, a ¼ 0.2 and b ¼ 0.3. Table 5 shows the variance
varP, rmseP, and totalP are, the better is the performance measure ‘‘varP’’; for all ordering
performance of BM+Holt’s and BM+Brown’s policies where Holt’s and DES are beneficial.
models over the other two models. The result of All the values are substantially o1, so the result
optimal ordering policy identification is given in in Fig. 7 applies to a range of ordering policies.
Fig. 6. The darker the grid, the better is the Tables 6 and 7 show similar results for RMSE and
performance; for instance, the value in the grid cost. The total performance measure, totalP, (23)
(0.2,0.4) is 0.456 that is better than that of grid can be found in Fig. 6, again indicating a similar
(0.4,0.5), which is 0.815. result. We can therefore conclude that Holt’s and
ARTICLE IN PRESS
594 D. Wright, X. Yuan / Int. J. Production Economics 113 (2008) 587–597
Y-Axis --Beta--
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.0
0.9
0.8
0.7
0.1
X-Axis --Alpha--
Fig. 6. Optimal ordering policy identification, with TotalP from Eq. (23).
400 10
5
200
0
0 Retailer Whole Distributor Factory
Customer Retailer Whole Distributor Factory Saler
Saler
BM 7.038542 13.423755 21.148637 29.970608
BM 104.64434 149.99374 322.40219 651.23542 1106.8314
BM+MA 104.64434 139.56251 304.88606 638.91504 1105.6556 BM+MA 6.6784208 13.227441 21.818073 31.367546
BM+Holts 104.64434 107.49647 126.47865 171.807 284.21646 BM+Holts 5.2482415 6.7339632 9.4617702 14.124073
BM+Browns 104.64434 108.71322 123.46605 191.10868 361.51595 BM+Browns 5.3108915 6.9778109 10.903824 16.997989
Fig. 7. Variance of order rate for a ¼ 0.2 and b ¼ 0.3. Fig. 8. RMSE for a ¼ 0.2 and b ¼ 0.3.
ARTICLE IN PRESS
D. Wright, X. Yuan / Int. J. Production Economics 113 (2008) 587–597 595
Cost Order
60000 70
BM 60 Retailer
50000 BM+MA Whole Saler
50 Distributor
BM+Holts
40000 Factory
BM+Browns 40
Customer
30
30000
20
20000
10
10000 0
1 9 17 25 33 41 49 57 65 73 81 89 97
0
Retailer Whole Distributor Factory Total Cost Fig. 12. Order rate using BM+Holt’s for a ¼ 0.2 and b ¼ 0.3.
Seller
40
Table 5
20 Variance performance measure report
0
VarP at Value varP at Value
1 9 17 25 33 41 49 57 65 73 81 89 97
(0.2,0.2) 0.329 (0.3,0.5) 0.519
Fig. 10. Order rate using BM for a ¼ 0.2 and b ¼ 0.3.
(0.2,0.3) 0.292 (0.3,0.6) 0.675
(0.2,0.4) 0.293 (0.4,0.5) 0.745
(0.2,0.5) 0.318 (0.5,0.2) 0.454
Order
(0.2,0.6) 0.466 (0.6,0.2) 0.499
120
(0.3,0.2) 0.699 (0.7,0.2) 0.659
Retailer
100 (0.3,0.3) 0.810 (0.8,0.2) 0.494
Whole Saler (0.3,0.4) 0.475
80 Distributor
Factory
60
Customer
40 3.2.2. Ordering policy analysis
Tables 5–7 give a consistent result that small
20
values of a such as a ¼ 0.2 show good performance
0 with Holt’s and Brown’s forecasting, compared with
1 9 17 25 33 41 49 57 65 73 81 89 97 higher values such as a ¼ 0.8. For the same value of
Fig. 11. Order rate using BM+MA for a ¼ 0.2 and b ¼ 0.3.
a, e.g. a ¼ 0.2 or 0.3, a medium value of b such as
b ¼ 0.4 is preferable. Therefore, slow adjustment on
the discrepancy of stock with a slightly larger
DES have a substantial impact on reducing the adjustment on the discrepancy of the supply line is
bullwhip effect independent of the performance a best practice for alleviating the bullwhip effect.
measure. For example, ordering policies such as (a ¼ 0.2,
ARTICLE IN PRESS
596 D. Wright, X. Yuan / Int. J. Production Economics 113 (2008) 587–597