Professional Documents
Culture Documents
www.elsevier.com/locate/omega
Abstract
A reduction in the inventory replenishment lead-time allows reducing safety stock requirements and improving customer
service. However, it might be accompanied by increased procurement costs because of premium charges imposed by suppliers,
or higher transportation costs. This paper studies a single-stage variable lead-time inventory system with lead-time dependent
procurement cost. Selection of the lead-time value represents finding the trade-off between benefits of lead-time reduction and
increase in the procurement cost. A model for joint optimization of inventory and procurement costs is developed. Numerical
studies are conducted to identify conditions under which lead-time reduction is favorable compared to procuring at the lowest
cost.
䉷 2006 Elsevier Ltd. All rights reserved.
0305-0483/$ - see front matter 䉷 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.omega.2006.04.009
878 C. Chandra, J. Grabis / Omega 36 (2008) 877 – 887
their disruptive impact on suppliers’ operations. These supplier and a manufacturer depending upon the supply
observations lead to the problem of finding the trade- chain power structure and the frame of commitment.
off between lead-time reduction and additional procure- Information about the external demand has two lev-
ment costs. The trade-off between purchasing cost and els of accuracy (not necessarily no information and
delivery lead-time from the manufacturing point is dis- completely accurate information). This level of accu-
cussed by Elhafsi [8] and Slotnick and Sobel [9] among racy depends upon a choice between using either early
others. These authors show that manufacturers are inter- commitment or delayed commitment. In the case of
ested in providing flexible pricing in response to more balance of power in the supply chain and, if accuracy
attractive delivery time conditions. of demand information improves quickly, the delayed
commitment is preferable. However, if improvement
1.2. Literature review of forecasting accuracy is small, the early commitment
is preferable. The author also provides an excellent
Ouyang and Wu [10] and Lan et al. [11] consider a summary of literature related to early commitment. His
continuous review re-order point policy with variable main observations are that existing models usually con-
lead-time. Costs associated with lead-time reduction are sider short planning horizons, two levels of information
represented by the lead-time crashing cost, which is in- accuracy or two procurement cost levels.
dependent of the order size. They consider the random Buyer–supplier interactions are also analyzed using
demand. The above-mentioned papers focus on compu- dual sourcing models (for instance, [16–18]), where the
tation of the optimal policies. Hariga and Ben-Daya [12] buyer allocates orders between suppliers with different
propose similar variable lead-time periodic review and characteristics including lead-time. The dual sourcing
base stock policies. Similarly, lead-time crashing has models are directed towards minimizing risk of suppli-
been investigated in two stage systems involving buyer ers’ non-performance while this paper focuses on set-
and supplier [13]. The integrated inventory model with ting the supply lead-time in the long-term collaboration
controllable lead-time is used in demonstrating that the framework.
buyer–suppliers system as a whole benefits from the
lead-time reduction while gains for individual partners 1.3. Contribution
are not distributed uniformly.
Burnetas and Gilbert [5] have analyzed the trade-off In this paper, a single stage, long horizon, reorder
between higher procurement costs against the benefit of point inventory system is studied; wherein the pro-
making ordering decisions using better demand related curement cost increases with decreasing lead-time. An
information. The authors develop an optimal ordering instance of such inventory problems is routine procure-
policy for a short life-cycle product. The procurement ment of a product for which there are multiple modes
cost increases along the product life cycle, while a pa- of transportation at different cost levels. The objective
rameter of the demand process (modeled as a Bernoulli of this paper is to find the lead-time value representing
process) is continuously updated as more accurate in- the trade-off between benefits of lead-time reduction
formation becomes available. The authors demonstrate and increase in the procurement cost. The lead-time
that ordering pattern depends upon characteristics of the reduction is expected to reduce the inventory cost be-
procurement cost function. cause of more accurate demand information and lower
There have been numerous investigations on early safety stock requirements. Yet, it increases the procure-
commitment policies, which are closely related to vari- ment cost because the supplier sets a higher price for
able lead-time inventory policies in the supply chain en- short lead-time orders, limited supplies force to seek
vironment (e.g., [6,14]). Under these policies, a supplier more costly alternative products or a more expensive
offers a discount, if an internal supply chain customer shipping mode is used to ensure the shorter lead-time.
places orders in advance. The customer increases its in- In order to describe this trade-off, a model for joint
ventory carrying risk by committing early (i.e. using the optimization of inventory and procurement costs with
longer lead-time) or loses an opportunity to procure at respect to the lead-time is developed. It is based on
the lower cost. The primary objective of early commit- the traditional continuous review re-order point model.
ment policies is finding a discount level optimizing the Demand is modeled as an autoregressive process. This
supply chain performance. model differs from the traditional model by inclu-
Early commitment policies under evolving forecast- sion of the lead-time dependent procurement cost. A
ing accuracy are investigated by Ferguson [15]. He procurement cost function is used to describe the de-
analyzes negotiation of an exchange price between a pendence of the procurement cost upon the lead-time.
C. Chandra, J. Grabis / Omega 36 (2008) 877 – 887 879
where = (1 − ) and the noise process t ∼ is the last observed demand value. The variance of the
N(0, ). lead-time demand is obtained as 2L = (2 /(1 − )2 )
L i 2
i=1 (1 − ) (see [22]). This expression is trans-
7. The procurement cost is a decreasing function of the
lead-time. It is calculated per unit purchased (Ca- formed to a form without summation for use in further
chon [6] also determines the procurement cost for computations
orders with a shorter lead-time on per unit purchased
basis. Lan et al. [11] calculate the lead-time crashing
2 L − 1 2
2L
−1
cost on per order basis, while the lead-time crash- 2L = L − 2 + 2 .
ing cost has both order cycle and per item specific (1 − )2 −1 −1
components according to Pan et al. [21]).
The total cost function uses widely used approximation
In this paper, the time unit is one year divided in weeks. for determining average inventory just before inventory
A week represents the period between consecutive up- replenishment and the number of backlogged orders.
dates of demand forecasts. The base measurement unit This approach does not explore impact of short-term de-
for the lead-time is also a week. Other time units can mand fluctuations characteristic to autoregressive pro-
be used, if necessary. cesses. Urban [20] has shown that by using the variable
re-order point, the average number of backlogged units
2.3. Model formulation under autoregressive demand is controlled in the same
way as under random demand. Therefore, accuracy of
The inventory management cost over the unit time the approximation is comparable in both situations. In
interval is comprised of ordering, holding and shortage order to assess the impact of short-term fluctuations on
costs. This traditional cost structure is augmented to in- the inventory level just before the replenishment arrives,
clude the purchasing cost expressed as a multiple of the a more elaborate expression for determining the inven-
lead-time dependent procurement cost c(L) and D units tory holding cost (IC) is
purchased over the unit time interval. The unit variable
cost is expressed as a sum of the variable procurement Q Q ∞
int(D/Q)
IC = v r + v r (st − YtL )f dYtL
cost and the added value v = v + c(L) because the 2 D 0
i=1
variable procurement cost has substantial impact on the
shelf value of the product (see [19, p. 44]). Q Q ∞
int(D/Q)
The total inventory and procurement cost is expressed = v r + vr (ŶtL + kL − YtL )
2 D 0
as i=1
the computationally more attractive approximation is Step 5: Repeat steps 3 and 4 until values of k, Q
used in further presentation. and L do not change, similar to using the simultaneous
procedure for finding k and Q only (see [19, p. 325]).
2.4. Model solving Step 6: Calculate the optimal total cost using expres-
sion (2).
In order to minimize the total cost, the total cost
function is differentiated with respect to each decision 2.5. Procurement cost function
variable. Partial derivatives obtained are set to zero, re-
sulting in following expressions used in finding optimal Solving model (2) requires knowing the procurement
values of decision variables cost function, which should be provided by a supplier.
In practice, suppliers are more likely to provide a few
rQ
pu (k) = , (4) discrete values of the procurement cost for specific lead-
B2 D times. Existing research is restricted mainly to using
just two levels of the procurement cost [6,14]. How-
2[AD + B2 v L Gu (k)]
Q= , (5) ever, it is plausible to have more than two levels, espe-
vr
cially if the procurement cost depends upon the mode of
and transportation used. One could use these discrete values
to conduct optimization by enumeration of alternatives.
Qr jc(L) B2 Gu (k)D
+D ∗ + kr+ h = 0, (6) Alternatively, the procurement cost function can be fit-
2 jL Q ted using the values provided. This approach is appeal-
where ing because the retailer can find any lead-time value to
be optimal from the range of plausible lead-time val-
jL jc(L) jL
h=v + L + c(L) . ues. Even though the supplier probably does not quote
jL jL jL this lead-time value, the retailer can attempt to agree
For purposes of further derivations, Qr/2+D is denoted with the supplier on providing it at the corresponding
by and kr + B2 Gu (k)D/Q is denoted by . purchasing cost level. Such situations occur if products
Evaluation of jc(L)/jL depends upon the choice of are purchased directly from manufacturers with limited
PCF. The partial derivative of L is found as capacity. Manufacturers often have options to alternate
their manufacturing schedules to balance manufactur-
jL 1 2 ing costs and delivery lead-time requirements.
= For example, the supplier offers L = 2, c(L) = 9 and
jL 2 2 (1 − )2
L=5, c(L)=7.5. The retailer finds the optimal lead-time
2L+1 ln() L equal to 1 with the corresponding purchasing cost equal
× 1− 1− .
−1 +1 to 9.5 per item. Both sides can attempt to settle on L=1.
In this paper two types of PCF are considered:
The following procedure is used to find values of the
1. linear PCF c(L) = b0 − b1 L;
decision variables if demand is a first order autoregres-
2. nonlinear PCF c(L) = b0 + b1 La .
sive process:
Step 1: Identify a range of lead-time values. The linear PCF implies a controlled growth of the pro-
Step 2: Initialize values of L and Q by setting L = 1 curement cost while the nonlinear PCF implies an un-
and Q to economic order quantity. controlled growth. The coefficient b1 determines the
Step 3: Calculate k using expression (4). Update Q speed of growth of the procurement cost. Coefficients
using expression (5). b1 and a should be selected in a way to obtain a declin-
Step 4: Numerically solve expression (6). If it has no ing function in the case of nonlinear PCF. That occurs
solutions, calculate the total cost using lead-time values if b1 > 0 and a < 0 or b1 < 0 and a > 0. For the latter
defined by endpoints of the range of lead-time values case, a = 1 yields linear PCF and a > 1 yields concave
and select the lead-time value yielding the lowest total PCF, which is less likely to occur in practice. Burnetas
cost. The endpoints of interval are analyzed to find the and Gilbert [5] have used linear and step-wise purchas-
optimal lead-time in cases when the total cost is not a ing cost functions. Ray and Jewkes [23] in their paper
convex function of the lead-time or expression (6) has on dependence of external customer demand upon the
no solution within the specified lead-time range (Section lead-time also express the product price as a linear de-
2.6 discusses properties of the total cost function). creasing function of the lead-time.
882 C. Chandra, J. Grabis / Omega 36 (2008) 877 – 887
2.6. Model analysis concave function, the minimum TC occurs at the end-
points of the lead-time range, similarly as for the linear
In order to gain preliminary insights into the influence PCF.
of lead-time on the total inventory and procurement
cost and to predict performance of the proposed solu- 3. Numerical results
tion procedure, dependence of the total cost function
3.1. Experimental design
upon the lead-time is analyzed. The impact of autocor-
relation on the lead-time demand variability is ignored Numerical studies are conducted to test the accuracy
to avoid too cumbersome derivations. Despite omitting of approximations used, to evaluate the proposed model
the impact of autocorrelation, the lead-time demand and to identify parameters of the inventory system and
variance is still a monotonically increasing function of properties of the demand process having the most pro-
the lead-time. The linear and nonlinear procurement found impact on the trade-off between the inventory cost
cost functions as defined in the previous section are and the procurement cost. The demand process is de-
analyzed. scribed using expression (1). Factors characterizing the
Given that the linear PCF is just a special case of the inventory system included in the experimental design
nonlinear power PCF, the model analysis is conducted are fixed ordering cost, added value and backlogging
for the general case at first. The first partial derivative penalty. Factors characterizing the demand process are
is found for analysis of extreme points of the total cost the coefficient of the autoregressive demand process
function, and the second partial derivative is found to and the signal-to-noise ratio. The signal-to-noise ratio,
analyze convexity properties of the total cost function. which describes variability of the demand process, is
The first partial derivative of the total cost according defined as the average demand divided by the demand
to the lead-time is ab1 La−1 + b1 (a + 21 )La−1/2 + standard deviation . Values of the considered experi-
√
(b0 + v) /2 L. The second partial derivative of the mental factors are given in Table 1. Values of other pa-
total cost according to the lead-time subsequently mul- rameters are set to D = 5000 (a relatively fast moving
tiplied by L−3/2 is a(a − 1)b1 La−1/2 + b1 (a − 41 )2 product) and r = 0.25. These values are chosen similar
to those often considered in the published literature. The
La − (b0 + v) /4.
values of the demand properties are chosen to cover both
In the case of linear PCF a = 1, the second partial
strongly and weakly correlated processes. The base pro-
derivative is clearly negative (after accounting for the
curement cost b0 is set to obtain equal values of c(Lmax )
sign change). Therefore, TC is the concave function
for all PCFs considered. The parameter a for the non-
with respect to the lead-time. Subsequently, the optimal
linear PCF is set equal to −1. The range of lead-time
lead-time is either the shortest one or the longest one.
values is 1 to 12. TC(Lopt , Q, k) and TC(Lmax , Q, k)
This effect can be perceived as dominance of either the
are calculated separately, for all 108 experimental cells
inventory cost or the procurement cost in the decision
for three PCFs. The difference between these two quan-
making process.
tities TC = TC(Lmax , Q, k) − TC(Lopt , Q, k) is used
The situation is more complex in the case of non-
in evaluating performance of the proposed model. It
linear PCF. The convexity properties of the total cost
shows the reduction of the total cost by using the opti-
function strongly depend upon the values of parameters
mal lead-time, compared to procuring at the lowest cost.
of inventory system and demand. In the general case,
Values larger than zero indicate that Lopt < Lmax and
the total cost function is neither convex nor concave. If
reduction of the inventory cost by using more accurate
b1 > 0 and a < 0, TC is convex for small values of L
demand information outweighs additional procurement
and it becomes concave as L increases due to rapid de-
expenses. The shorter lead-time reduces the inventory
cline of the power function. Location of inflection point
cost because the resulting smaller lead-time variance al-
strongly depends upon the parameter v (its large val-
lows carrying less safety stock and improving the cus-
ues causes inflection at smaller L). The inflection point
tomer service.
is also a global minimum of TC for the given range
of lead-time values because the first derivative found 3.2. Example
above is composed of monotonic functions. It can be
perceived as a point of trade-off between the inventory In order to illustrate inventory management under
cost and the procurement cost. autoregressive demand with lead-time dependent pro-
If b1 < 0 and a > 0, TC is likely to be concave. It curement cost, the following example is elaborated. The
is certainly concave for b1 < 0 and a > 1 . If TC is a parameters are set to the following values: A=100$ per
C. Chandra, J. Grabis / Omega 36 (2008) 877 – 887 883
Table 1
Experimental design parameters
Factors A v B2 1 /
Table 2
Sample optimization results for each iteration and at the endpoints of the lead-time range
Iterations k Q L TC L v
order, D = 5000 units/year, v = 40$/item, B2 = 1, r = noise ratio has the most significant impact on the trade-
0.25 per year, weekly demand is yt =38.4+0.6yt−1 +t , off between the inventory cost and the procurement cost
where t ∼ N(0, 47.5), Lmax = 20 weeks and procure- (distribution of TC is skewed to the left causing con-
ment cost per unit is determined as c(L)=9.583+5L−1 . cern for applicability of ANOVA. Therefore, ANOVA
The optimal result is found after five iterations was also conducted after applying the logarithmic
(Table 2) (the maximum number of iterations encoun- transformation to TC. The same results concerning
tered among other cases analyzed is 11). Variation of significance of factors were obtained, suggesting that
lead-time values over the iterations is small, which the procedure is sufficiently robust. The results for the
can be explained by relatively minor impact of Q and non-transformed case are reported because the averages
k on the optimal lead-time. The table also shows re- obtained are more informative). The added value is the
sults obtained at the endpoints of the lead-time range second most significant factor and the most significant
and using the integer-valued optimal lead-time. Min- out of all parameters for the inventory system consid-
imization of the total cost yields Q = 317, k = 2.07 ered. The backlogging penalty and the fixed ordering
and Lopt = 5.2. The optimized total inventory man- cost do not have significant impact on the trade-off. The
agement and purchasing cost is 62, 747$/year while second level interactions between factors are significant
TC(Lmax , Q, k) = 68, 006$/year. Thus, using the opti- only for combinations among v, 1 and /. The largest
mal lead-time results in cost savings of about 8%. absolute cost difference between using optimal L and
Lmax is observed for the high value of the signal-to-
3.3. ANOVA analysis noise ratio. The average absolute difference is relatively
small because in many cases Lopt =Lmax as can be seen
The ANOVA analysis shows whether the cost dif- in Table 4 , which reports fraction of the experimental
ference TC is statistically significant and highlights cells with TC > 0. These results are reported for the
factors that most substantially influence the trade-off most important factors. There are few cases when the
between procurement cost and lead-time reduction. optimal lead-time is smaller than Lmax in the case of
Results of the ANOVA analysis for each of PCFs the linear PCF. The fraction is substantially larger for
separately are reported in Table 3. The mean TC the nonlinear PCF with b1 = 5. The lead-time reduc-
column shows the average reduction in total cost for tion is almost never beneficial, if demand is weakly
low and high values of the factors (the mean value for autocorrelated because the speed of lead-time variance
combinations of both low values and both high values, increase is relatively low.
respectively, are reported in the case of second order The total cost also has been calculated for all ex-
interactions). These results suggest that the signal-to- perimental cells using the integer value of the optimal
884 C. Chandra, J. Grabis / Omega 36 (2008) 877 – 887
Table 3
F -Ratio with p -values given in parenthesis and mean cost difference estimates (for low and high levels) for each factor
Table 4
A fraction of experimental cells with TC > 0
v = 10 v = 20 v = 40 v = 10 v = 20 v = 40 v = 10 v = 20 v = 40
10 0.2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.8 0.00 0.00 0.00 0.00 0.00 0.83 0.00 0.00 0.00
5 0.2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
0.8 0.00 0.00 0.00 0.00 0.83 1.00 0.00 0.00 0.67
2 0.2 0.00 0.00 0.00 0.00 0.00 0.83 0.00 0.00 0.17
0.8 0.00 0.50 1.00 0.83 1.00 1.00 0.00 0.83 1.00
C. Chandra, J. Grabis / Omega 36 (2008) 877 – 887 885
85000 85000
μ /σ = 2
79000 79000
73000 μ /σ = 2
μ /σ = 5 73000
TC
TC
67000 67000 μ /σ = 5
61000 μ /σ = 10 61000 μ /σ = 10
55000 55000
0 3 6 9 12 0 3 6 9 12
(a) L (b) L
Fig. 1. TC according to noise: (a) linear procurement cost function; and (b) nonlinear procurement cost function, b1 = 5. The value added is
40$/unit and = 0.8.
85000 85000
v = 40
79000 79000
73000 v = 20 73000 v = 40
TC
TC
v = 10
67000 67000 v = 20
61000 61000 v = 10
55000 55000
0 3 6 9 12 0 3 6 9 12
(a) L (b) L
Fig. 2. TC according to added value; (a) linear procurement cost function; and (b) nonlinear procurement cost function. The signal-to-noise
ratio is 47.5 and = 0.8.
12 73000 12 73000
->
9 9
-> 67000 67000
TC
TC
6 6 <-
L
Fig. 3. Lead-time and total cost according to coefficient of: (a) liner procurement function and; (b) nonlinear procurement function. The value
added is 40$/unit, the signal-to-noise ratio is 47.5 and = 0.8.
lead-time. The average relative difference between this 3.4. Characteristics of the total cost function
total cost and TC(Lopt , Q, k) is of order 10−5 and the
maximum absolute difference observed is of order 10−3 . Figs. 1 and 2 show the total cost according to the
Therefore, using the integer-valued optimal lead-time lead-time for various levels of the signal-to-noise ratio
results in negligible additional costs. However, this re- and the added value, respectively. The figures confirm
sult depends upon forecasts updating period. that the previously identified convexity properties of the
Similarly, the total cost with the inventory holding total cost function according to the lead-time also hold,
cost determined using expression (3) has been calcu- if the impact of the demand autocorrelation is accounted
lated for all experimental cells. Previously found op- for in calculating the lead-time demand variance.
timal values of the decision variables have been used. Results depend upon the speed of increase of the
Expression (3) has been evaluated using simulation and procurement cost. For the particular PCFs, the speed is
the total cost is averaged over 50 replications. The av- defined by b1 . Fig. 3 shows Lopt and TC according to b1
erage absolute difference between the total cost calcu- with the factors fixed at their upper values. In the case
lated using the inventory cost approximation as in (2) of the linear PCF, the shift from Lopt = 1 to Lopt = Lmax
and using expression (3) is 103$ and the maximum is occurs abruptly as suggested by the analytical analysis.
1514$, while the total cost TC is of order 104 . The width It occurs when the procurement cost at Lopt = 1 has
of 95% confidence interval of the simulated total cost increased by more than 44% relative to b0 . The shift is
estimate is in the range between 18$ and 1288$. gradual in the case of the nonlinear PCF.
886 C. Chandra, J. Grabis / Omega 36 (2008) 877 – 887
[14] Gilbert SM, Ballou RH. Supply chain benefits from advanced [21] Pan JCH, Lo MC, Hsiao YC. Optimal reorder point inventory
customer commitments. Journal of Operations Management models with variable lead time and backorder discount
1999;18(1):61–73. considerations. European Journal of Operational Research
[15] Ferguson ME. When to commit in a serial supply chain with 2004;158:488–505.
forecast updating. Naval Research Logistics 2003;50:1–20. [22] Chandra C, Grabis J. Application of multiple-steps forecasts to
[16] Ryu SW, Lee KK. A stochastic inventory model of dual sourced restrain the bullwhip effect. European Journal of Operational
supply chain with lead-time reduction. International Journal of Research 2005;166(2):337–50.
Production Economics 2003;81/82:513–24. [23] Ray S, Jewkes EM. Customer lead time management when both
[17] Tyworth JE. Transportation’s role in the sole-versus dual- demand and price are lead time sensitive. European Journal of
sourcing decision. International Journal of Physical Distribution Operational Research 2003;153:769–81.
and Logistics Management 2000;30:128–44. [24] Chandra C, Grabis J. On trade-off between inventory and
[18] Ruiz-Torres AJ, Mahmoodi F. The optimal number of suppliers procurement costs under seasonal demand. Proceedings of
considering the costs of individual supplier failures. Omega, in the 34th annual decision sciences institute meeting; 2003,
press. November 22–25, Washington, USA. p. 1939–44.
[19] Silver EA, Pyke DF, Peterson R. Inventory management and
production planning and scheduling. New York: Wiley; 1998.
[20] Urban TL. Reorder level determination with serially-
correlated demand. Journal of the operational Research Society
2000;51(6):762–8.