You are on page 1of 21

Improved Base-Stock Policies Under Order Crossover

James R. Bradley Johnson Graduate School of Management Cornell University Ithaca, NY 14853-6201 jrb28@cornell.edu Phone: 607.255.9594 Fax: 607.254.4590 Lawrence W. Robinson lwr2@cornell.edu Phone: 607.255.4721 July 22, 2003

Abstract Order crossover occurs whenever replenishment orders do not arrive in the sequence in which they were placed. When order crossover is possible, Robinson et al. (2001) showed that its eect can be correctly taken into account by setting inventory levels using the inventory shortfall rather than the lead-time demand. For convenience, a heuristic base-stock level can be found by assuming that the shortfall is Normally distributed. This calculation, however, depends on the variance of the number of orders outstanding, which is cumbersome to compute and requires more information than may be available in practice. This paper makes the Normal approximation to the shortfall distribution tractable by computing bounds on the variance of the number of orders outstanding. Specically, we show that the variance of the number of orders outstanding is bounded above by the minimum of the lead-time mean, the lead-time variance, and the standard deviation of the lead time divided by 3. This easily-computed upper bound can be used in Normal approximations of the shortfall distribution to take order crossover into account, and so will improve signicantly upon policies set according to the lead-time demand distribution.

Introduction

In this paper, we consider the standard periodic review inventory model, where an order is placed at the start of each period to return the inventory position to some base-stock level S . When the replenishment lead time is stochastic, the probability distribution of lead-time demand the quantity of demand that occurs between the placement and receipt of an order is commonly used in calculating S . The lead-time demand distribution is a mixture, which depends on the p.m.f. of the lead time and the convolution of demand over multiple periods. Because this mixture is inconvenient (albeit straightforward) to compute, it is often approximated by the Normal distribution with the same mean and variance. Using the lead-time demand distribution to set the base-stock level is appropriate when orders always arrive in the same sequence in which they were placed. But if out-of-sequence order crossovers are possible, then it is no longer correct to use the lead-time demand distribution as the basis for setting S . Instead, the distribution for the inventory shortfall (the outstanding inventory still on order, plus the current periods demand) should be used. Robinson et al. (2001) show that the cost of incorrectly setting the base-stock level S from the lead-time demand distribution, rather than from the shortfall distribution, can be quite high even if crossovers are infrequent. The distribution for the shortfall, like the lead-time demand, is a mixture of demand over multiple periods, but with mixing probabilities derived from the distribution of the number of outstanding orders rather than from the lead time. Although the optimal base-stock level S can be computed from the aforementioned mixture, this computation is quite cumbersome in practice, and requires information about the entire lead-time and demand distributions, which may not always be available. For tractability, a Normal approximation of the lead-time demand distribution is commonly used to set the base-stock level. It is considerably more dicult to construct a Normal approximation for the shortfall distribution (in order to reect order crossovers) however, because its variance depends on the variance of the number of orders outstanding, whose computation still depends on the entire distribution of the lead time. 1

Our main contribution is to facilitate the use the Normal approximation to the shortfall distribution through the development of simple and tractable bounds and approximations for the variance of the number of orders outstanding. The resulting heuristic base-stock level requires no more information or computational eort than would be required for one based on the Normal approximation to the lead-time demand distribution. Specically, we prove that the variance of the number of orders outstanding is bounded above by the standard deviation of the lead time divided by 3. This implies that when the lead time is highly variable, the number of orders outstanding will be much less variable. In turn, the shortfall could be much less variable than the lead-time demand, so that inventory policies based on the lead-time demand could carry far too much safety stock. We also show that the variance of the number of orders outstanding is bounded above by both the mean and the variance of the lead time. When the lead-time distribution is compact, we develop a lower bound on the variance of the number of orders outstanding, which is again proportional to the standard deviation of lead time. For any xed mean and variance of the lead time, we show that the distribution of the number of orders outstanding is remarkably similar, regardless of its exact distribution. In particular, when the lead time distribution belongs to the family of binomial / Poisson / negative binomial distributions, we show that the upper bound is generally quite close to the true variance of the number of orders outstanding. We present an example that compares the performance of four heuristic policies with the optimal policy, where the optimal base-stock level S is based upon the actual shortfall distribution. In this example, two Normal shortfall-based heuristics outperform the common Normal lead-time demand approximation and a fourth heuristic based on the Poisson distribution. These results hold for all but the very highest critical fractiles, for which the skewness of the shortfall distribution is poorly matched by any Normal approximation. In the remainder of the paper, we rst summarize the relevant literature in Section 2. We then describe our inventory replenishment model and highlight the importance of the

variance of the number of orders outstanding in Section 3.

We compute bounds on, and

approximations to, the variance of the number of orders outstanding in Section 4. We evaluate the performance of our shortfall-based heuristic policies in Section 5, and conclude in Section 6. Although our bounds can be stated concisely, their proofs are long and in some cases tedious. They have therefore been relegated to the supporting working paper Im-

proved Base-Stock Policies Under Order Crossover: Proofs (Author-to-be-revealed, 2003), which is available from the rst authors web site, www.to_be_revealed.edu.

Literature Review

The mutually inconsistent assumptions that orders do not cross and that lead times are independent are made often in the literature. This inconsistency has generally been resolved in either of two ways. First, the possibility and consequence of order crossover has been assumed to be suciently small so that it can be ignored. Or second, a structure is invoked for the lead times that precludes order crossover. Examples of the second are the models by Kaplan (1970), Nahmias (1979), and Ehrhardt (1984). Song and Zipkin (1996b) construct a general model in which the replenishment process is Markovian. This work is most closely related to that of Zalkind (1976, 1978) and Robinson et al. (2001), who examine the eects of order crossover in a periodic review base-stock model. They show that the optimal order-up-to point S should be based on the distribution of the shortfall rather than the lead-time demand. The shortfall is dened to be the amount of inventory on order plus the current periods demand, and measures the dierence between S and the ending inventory level. Under a base-stock policy, the shortfall distribution depends on the distribution of the number of orders outstanding. Robinson et al. (2001) provide an iterative algorithm for calculating this distribution, based on the distribution of the lead time. Both they and Zalkind (1976) show that number of orders outstanding has the same mean, and the same or smaller variance, than the lead time. And nally, many authors have analyzed the eect of order crossover in continuous-review

(Q, r) models, including Hayya et al. (1995), Song and Zipkin (1996a), He et al. (1998), and Hayya et al. (2001).

The Model and Base Stock Approximations

After introducing our assumptions, notation, and model formulation, we will review the importance of setting the base-stock level from the shortfall distribution rather than the leadtime demand distribution. We also review the common heuristic of setting base-stock levels based on a Normal approximation of lead-time demand, and introduce similar approaches based on Normal and Poisson approximations to the shortfall distribution. Consider a periodic review model where Dt is the stochastic and stationary demand in any period t, and Lt is the stochastic and stationary lead time for an order of size Qt that is placed at the start of period t. We assume that Lt is measured in periods, is non-negative, and is integer-valued. We assume that Dt and Lt are jointly independent over time. And . nally, we dene fl = Pr{L = l} to be the p.m.f. of the lead time, with a c.d.f. of Fl . The rst event in a period is the possible arrival of outstanding orders, after which there is the opportunity to order. Under a base-stock policy (as long as the initial inventory is not too high), the order quantity Qt will equal the demand in the previous period, Dt1 . After the order is placed (and possibly received, if Lt = 0), then the demand Dt is realized, after which inventory holding and penalty costs are incurred. We assume that these costs are minimized at some known fractile r of the appropriate probability distribution. We let z denote a realization of a random variable Z . Furthermore Z = E(Z ) denotes
(L) to be the L-fold expected value, and 2 Z = Var(Z ) denotes variance. Also, we dene Z

convolution of a random variable Z . Dene SFt to be a random variable denoting the inventory shortfall at the end of period t; it will equal the total amount of inventory that has been ordered in or prior to period t but which has not arrived, plus the demand in period t, Dt . Note that SFt is invariant with the base-stock level S and that the ending inventory level is S SFt . If we dene the 4

random variable N to be the number of orders still outstanding after the very beginning of a period (i.e., after any earlier orders might arrive and a new order is placed), and let {gn } be its p.m.f., then the inventory shortfall SF is a mixture of n + 1 periods worth of demand, with mixing probabilities gn Pr {SF = sf } =
X n=0

where the + 1 reects the demand occurring in the most recent period. Both Robinson et al. (2001) and Zalkind (1976, 1978) show that if orders cannot strictly cross (e.g., when L is either deterministic or can assume one of two adjacent values), then N is identical in distribution to L and, moreover, the shortfall SF is identical in distribution to the demand over the lead time, LT D: Pr {LT D = ltd} =
X l=0

gn Pr D(n+1) = sf ,

(1)

In this case the critical fractile solutions with respect to LT D and SF are of course identical and optimal. But when orders can cross over one another, Robinson et al. (2001) show that incorrectly using the distribution of LT D to set the base-stock level S will typically result
2 2 2 in unnecessarily high inventory levels (and costs) because 2 L > N , and so LT D > SF .

fl Pr D(l+1) = ltd .

(2)

Robinson et al. (2001) provide a specic example where using the LT D rather than the SF distribution causes inventory cost to increase by more than 60% even though the probability of order crossover is less than 10%. In order to avoid the mixture calculations, the base-stock level is often approximated from a Normal approximation to the LT D distribution, . SLT D = LT D + LT D 1 (r) , (3)

where () is the unit normal c.d.f., and the mean and variance can be computed using the well-known result for the sum of a random number N + 1 of i.i.d. random variables D: LT D = (L + 1) D
2 2 2 2 LT D = (L + 1) D + D L .

(4) (5)

Despite the dangers of this approach outlined by Eppen and Martin (1988), it does oer two distinct advantages. First, the cumbersome mixture calculations in (2) are avoided. And second, there is no longer any need to know the entire distribution of L; only its rst two moments are needed in order to calculate LT D and 2 LT D in (4) and (5). One obvious improvement to this approach when order crossover is possible is to approximate the base stock using a Normal approximation to the shortfall SF rather than to LT D, . SSF = SF + SF 1 (r) , with SF = (N + 1) D
2 2 2 2 SF = (N + 1) D + D N .

(6)

(7) (8)

These computations require the rst two moments of the distribution of N , which Robinson et al. (2001) compute as: N = L X 2 N = Fl (1 Fl ) .
l=0

(9) (10)

Note that SF = LT D , which is easily calculated, and that the only dierence between
2 2 2 2 2 LT D and SF is that N in (8) replaces L in (5). Computing N in (10), however, requires

knowledge of the entire distribution of the lead time L. Thus the two major advantages of using a Normal approximation easier calculations and reduced data requirements are both lost when the shortfall distribution is used in place of the lead-time demand distribution. The primary goal of this paper is to develop a tractable and robust approximation for 2 N, so that the distribution for the shortfall SF can be approximated by the Normal distribution just as easily as can the lead-time demand distribution LT D. Thus we make computing a base-stock policy from the Normal shortfall approximation just as accessible to practitioners as the generally-familiar Normal lead-time demand approximation. Because the shortfall distribution is unimodal (Zalkind 1976), albeit not Normal, much of Eppen and Martins 6

concern with the Normal approximation is obviated. We will show in an example that our base-stock policy using an approximation to 2 N is superior to the Normal lead-time demand approximation and is not signicantly dierent than one based on the true variance 2 N. We also evaluate our shortfall approximation relative to another heuristic based on the optimal continuous-review base-stock policy for Poisson demand. In that case, Palm (1938) shows that the inventory shortfall will have a Poisson distribution with a mean of SF . Surprisingly, the shortfall distribution does not depend on the variance of either the lead time or the number of orders outstanding. Dene (|) to be the c.d.f. of a Poisson distribution with mean , and dene the base-stock level through the inverse c.d.f. . SP = min {S | (S |SF ) r} , (11)

which will be asymptotically optimal (with Poisson demand) as the period length goes to zero. Robinson et al. (2001) generalized this approach to the family of binomial / Poisson / negative binomial distributions, although that generalization required calculating 2 N. One limitation of this result is that, in practice, the inverse Poisson (or negative binomial) distribution is not as tractable as the inverse Normal distribution.

Bounds on the Variance of the Number of Orders Outstanding

In this section we develop upper and lower bounds on the variance of the number of orders
2 outstanding, 2 N , based on the rst two moments of the lead time distribution, L and L .

Although a wide variety of distributions {fl } can match these two moments, the distributions {gn } that they generate tend to be remarkable similar. For example, consider the three distinctly dierent lead time probability distributions shown in Figure 1; all have L = 2 L = 2. However, the three corresponding distributions of the number of orders outstanding {gn }, as shown in Figure 2, are remarkably similar to one another given their disparate origins, with variances 2 N in the relatively narrow range between 0.667 (for the two-point distribution) to 0.800 (for the uniform distribution). This relatively narrow range for 2 N engenders hope that 7

a reasonable approximation to it will facilitate good base-stock approximations, particularly


2 given that 2 N contributes to only one of two terms in SF , which in turn contributes to only

one of two terms in the Normal base-stock approximation.

4.1

An Upper Bound on 2 N

The problem of nding the distribution of the lead time {fl } that maximizes 2 N given the rst two moments L and 2 L can be formulated as the following quadratic program: (P1) max 2 N =
l=0 P P

l=0

Fl (1 Fl )

s.t.

fl = 1

l=0

l=0 P

l fl = L
l P

2 l2 fl = 2 L + L

Fl =

fk

l = 0, 1, ... l = 0, 1, ...

k=0

fl 0

Proposition 1 For any lead-time distribution with variance 2 L, L 2 N . 3 Proof. The details of this proof are available in the supporting working paper Improved Base-Stock Policies Under Order Crossover: Proofs (Author-to-be-revealed, 2003), which is available from the rst authors web site, www.to_be_revealed.edu. We show there that 2 N is maximized when the lead time L follows a quasi-uniform probability distribution, which generalizes the discrete uniform distribution by allowing both endpoints to have dierent mass than the middle points. We can tighten this upper bound of L / 3 by incorporating two other simple upper bounds. First, both Robinson et al. (2001) and Zalkind (1976, 1978) have shown that
2 2 2 N L for all distributions of lead time L. And second, it is trivial to show that N L :

2 N =

X l=0

Fl (1 Fl ) 8

X l=0

(1 Fl ) = L .

Combining these three bounds yields the tight upper bound of: 2 N L . 2 . = min L , L , 3 (12)

Figure 3 shows which of these three bounds is the tightest for various L and L . The newly-developed bound of L / 3 is the tightest for most cases, while L will be the tightest . for distributions with large coecients of variation cv = L /L 3, and 2 L will be the tightest when the lead time variability is quite small, L < 1/ 3. (Note that the scalloped shaded region represents those L that are infeasible due to noninteger means L of the p . integer-valued random variable L. To be exact, L (1 ), where = L bL c .)

One of the most striking aspects of Proposition 1, besides its simplicity, is that this upper

bound on the variance of the number of orders outstanding is proportional to the standard deviation of the lead time. Thus whenever the lead time is highly variable, the variance of the number of orders outstanding will be signicantly less than the variance of the lead time. The shortfall will in turn be less variable than the lead-time demand, and so basing the Normal approximation on the shortfall should be signicantly better than basing it on the lead-time demand distribution. Note also that 2 N depends only on the mean and variance of the lead time; using it within an approximation for the variance of the shortfall requires no more information than is required for the commonly-calculated variance of the lead-time demand. We will examine the tightness of this upper bound 2 N in Section 5.

4.2

Lower Bounds on 2 N

In Improved Base-Stock Policies Under Order Crossover: Proofs (Author-to-be-revealed, 2003), we show that 2 N can always attain its lowest possible value, (1 ), if arbitrarily long lead times are possible. However, if the lead-time distribution support is compact, we show in this section that we can obtain a tighter lower bound.

Proposition 2 For any L and 2 L , if the replenishment lead time L falls within the compact . range [L kl L , L + ku L ], then with k = max {kl , ku }, 2 N k L. k2 + 1 (13)

Proof. The details of this proof are also available in the supporting working paper Improved Base-Stock Policies Under Order Crossover: Proofs (Author-to-be-revealed, 2003). The proof is based on showing that the probability distribution that minimizes 2 N has no more than three points with positive probability mass. There are two very interesting conclusions from Proposition 2. First, note that the lower bound on the variance of the number of outstanding orders N , 2 N , is again a linear 2 function of the standard deviation of the lead time L, L , just as was the upper bound N found in Proposition 1. And second, the upper and lower bounds on 2 N are fairly narrow. Consider our ongoing example where L = 2 L = 2. Assume that we can safely impose the additional restriction L [0, 8]. (When L follows the Poisson distribution, for example, it will fall within this range 99.98% of the time.) Then kl = 2 and k = ku = 3 2, From (12), the upper bound on 2 and the lower bound on 2 N from (13) is 0.3156. N is n o p p 2/3 = 2/3 = 0.8165. So, in this example, imposing the very mild 2 N = min 2, 2, constraint L 8 is sucient to narrow the bounds on 2 N to the range [0.3156, 0.8165].

Performance Evaluation

Table 1 below summarizes the ve methods for setting the base-stock level S under evaluation. Each method employs a dierent representation of the shortfall distribution; the rst approach uses the exact distribution, while the next three use Normal approximations, and the nal approach uses the Poisson approximation. We are particularly interested in 2 the third heuristic, SSF c , which is introduced in this paper. Using the upper bound N to

2 approximate 2 N should provide a substantial improvement over using L , and is just as

convenient; it has the same low data requirements, and is equally simple to calculate.

10

Shortfall Distribution True Mixture

Normal Approximation

Normal Approximation Normal Approximation

Poisson Approximation

Base Variance Stock Used Comments 2 S N Optimal solution. Mixture distribution dicult to calculate. Requires complete knowledge of demand and lead time distributions. 2 SSF N 2 N is tedious to calculate. Requires complete knowledge of lead time distribution. 2 SSF N Proposed heuristic policy. Easy to c compute from L and 2 L. 2 SLT D L Commonly used variance is too high when orders can cross over. 2 L known initially. SP Asymptotically optimal with Poisson demand, as the period length 0.

Table 1: Optimal and Heuristic Policies We perform two numerical analyses in this section. In the rst, we examine the tightness of the upper bound 2 N for a variety of lead time distributions. In the second, we provide an example of the cost performance of the four heuristic policies relative to the optimal policy. As was shown in the previous section, the tightness of the upper bound 2 N depends on the particular distribution of the lead time. If the lead time follows a quasi-uniform distribution, then the upper bound will be very tight. Conversely, if the lead time follows a three-point distribution, then it will be much looser. Thus there is little hope of nding any universal performance results. Instead, we investigate the tightness of this upper bound for the family of negative binomial / Poisson / binomial distributions, corresponding to whether the variance of the lead time 2 L is greater than, equal to, or less than its mean L , respectively. The tightness of the upper bound 2 N is measured by the nonnegative relative parameter , 2 2 . = N 2 N . N (14)

This measure is graphed in Figure 4 for the Poisson distribution, and for several dierent values of the coecient of variation cv . In the worst case for the Poisson distribution ( 2 L = 11

2 1/3), the upper bound 2 N is almost 33% higher than the true variance N . However, is

generally small; the upper bound is within 5% of the true variance, as long as 2 L is outside of the range (0.05, 2.56). Note too that as 2 L becomes large it appears that approaches an asymptote of approximately 2.33%. Keeping the coecient of variation constant, Figure 4 also shows that is fairly small (< 15%) when the coecient of variation cv is small or very large (1/3, 1, 3). However, when 2 the coecient of variation is moderately large (cv = 2), the upper bound N can be almost 30% more than the actual variance 2 N . This is because neither of the upper bounds L / 3
2 2 nor L = L /cv are especially tight. Regardless, N is still a much closer estimate of N than

is 2 L . (In Figure 4, the intersections of the four cv curves with the Poisson curve identify where 2 L = L . Points to the left of these points correspond to the binomial distribution, while points to the right correspond to the negative binomial distribution.) Perhaps the most relevant performance measure of the upper bound 2 N is its eect on the base-stock level S set with a Normal approximation to the shortfall distribution, and the resulting cost. The heuristic that we propose using 2 N is to set . SSF SF 1 (r), c = SF + (15)

where SF is found by replacing 2 2 N in (8), N with its upper bound of . 2 2 2 2 SF = (N + 1) D + D N . (16)

The three alternate heuristic policies are SSF , SLT D , and SP , which are dened in (3), (6), and (11), respectively. The performance of the three Normal-based heuristic policies depends on two common factors. The rst is the degree to which the mixture-based shortfall distribution resembles the Normal distribution. The second factor is the relative magnitude of the two terms comprising the variance. If the average lead time (L ) or variance in demand ( 2 D ) is large, then the rst term in the variance will be much larger than the second, so that any dierences among

12

2 2 2 N , and L will be relatively insignicant. Conversely, if the rst term is small, then N,

the distinctions among these three estimates should be more pronounced. Consider again our ongoing example where the lead time L follows a Poisson distribution 2 with L = 2 N = 0.8165. The distribution of the L = 2, which yields an upper bound number of orders outstanding, N , (computed using the procedure in Robinson et al. 2001) is shown in Figure 2. In this case we nd that N = 2 and 2 N = 0.7715, so that the upper
2 bound 2 N exceeds N by about 5.8%. For computational tractability, we assume that period

demand Dt also has a Poisson distribution; we set D = 3.8885 so that the two terms in (16) have equal weight. Calculation (1) yields the skewed shortfall distribution shown in Figure 5. The optimal (S ) and four heuristic base-stock policies (SSF , SSF c , SLT D , and SP )

are shown in Figure 6 for critical ratios r [0.8, 1.0). The Poisson approximation SP has a variance of SF , which is by construction half of the true variance 2 SF = 23.331; thus SP is always signicantly lower than S . SSF and SSF c are very similar, and match the optimal

base-stock level S quite closely for r 0.96. For extremely high critical ratios, however, the skewness of the shortfall distribution causes the optimal base-stock level to exceed these two Normal-based heuristics by an increasing amount. Within this range, the higher variance of LT D osets the skewness of SF , and so SLT D provides a closer t to S . The consequences of these dierences in base-stock levels are shown in Figure 7, which shows the percentage increase in inventory cost from using a heuristic base-stock policy S rather than the optimal policy S , for critical ratios r [0.8, 1.0). As would be expected, the low base-stock level outperform SLT D except for the very extreme critical ratios r 99%. SP can be quite costly. SSF and SSF c yield costs within 1% of optimal for r 96%, and

Conclusions

Whenever stochastic lead times allow for the possibility of replenishment orders crossing over one another, the base-stock level should be calculated from the distribution of the shortfall rather than the distribution of the lead-time demand. Because both distributions are

13

mixtures, their derivations are quite involved, and depend upon the complete distributions of both demand and the lead time. Thus for expediency, a Normal approximation of the lead-time demand distribution is often used in its stead. When order crossover is possible, the Normal approximation of the shortfall distribution has heretofore been dicult to use because its variance depends on the variance of the number of orders outstanding, 2 N, the determination of which requires knowledge of the entire lead time distribution and the computation of an innite sum. The simple bounds that we develop in this paper serve two purposes: rst, they make the Normal approximation of the shortfall distribution tractable; and second, they highlight the importance of setting base-stock levels with the shortfall rather than the lead-time demand distribution. The upper and lower bounds on 2 N are linear functions of the standard deviation of the lead time, L , and the mean L . Thus setting a base stock with

a Normal approximation of the shortfall distribution using the upper bound 2 N requires no additional information or computation than that required for the traditional LT D ap proximation. Moreover, the upper bound, 2 N = L / 3, shows that the variance of N is proportional to the standard deviation of L. Thus when 2 L is large, the dierence be2 tween 2 N and L can be dramatic, which demonstrates the importance of using the shortfall

distribution rather than the lead-time demand distribution to set base-stock levels using shortfall facilitates more appropriate, and less costly inventory policies than does the traditional Normal approximation of lead-time demand. Our numerical examples imply that the skewness of the true shortfall distribution makes its approximation by any Normal distribution inappropriate as the critical fractile r approaches one. But apart from these extreme values, the Normal approximation of shortfall based on the upper bound is comparable to the one based on the true value of 2 N . Particularly when the variability in the number of orders outstanding is relatively large, these two approximations perform within a few percent of optimal.

14

References
[2003] Author-to-be-revealed. 2003. Improved base-stock policies under order crossover: proofs. Working Paper. [1984] Ehrhardt, R. 1984. (s,S ) policies for a dynamic inventory model with stochastic lead times. Oper. Res. 32 121-132. [1988] Eppen, G. D., R. K. Martin. 1988. Determining safety stock in the presence of stochastic lead time and demand. Management Sci. 34 1380-1390. [2001] Hayya, J. C., J. G. Kim, and X. He. 2001. On stochastic lead-time inventory models. MS&IS Working Paper 01-04, Pennsylvania State University. [1995] Hayya, J. C., S. H. Xu, R. V. Ramasesh, X. X. He. 1995. Order crossover in inventory systems. Sto. Mod. 11 279-309. [1998] He, X. X., S. H. Xu, J. K. Ord, J. C. Hayya. 198. An inventory model with order crossover. Oper. Res. 46 S112-S119.9 [1970] Kaplan, R. 1970. A dynamic inventory model with stochastic lead times. Management Sci. 16 491-507. [1979] Nahmias, S. 1979. Simple approximations for a variety of dynamic leadtime lost-sales inventory models. Oper. Res. 27 904-924. [1938] Palm, C. 1938. Analysis of the Erlang trac formulae for busy-signal arrangements. Ericson Technics 4 39-58. [2001] Robinson, L. W., J. R. Bradley, L. J. Thomas. 2001. Consequences of order crossover under order-up-to policies. Mfg. and Service Oper. Mgmt., 3 175-188. [1996a] Song, J. S., P. H. Zipkin. 1996. The joint eect of leadtime variance and lot size in a parallel processing environment. Management Sci. 42 1352-1363. 15

[1996b] Song, J. S., P. H. Zipkin. 1996. Inventory control with information about supply conditions. Management Sci. 42 1409-1419. [1976] Zalkind, D. 1976. Further results for order-level inventory systems with independent stochastic leadtimes. Technical Report 76-6 Department of Health Administration and Curriculum in Operations Research and Systems Analysis, University of North Carolina at Chapel Hill. [1978] Zalkind, D. 1978. Order-level inventory systems with independent stochastic leadtimes. Management Sci. 24 1384-1392.

16

70% fl 60% 50% 40% 30% 20% 10% 0% 0 1 2 3 4 5 6 l 7 Poisson Uniform Two-Point

Figure 1: Lead Time Distributions with L = 2 L = 2

70% gn 60% 50% 40% 30% 20% 10% 0% 0 1 2 3 4 5 6 n 7 Poisson Uniform Two-Point

Figure 2: Distributions on the Number of Orders Outstanding

17

L
L

L
1

2 L

0 0 1 2

Figure 3: Tightest Bounds on 2 N

35% 30% 25% 20% 15% 10% 5% 0% 0.1 1 10 100 1,000


L2

cv = 1/3 cv = 1 cv = 2 cv = 3 Poisson

10,000

Figure 4: Increases of the Upper Bound for the Binomial Family of Distributions 18

10

15

20

25

30

sf

35

Figure 5: Shortfall Distribution (L = 2, D = 3.8885)


34

30

26

SSF c

22 SLT D 18 S SP 14 80% 82% 84% 86% 88% 90% 92% 94% 96% 98% SSF r

100%

Figure 6: Optimal and Heuristic Base-Stock Levels (L = 2, D = 3.8885)

19

16% 14% 12% 10% 8% 6% 4% 2% 0% 80% 82% 84% 86% 88% 90% 92% 94% 96% SLT D SSF SSF c 98% r 100% SP

Figure 7: Relative Cost Performance of the Heuristics (L = 2, D = 3.8885)

20