European Journal of Operational Research 173 (2006) 617–636 www.elsevier.com/locate/ejor Production, Manufacturing and Logistics Quantifying the bullwhip effect in a supply chain with stochastic lead time Jeon G. Kim a, Dean Chatfield b, Terry P. Harrison c, Jack C. Hayya a,* a c Department of Supply Chain and Information Systems, The Smeal College of Business Administration, Pennsylvania State University, 303 Beam Building, University Park, PA 16802, USA b Department of Management Science and Information Technology, Virginia Tech, 1007 Pamplin Hall (0235), Blacksburg, VA 24061, USA Department of Supply Chain and Information Systems, The Smeal College of Business Administration, Pennsylvania State University, 509-L BAB, University Park, PA 16802, USA Received 12 August 2003; accepted 10 January 2005 Available online 31 March 2005 Abstract In a recent paper, Dejonckheere, Disney, Lambrecht, and Towill [European Journal of Operational Research 147 (2003) 567] used control systems engineering (transfer functions, frequency response, spectral analysis) to quantify the bullwhip effect. In the present paper, we, like Chen, Ryan, Drezner, and Simchi-Levi [Management Science 46 (2000) 436], use the statistical method. But our method extends Dejonckheere et al. and Chen et al. in that we include stochastic lead time and provide expressions for quantifying the bullwhip effect, both with information sharing and without information sharing. We use iid demands in a k-stage supply chain for both. By contrast, Chen et al. provide lower bounds using autoregressive demand for information sharing and for information not sharing (with zero safety factor for stocks). Dejonckheere et al. validate Chen et al.Õs results for a 2-stage supply chain without information sharing, using both autoregressive and iid normally distributed demands. We estimate the mean and variance of lead-time demand (LTD) from historical LTD data, rather than from the component period demands and lead time. Nevertheless, we also calculate the variance amplification like Chen et al., but with gamma lead times. With constant lead times, which Chen et al. used, our method yields lower variance amplification. As for the effect of information, we find that the variance increases nearly linearly in echelon stage with information sharing but exponentially in echelon stage without information sharing. Ó 2005 Elsevier B.V. All rights reserved. Keywords: Supply chain management; Stochastic lead times; Bullwhip effect; Information sharing; Demand during lead time * Corresponding author. Tel.: +1 8148651461; fax: +1 8148632381. E-mail addresses: [email protected] (J.G. Kim), [email protected] (D. Chatfield), [email protected] (T.P. Harrison), [email protected] (J.C. Hayya). 0377-2217/$ - see front matter Ó 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.ejor.2005.01.043 618 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 1. Introduction 1.1. Background We shall not review the literature on the bullwhip effect, as Dejonckheere et al. (2003) have done an excellent job of that. Our object is to quantify the bullwhip effect with and without information sharing. By the bullwhip effect, we mean variance amplification of demand up a supply chain. By information sharing, we mean that customer demand at the lowest node of the supply chain is immediately transmitted to all the upstream nodes. (ÔInformation sharingÕ is what Chen et al. (2000, p. 439) call Ôcentralized demand informationÕ.) Like Dejonckheere et al. and Chen et al., we use order-up-to policies. In these papers lead time was constant, and lead-time demand (LTD) was estimated from its component lead time (L) and period demand (D). In our paper, L is stochastic and LTD is estimated from past observations of LTD. Let D and L be random variables. Then the demand during the lead time is L1 X X t ¼ Dt þ Dtþ1 þ þ DtþL1 ¼ Dtþj : ð1AÞ j¼0 Eq. (1A) is the conventional definition of LTD, in that the lead time is forward, but it is not ‘‘physically realizable’’ in trying to estimate Xt from past data at time t. In that situation, the appropriate definition is1 L X Xt ¼ Dtj ; ð1BÞ j¼1 because the definition in Eq. (1A) is not ‘‘physically realizable.’’ Nevertheless, the two definitions are mathematically equivalent in terms of a priori statistical analysis. As Eqs. (1A) and (1B) are mirror images of each other, we can just substitute the subscript j in (1B) by +j to achieve Eq. (1A). Now when both D and L are random, X is a random sum of random variables, and demand during lead time is the convolution (Ord, 1972, pp. 64–66; Mood et al., 1974, p. 186; Bagchi et al., 1986, p. 180) of demand rate and lead time. The LTD can also be said to be a Ôcompound distributionÕ of demand rate and lead time (Bagchi et al., 1984). Now suppose that D and L are statistically independent. Then Var ðX Þ ¼ r2X ¼ lL r2D þ l2D r2L and EðX Þ ¼ lX ¼ lL lD ; ðlL ; r2L Þ ð2AÞ ðlD ; r2D Þ are the theoretical mean and variance of the lead time and where variance of the demand per unit time. But if the lead time were fixed, then Var ðX Þ ¼ r2X ¼ Lr2D and EðX Þ ¼ lX ¼ LlD : the theoretical mean and ð2BÞ 1.2. Organization of the paper In Section 2, we quantify the bullwhip effect with stochastic lead time and with deterministic lead time, as a special case. We show how subtle differences in estimation could lead to dramatically different variance amplifications. In Section 3, we compare our results for variance amplification for information sharing versus information not sharing. In that section we also present comparisons with Chen et al. and Dejonckheere et al., using our method of estimating the order-up-to level. We provide two appendices that contain the analytical proofs for the material presented in Section 2. The long Appendix A contains Lemmas 1 and 2 and the attendant propositions needed for their proofs. Appendix B repeats the independently and identically distributed (iid) derivations in the text, but for autoregressive demand. 1 We are grateful to an anonymous referee for pointing out that our original use of Eq. (1A) was not physically realizable and that the logical expression for estimation is of the form in Eq. (1B). J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 619 2. Quantification of the bullwhip effect We quantify the bullwhip effect in an environment of stochastic demand and lead times. We show that the bullwhip effect increases in approximately linear fashion with supply chain echelon, k, when customer demand information is shared and exponentially with echelon (or as a k-power function of mean lead time, lL) when customer demand information is not shared. 2.1. Assumptions and notation (1) Lead time: Let Lt(k) be the lead time of the order placed at the beginning of period t at stage k. Assume that lead times at any stage and at the different stages are iid. In other words, Lt(k) are identical and independent for k = 1, 2, . . . , and for t = 1, 2, . . . , with mean lL and variance r2L . (2) Demand rate: Period demands, D1, D2, . . . , Dt, . . . are stationary and iid, with mean lD and variance r2D . (We use iid period demands, because the iid structure provides a basic building block for correlated period demands, as we show in the forthcoming analysis. Also, see, for example, Chen et al. (2000, Theorem 2, p. 44).) (3) Statistical independence: D and L are independent. (4) The length of the moving average, p: Following Chen et al. (2000) and Dejonckheere et al. (2003) we concentrate on the more practical case where p P L. (5) Sequence of events: at the beginning of each period, t, the inventory manager observes the inventory level and places an order, Qt. After the order is placed, customer demand, Dt, is observed. Let lX(k) and r2X ðkÞ be the mean and variance of lead-time demand at stage k. Then lX(k) = lX and ¼ r2X , because Lt(k) is iid for all k and t. Now define r2X ðkÞ Qkt = order quantity at stage k, S kt = the order-up-to level at stage k, st(k) = sample standard deviation of lead-time demand at stage k, Xt(k) = lead-time demand (LTD, also called Ôdemand during lead timeÕ) at stage k, zk = standardized normal variable for setting safety stocks at stage k. To further investigate Var(Qt) when the lead time is stochastic, denote the mth moment of X about l by ðmÞ m lX ¼ EðX lX Þ ; ð3Þ with the superscript m not to be confused with k, the stage number in the supply chain. 2.2. Subtle differences We show that subtle differences in the way estimates are made when updating the order-up-to-level could lead to marked differences in the magnitude of the bullwhip effect. To illustrate, let k = 1, z = 0, L be deterministic, and D iid. Calculating moving averages of demand rate and multiplying that average by L to estimate mean demand during lead time leads to (Chen et al., 2000, p. 438) Var ðQÞ 2L 2L2 ¼1þ þ 2 ; Var ðDÞ p p ð4Þ 620 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 whereas using our method of calculating moving averages of the demand during lead time (not the demand rate) leads to Var ðQÞ 2 2L ¼1þ þ 2 ; Var ðDÞ p p ð5Þ a drop in variance amplification of the order of L. However, both estimation methods are legitimate, with Chen et al.Õs perhaps the more common, because it may be easier to keep two sets of data files (D and L), rather than three (D, L, and X). Note that the difference between the two methods is that in the Chen et al. construction ðX t ¼ L:Dt Þ the lead time, L, is explicitly multiplicative (understandably so, because L is PL1 Pp Dtiþj constant), whereas in our model X t ¼ j¼0 p i¼1 , L is implicitly additive. Thus, the difference between Eqs. (4) and (5). Chen et al. (2000, pp. 437, 440) use the following construction: S t ¼ LDt þ z sLet ; ð6Þ where sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi Pp 2 ðe Þ ti i¼1 sLet ¼ cL;q p is the standard deviation of the forecast error over the lead time, with et ¼ Dt Dt the one-period-ahead forecast error; cL,q a constant function of lead time, L; q the autoregressive correlation; and p the number of observations in the moving average. Chen et al. (2000, Theorem 2, p. 438) show that the estimate in Eq. (6) leads to Eq. (4). But in the present paper, we use the alternative estimate of the order-up-to level, S t ¼ X t þ z st ðX Þ; ð7Þ where Pp i¼1 Xt ¼ p X ti PL1 Pp ¼ j¼0 i¼1 Dtiþj p ¼ L1 X Dtþj ; j¼0 and p 1 X 2 ðX t X t Þ ; p 1 t¼1 " # p X p X L1 L1 X X 1 2 ¼ ðDtiþj Dtþj Þ þ ðDtiþj Dtþj ÞðDtiþm Dtþm Þ : p 1 i¼1 j¼0 i¼1 j6¼m s2t ðX Þ ¼ Let k = 1, z = 0, L deterministic and D iid. This is a special case of Lemma 1 that we present later. Like in Chen et al. (2000, p. 438), we begin with Qt ¼ X t X t1 þ Dt1 ; but we take moving averages of demand during lead time, X, rather than of L multiplied by the moving average of period demand. Thus, 1 1 Qt ¼ ½X t1 X tp1 þ Dt1 ¼ Dt1 þ Dt þ þ DtþL2 Dtp1 Dtp DtpþL2 þ Dt1 : p p J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 621 Hence, 2L Var ðDÞ 2 þ Var ðDÞ þ Cov ½ðDt1 þ Dt þ DtþL3 DtþL2 Þ; Dt1 2 p p 2 2L ¼ Var ðDÞ 1 þ þ 2 for p P L p p 2 for p < L: ¼ Var ðDÞ 1 þ p Var ðQt Þ ¼ ð8Þ 2.3. k-Stage supply chain We consider k echelons in the supply chain: retailer, wholesaler, distributor, factory, and tiers of suppliers. We shall put k as a superscript or in parenthesis to denote kth stage. 2.3.1. When information on customer demand is shared In a paper on Ôinformation enrichmentÕ, Mason-Jones and Towill (1997, p. 138) say that ‘‘. . . market information notoriously suffers from delay and distortion as it moves through the supply chain,’’ with two main remedies being the reduction of uncertainty and the reduction of lead time. Here we deal with the first remedy: reduction of variance amplification, where the information on customer demand is shared by all stages in the supply chain, so that each stage can utilize this demand information for setting up its order-up-to level and its order quantity. Therefore, when the information on customer demand is shared, the order-up-to level at stage k is S kt ¼ X t ðkÞ þ zk st ðkÞ; ð9Þ which is the same as Eqs. (2A) and (2B) in Chen et al. (2000, p. 572) and Dejonckheere et al. (2003, p. 437). We also a use p-period moving average (MA(p)) to estimate means and variances of lead-time demand, Xt, based on observations from the previous p periods. Let stage zero be the customer in the supply chain. Since the demand at stage k is the orders from stage (k 1), the order quantity at stage k is k k k1 k1 1 1 0 Qkt ¼ S kt S kt1 þ Qk1 t1 ¼ S t S t1 þ S t1 S t2 þ þ S tkþ1 S tk þ Qtk : ð10Þ Since stage zero is the customer, let Q0tk ¼ Dtk . For convenience and without loss of generality, assume that zi = z, for all i. Then the order quantity at stage k is Qkt ¼ k k k X 1X 1X X t1kþi ðiÞ X t1pkþi ðiÞ þ Dtk þ z ðstkþi ðiÞ stk1þi ðiÞÞ: p i¼1 p i¼1 i¼1 ð11Þ Hence, the variance of the order quantity at stage k is Var ðQkt Þ ¼ k k 1 X 2 X Var ðX ðiÞÞ þ Cov ðX t1kþi ðiÞ; X t1kþj ðjÞÞ t1kþi p2 i¼1 p2 i<j þ k k 1 X 2 X Var ðX ðiÞÞ þ Cov ðX t1pkþi ðiÞ; X t1pkþj ðjÞÞ þ Var ðDtk Þ t1pkþi p2 i¼1 p2 i<j þ z2 k X i¼1 Var ðstkþi ðiÞ stk1þi ðiÞÞ þ 2z2 k X i<j Cov ½stkþi ðiÞ stk1þi ðiÞ; stkþj ðjÞ stk1þj ðjÞ 622 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 ! ! k k k X X X 2 2 2 Cov X t1kþi ðiÞ; X t1pkþi ðiÞ þ Cov X t1kþi ðiÞ; Dtk p p i¼1 i¼1 i¼1 ! ! k k k X X X 2z 2 þ Cov X t1kþi ðiÞ; ½stkþi ðiÞ stk1þi ðiÞ Cov X t1pkþi ðiÞ; Dtk p p i¼1 i¼1 i¼1 ! k k X X 2z Cov X t1pkþi ðiÞ; ½stkþi ðiÞ stk1þi ðiÞ p i¼1 i¼1 ! k X ½stkþi ðiÞ stk1þi ðiÞ : ð12Þ þ 2z Cov Dtk ; i¼1 We simplify Eq. (12) to the following lemma, whose proof is given in Appendix A. Lemma 1. When information on customer demand is shared, the variance of the order quantity at stage k is VarðQkt Þ ð3Þ 2 2kðk 1Þ kþ1 2k 2 2kz lX kz 2ðlL 1Þ ð3Þ ð3Þ 2 2 ½lD þ lD rD ¼ 1þ þ lL þ l rD þ 2 rX þ 2 p p2 3 p p rX prD D p1 " # " # ð4Þ ð4Þ z2 p 2 lX 2 2 ðk 1Þz2 ðp2 3p þ 3Þ lX ðp2 6p þ 3Þ 2 þ r : r þ 3 ð13Þ p3 ðp 1Þ r2X p ðp 1Þ X 2 p3 r2X p2 X 2 Corollary. When lead time is deterministic, Eq. (13) reduces to Var ðQkt Þ ¼ ð3Þ 2 2kðk 1Þ kþ1 2k 2 2kz LlD 2 1þ þ L þ r þ r D p p2 3 p2 X p2 rX n o 2 3 ð4Þ 2 L 3ðL 1Þr4D þ lD kz 2ðL 1Þ z p 2 2 ð3Þ ð3Þ ½lD þ lD r2D þ 4 3 þ lD 2 r2X 5 prD p1 p p 2 r2X n o 2 3 ð4Þ 4 z2 4p 2 L 3ðL 1ÞrD þ lD 2 25 2 rX þ p3 p r2X 2 n o 2 3 ð4Þ 4 ðk 1Þz2 4ðp2 3p þ 3Þ L 3ðL 1ÞrD þ lD ðp2 6p þ 3Þ 2 5 rX : þ p3 ðp 1Þ p3 ðp 1Þ 2 r2X ð14Þ As we see from Eqs. (13) and (14), the variance of the order quantity at stage k is amplified with the amplification mostly linear in k, except for quadratic and cubic terms in some system parameters. Thus, information sharing of customer demand per se does not eliminate the Bullwhip Effect. This reinforces what Chen et al. (2000, p. 442) have said. However, if the manager at each stage does not update the order-up-to level, i.e., if ki S ki ti ¼ S t1i ; i ¼ 0; 1; . . . ; k 1; then the order quantity is the same as the previous demand, i.e., Qkt ¼ Dtk . Hence, in this case, there will be no bullwhip effect, with Var(Qk) = Var(D), which conforms to Chen et al. and Dejonckheere et al. J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 623 2.3.2. Simple supply chain: One retailer and one immediate upstream agent As a special case, let us consider a simple supply chain, with only one retailer and one immediate upstream agent: this is the case of k = 1. If in Eq. (12) we put k = 1, that equation reduces to Var ðQt Þ ¼ Var ðDt1 Þ þ 1 1 2 Var ðX t1 Þ þ 2 Var ðX t1p Þ þ z2 Var ðst ðX Þ st1 ðX ÞÞ þ Cov ðDt1 ; X t1 Þ p2 p p 2 2 Cov ðDt1 ; X t1p Þ þ 2z Cov ðDt1 ; ðst ðX Þ st1 ðX ÞÞÞ 2 Cov ðX t1 ; X t1p Þ p p þ 2z 2z Cov ðX t1 ; ðst ðX Þ st1 ðX ÞÞÞ Cov ðX t1p ; ðst ðX Þ st1 ðX ÞÞÞ; p p ð15Þ and Lemma 1 becomes 2 2 2 2 z 2ðlL 1Þ ð3Þ ð3Þ 2 ½lD þ lD rD l Var ðQt Þ ¼ 1 þ rD þ 2 rX þ p p prD D p1 2z ð3Þ z2 p 2 ð4Þ 2 4 þ 2 lX þ 2 l 2 rX : p rX p3 X p 2rX ð16Þ Now if, as in Chen et al. (p. 437), we used the autoregressive process, Dt ¼ l þ qDt1 þ et ; ð17Þ and their estimation method, then Eq. (15) would reduce to a result similar to Chen et al.Õs (p. 438): 2L 2L2 2L p 2 þ 2 ð1 q Þ rD þ 2z 1 þ Var ðQt Þ ¼ 1 þ Cov ðDt1 ; ðst st1 ÞÞ þ z2 Var ðst st1 Þ: p p p ð18Þ A difference between our Eq. (18) and Chen et al.Õs is that Chen et al. use the standard deviation of forecast error, set, in setting safety stocks, whereas we use the standard deviation of lead-time demand, st. With our method of estimation and with autoregressive demand, Eq. (15) would become (see Appendix B), " ( )# 2 ð1 qpLþ1 Þð1 qL Þ 2 L 2qð1 qÞ þ 2q2 ð1 qL1 Þ þ qpLþ1 ð1 qL Þ2 VarðQt Þ ¼ 1 þ þ 2 r2D 2 p ð1 q2 Þð1 qÞ p ð1 q2 Þ ð1 q2 Þð1 qÞ 2 þ 2z CovðDt1 ; ðst st1 ÞÞ þ CovðX t1 ; ðst st1 ÞÞ þ z2 Varðst st1 Þ: ð19Þ p Setting q = 0, z = 0, Eq. (19) reduces to our previous Eq. (8). 2.3.3. When information on customer demand is not shared Mason-Jones and Towill (1997) speak of a ‘‘seamless supply chain,’’ where everyone in the supply chain gets the most recent market sales data. Utilization of this information, they say improves the responsiveness of the supply chain and reduces the bullwhip effect. Güllü (1997) demonstrates that information sharing in a two-echelon allocation model results in lower order-up-levels and diminished system costs. Lee et al. (2000) echo the same. With one supplier and multiple identical retailers, Cachon and Fisher (2000) find that supply chain costs are 2.2% lower on average with full information, with a maximum difference of 12.1%. Mitra and Chatterjee (2004) show through numerical examples that the optimal 624 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 expected total cost with demand information known throughout a one-warehouse, two-retailer, system is always lower than that with information withheld. Dejonckheere et al. (2004) employ control systems engineering to show that an information-enriched supply chain reduces the bullwhip effect, especially at the higher echelons. This is supported in a simulation study by Chatfield et al. (2004), where information sharing reduces variance amplification and protects a supply chain against ‘‘cascading failures’’, as occurs in an electrical power grid. When the information on customer demand is not shared, the kth stage relies on the order quantity from the (k 1)th stage. Because the demand at the kth stage is the order Qk1 , lead-time demand at stage k (we t now use Y instead of X to distinguish lead-time demand when customer demand information is not shared) is Y t ðkÞ ¼ L1 X Qk1 tþi : ð20Þ i The order quantity at stage k is 1 1 k1 Qkt ¼ S kt S kt1 þ Qk1 t1 ¼ Y t1 ðkÞ Y t1p ðkÞ þ Qt1 þ zðst ðY ðkÞÞ st1 ðY ðkÞÞÞ: p p ð21Þ For simplicity, let st(Y(k)) = st(k). Then, 1 1 Var ðY t1 ðkÞÞ þ 2 Var ðY t1p ðkÞÞ þ z2 Var ðst ðkÞ st1 ðkÞÞ p2 p 2 2 k1 Cov ðQk1 þ Cov ðQk1 t1 ; Y t1 ðkÞÞ t1 ; Y t1p ðkÞÞ þ 2z Cov ½Qt1 ; ðst ðkÞ st1 ðkÞÞ p p 2 2z 2 Cov ðY t1 ðkÞ; Y t1p ðkÞÞ þ Cov ½Y t1 ðkÞ; ðst ðkÞ st1 ðkÞÞ p p 2z Cov ½Y t1p ðkÞ; ðst ðkÞ st1 ðkÞÞ: p Var ðQkt Þ ¼ Var ðQk1 t1 Þ þ ð22Þ Eq. (22) reduces to (16) for k = 1. Using Eq. (16), we obtain the following iterative approximation of the bullwhip effect when information on customer demand is not shared. Lemma 2. After starting with Eq. (16) for k = 1, we can approximate the variance of the order quantity at stage k for information not sharing by the following iterative equation: For k P 2, 2 2lL 2 2 2 2zlk1 z2 lk1 p 2 ð4Þ 2 4 ð3Þ k k1 L L Var ðQt Þ ffi 1 þ þ 2 Var ðQt Þ þ 2 lD rL þ 2 l þ l 2 rX p p p3 X p p p rX X 2r2X zlk1 2ðlL 1Þ ð3Þ ð3Þ ½lD þ lD r2D : þ L lD ð23Þ p1 prD The proof of Lemma 2 is in Appendix A. As we see from Eq. (23), the variance of the order quantity at stage k can be approximated as functions of the kth exponent of mean lead time when information on customer demand is not shared. This is with stochastic lead time and MA(p). As a special case, consider z = 0, a case that Chen et al. and Dejonckheere et al. explored. Then from Eqs. (22) and (23), 2 2lL 2 k Var ðQt Þ ¼ 1 þ þ 2 Var ðQk1 Þ þ 2 l2D r2L : ð24Þ t p p p J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 625 3. Numerical comparisons 3.1. Variance amplification: With and without information sharing In Table 1, we give an example of the behavior of the variance in a k-stage supply chain with and without information sharing. We have customer demand (k = 0) with a standard deviation of 20 and see that as a result of MA(p = 15) forecasting and the consequent periodic updating of the order-up-to level that the variance of the orders at the factory (k = 4) becomes inflated, rising to 52.5 with information sharing and 179.0 without information sharing. The agreement with the simulation results in Table 1 is heartening, except for Ôfactory,Õ where the error is nine percent, which we attribute to sampling variability. Fig. 1 illustrates the nearly linear behavior of the variance in echelon stage, k, with information sharing (is) and the exponential growth in k without it (nis). Merkuryev et al. (2003) produce a similar figure via simulation. 3.2. Comparison with Chen et al. and Dejonckheere et al. Chen et al. (2000, p. 441, and verified in Dejonckheere et al., 2003, p. 581, for k = 1) give a lower bound (LB) for the amplification of customer demand with deterministic lead time and information sharing, Pk Pk 2 Var ðQk Þ 2 i¼1 Li 2ð i¼1 Li Þ is: P1þ þ : ð25Þ 2 Var ðDÞ p p Table 1 Comparison of information sharing vs. no information sharing: Standard deviation and amplitude ratio (AR) Customer Retailer Wholesaler Distributor Factory Std. dev.: information sharing (is) ARis Std. dev.: no information sharing: (nis) ARnis 20 (20.18) 30.0 (30.21) 38.7 (37.22) 46.1 (43.34) 52.5 (47.72) 1 1.50 1.29 1.19 1.14 20 (19.77) 30.0 (30.47) 51.1 (51.85) 92.8 (92.04) 179.0 (162.32) 1 1.50 1.70 1.82 1.93 The bold numbers in parentheses are simulation results. AR ¼ std:dev:ðoutputÞ std:dev:ðinputÞ ; D N(50,400); is = information sharing; L gamma with mean 4 and variance 4; nis = no information sharing. Variance 200 nis 100 is 0 0 1 2 3 4 k Fig. 1. Variance amplification: no information sharing (nis) vs. information sharing (is). 626 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 Table 2 A comparison: Information sharing (z = 2) Our Lemma 1 (stochastic L) Our Lemma 1 (deterministic L) Chen et al.Õs LB Retailer k = 1 Wholesaler k = 2 Distributor k = 3 Factory k = 4 2.25 1.04 1.68 3.76 1.06 2.64 4.83 1.11 3.88 6.89 1.18 5.41 Table 3 A comparison: No information sharing (no safety stock case, z = 0) Our Lemma 2 (stochastic L) Our Lemma 2 (deterministic L) Chen et al.Õs LB Retailer k = 1 Wholesaler k = 2 Distributor k = 3 Factory k = 4 1.39 1.17 1.68 1.85 1.37 2.82 2.38 1.60 4.74 3.01 1.87 7.97 With no information sharing (and deterministic lead time), the LB would be k Y Var ðQk Þ 2Li 2L2i P þ 2 : nis: 1þ Var ðDÞ p p i¼1 ð26Þ Using L = 4, p = Tm = 15, k = 1, the LB given to variance amplification, reported by Chen et al. and Dejonckheere et al. would be 1.68 (deterministic lead time) > 1.04 (deterministic lead time) in our Table 2. Tables 2 and 3 compare our results with Chen et al.Õs for information sharing and information not sharing. We repeat that we and Chen et al. used different updating methods, both legitimate, and the comparison is merely intended to highlight the significant effect of different estimating procedures. 4. Summary We quantify the bullwhip effect in a serial supply chain, using periodic forecast updating schemes. We employ an (R, S) inventory system, where the review period R = 1, and where S is the order-up-to level, that is, S is revised every period. The revisions would be based on the realizations of the stochastic demand and lead time and also the lead-time demand that obtain at the different echelons of the supply chain. We conclude, as others before us have, that the bullwhip effect is caused by human intervention and by disruptions in information flow in the supply chain. As expected, lead times, whether deterministic or stochastic, exacerbate the bullwhip effect by inflating the variance of the demand at the upstream echelons. However, lead-time variability contributes further to the bullwhip effect. We illustrate how seemingly similar estimating methods could yield dramatically different variance amplifications and we reinforce the results in Chen et al. (2000) and Dejonckheere et al. (2003) that the bullwhip effect is due in part to demand forecasting. And as reported in earlier literature, the bullwhip effect is attenuated when information is shared. Beyond that, we find that attenuation reducing an exponential bullwhip when information is not shared to a linear effect when it is shared. Acknowledgment The authors acknowledge with thanks the support from the Center of Supply Chain Research, Smeal College of Business Administration, Pennsylvania State University. J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 627 Appendix A More detailed proofs are posted on http://www.psu.edu.personal/jch/EJORAPPENDICES.pdf Before proving Lemma 1, we begin with Propositions 1–7, which we need. Proposition 1. The following holds: Var ðst ðX Þ st1 ðX ÞÞ ¼ 1 ðp 2Þ ð4Þ 2 4 l r : X p3 p2 X 2r2X ðA:1Þ Proof. Since Var ðst ðX Þ st1 ðX ÞÞ ¼ Var ðst ðX ÞÞ þ Var ðst1 ðX ÞÞ 2 Cov ðst ðX Þ; st1 ðX ÞÞ; ð4Þ where ðp 1Þ2 lX ðp 1Þðp 3Þr4X 4p3 r2X ¼ Var ðst1 ðX ÞÞ ðKendall and Stuart, 1977, p. 296Þ; Var ðst ðX ÞÞ ¼ ðA:2Þ and Cov ðst ðX Þ; st1 ðX ÞÞ ¼ 1 Cov ðs2t ðX Þ; s2t1 ðX ÞÞ ðKendall and Stuart, 1977, p. 247Þ; 4r2X ðA:3Þ we obtain ð4Þ Var ðst ðX Þ st1 ðX ÞÞ ¼ ðp 1Þ2 lX ðp 1Þðp 3Þr4X 1 2 Cov ðs2t ðX Þ; s2t1 ðX ÞÞ: 2 3 2rX 2p rX ðA:4Þ Now consider Cov ðs2t ðX Þ; s2t1 ðX ÞÞ in (A.4) and, for convenience, let s2t ðX Þ ¼ s2t . Then, Cov ðs2t ; s2t1 Þ ¼ ðp3 4p2 þ 6p 3Þ ð4Þ p2 6p þ 3 4 lX rX : p3 ðp 1Þ p3 ðA:5Þ Finally, we prove Proposition 1 by putting Eq. (A.5) into (A.4), obtaining Var ðst ðX Þ st1 ðX ÞÞ ¼ 1 p 2 ð4Þ 2 4 l r : 2r2X p3 X p2 X Proposition 2. The following will hold: Cov ½X t1p ; st1 ðX Þ ¼ Cov ½X t1 ; st ðX Þ ¼ 1 ð3Þ l ; 2prX X ðA:6Þ and Cov ½X t1p ; st ðX Þ ¼ Cov ½X t1 ; st1 ðX Þ ¼ 0: ðA:7Þ Proof. Since Cov ½X t1p ; st1 ðX Þ ¼ 2r1X Cov ½X t1p ; s2t1 ðX Þ (Kendall and Stuart, 1977, 247–248), consider Cov ½X t1p ; s2t1 ðX Þ. Because 628 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 p X ðp 1Þ X t1p s2t1 ðX Þ ¼ X t1p ðX t1i X t1 Þ 2 i¼1 Pp1 2 2X 3t1p 2X t1p i¼1 X t1i ¼ þ X t1p p p i¼1 Pp1 Pp1 Pp1 2 4X t1p j¼1 X t1j 2X t1p i¼1 X t1i j6¼i X t1j X 3t1p þ p p p P Pp1 Pp2 2 2 X t1p p1 X 2X X 2X X t1p t1p j¼1 t1j j¼1 t1j j<k t1j X t1k þ þ ; þ p p p ðA:8Þ p1 X X 3t1p X 2t1i ð3Þ Cov ½X t1p ; s2t1 ðX Þ ¼ 1p lX . Hence, Cov ½X t1p ; st1 ðX Þ ¼ 1 ð3Þ l : 2prX X ðA:9Þ Since ðp 1Þ X t1p s2t ðX Þ ¼ X t1p p X X 2ti 2X t1p X t1p Pp 2 j¼1 X tj p 2 i¼1 X ti p i¼1 þ Pp þ X t1p 2X t1p Pp i¼1 X ti Pp j6¼i X tj p Pp j6¼k X tj X tk p ; Cov ½X t1p ; s2t ðX Þ ¼ 0: Therefore; Cov ½X t1p ; st ðX Þ ¼ 0: ðA:10Þ ðA:11Þ Similarly we can show that Cov ½X t1 ; st ðX Þ ¼ 1 ð3Þ l 2prX X and Cov ½X t1 ; st1 ðX Þ ¼ 0: Proposition 3 1 2ðlL 1Þ ð3Þ ð3Þ 2 ½lD þ lD rD : Cov ½Dt1 ; st ðX Þ st1 ðX Þ ¼ lD 2prD p1 ðA:12Þ Proof. Since Cov ½Dt1 ; st ðX Þ ¼ 2r1D Cov ½Dt1 ; s2t ðX Þ, consider Cov ½Dt1 ; s2t ðX Þ. Because E½s2t ðX Þ ¼ E½s2t1 ðX Þ, then Cov ðDt1 ; ðs2t ðX Þ s2t1 ðX ÞÞÞ ¼ E½Dt1 s2t ðX Þ E½Dt1 s2t1 ðX Þ: Also we can express ðp 1ÞDt1 s2t ðX Þ as ðp 1ÞDt1 s2t ðX Þo ¼ p L1 X L1 Xp XL1 p1 p1 X X Dt1 i¼1 j¼0 D2tiþj þ Dtiþj Dtiþk p p i¼1 j¼0 k6¼j p p p p L1 L1 X L1 1XXX 1XXX Dtiþj Dtlþj Dtiþj Dtlþk : p i¼1 l6¼i j¼0 p i¼1 l6¼i j¼0 k6¼j ðA:13Þ Similarly, ðp 1ÞDt1 s2t1 ðX Þ can be obtained by replacing t with t 1 of D in the summation at the RHS of Eq. (A.13). Hence, after some manipulation, we obtain J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 629 E½Dt1 s2t ðX Þ E½Dt1 s2t1 ðX Þ ¼ E½EfDt1 s2t ðX ÞjLg E½EfDt1 s2t1 ðX ÞjLg 1 ð3Þ 2ðlL 1Þ ð3Þ ½l þ lD r2D : ¼ lD p pðp 1Þ D Therefore, 1 2ðlL 1Þ ð3Þ ð3Þ 2 ½lD þ lD rD : Cov ½Dt1 ; st ðX Þ st1 ðX Þ ¼ lD 2prD p1 Proposition 4. The following holds: 2 2 2 2 z 2ðlL 1Þ ð3Þ ð3Þ 2 ½lD þ lD rD Var ðQt Þ ¼ 1 þ rD þ 2 rX þ l p p prD D p1 2z ð3Þ z2 p 2 ð4Þ 2 4 þ 2 lX þ 2 l 2 rX : p rX p3 X p 2rX ðA:14Þ Proof. Consider Cov (Dt1,Xt1) and Cov (Dt1,Xt1p): " Cov ðDt1 ; X t1 Þ ¼ E E Dt1 Lt1 X !# Dt1þi jL 2 ðlD Þ lL ¼ r2D : ðA:15Þ i¼0 Because we assume that L 6 p, " " Ltp1 Cov ðDt1 ; X t1p Þ ¼ E E Dt1 X i¼0 ## Dt1pþi L l2D lL ¼ 0: ðA:16Þ From Proposition 2, Cov ðX t1p; ðst ðX Þ st1 ðX ÞÞÞ ¼ 1 ð3Þ l ; 2prX X ðA:17Þ and Cov ðX t1 ; ðst ðX Þ st1 ðX ÞÞÞ ¼ 1 ð3Þ l : 2prX X ðA:18Þ From Proposition 3, 1 2ðlL 1Þ ð3Þ ð3Þ 2 ½lD þ lD rD : Cov ½Dt1 ; st ðX Þ st1 ðX Þ ¼ lD 2prD p1 Now " Cov ðX t1 ; X t1p Þ ¼ E E Lt1 X i¼0 Therefore, Lt1p Dt1þi X j¼0 !# l2L l2D ¼ 0: Dt1pþj L 2 2 2 2z ð3Þ Var ðQt Þ ¼ 1 þ rD þ 2 r2X þ z2 Var ðst ðX Þ st1 ðX ÞÞ þ 2 lX p p p rX z 2ðlL 1Þ ð3Þ ð3Þ 2 ½lD þ lD rD : þ lD prD p1 ðA:19Þ ðA:20Þ By applying Proposition 1 to (A.21), we prove Proposition 4. h ðA:21Þ 630 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 Proposition 5. The following holds: k X i<j k X kðk 1Þ kþ1 2 lL Cov ðX t1kþi ðiÞ; X t1kþj ðjÞÞ ¼ Cov ðX t1pkþi ðiÞ; X t1pkþj ðjÞÞ: rD ¼ 2 3 i<j ðA:22Þ Proof. We can express X in terms of D. For convenience, we simplify the notation to Xt1k+i PLi 1 (i) = Xt1k+i and Xt1k+j (j) = X . Let L (i) = L and L (j) = L . Then X ¼ t1kþi t1k+j t1k+i i t1k+j j m¼0 Dt1kþiþm and PLj 1 X t1kþj ¼ n¼0 Dt1kþjþn . Hence, X t1kþi X t1kþj ¼ Li 1 X Dt1kþiþm Lj 1 X ðA:23Þ Dt1kþjþn : n¼0 m¼0 Now consider j > i. When j = i + 1, the i can take (k 1) different values, i.e., i = 1, 2, . . . , k 1. For these (k 1) cases, Eq. (A.23) can be expressed as ðLi 1ÞD2 þ ½Li Lj ðLi 1ÞDa Db ; where a and b are arbitrary integers such that a 5 b. If we continue this procedure, then when j = i + (k 1), the i can take only one value, i.e., i = 1. For this case, Eq. (A.23) can be expressed as ðLi ðk 1ÞÞD2 þ ½Li Lj ðLi ðk 1ÞÞDa Db : Since E½X t1kþi ðiÞE½X t1kþj ðjÞ ¼ l2L l2D , k X i<j kðk 1Þ kþ1 2 lL Cov ðX t1kþi ðiÞ; X t1kþj ðjÞÞ ¼ rD : 2 3 Similarly we can show that k X i<j kðk 1Þ kþ1 2 lL Cov ðX t1pkþi ðiÞ; X t1pkþj ðjÞÞ ¼ rD : 2 3 Proposition 6. The following holds: Cov k X k X X t1kþi ðiÞ; ðstkþi ðiÞ stk1þi ðiÞÞ i¼1 ! ¼ i¼1 and Cov k X X t1kpþi ðiÞ; i¼1 k X k ð3Þ l ; 2prX X ! ðstkþi ðiÞ stk1þi ðiÞÞ ¼ i¼1 k ð3Þ l : 2prX X Proof. For notational convenience, let sr+i(i) = sr+i. Then, Cov k X k X X t1kþi ðiÞ; ðstkþi ðiÞ stk1þi ðiÞÞ i¼1 ¼ Xk i¼1 i¼1 Cov ðX t1kþi ; stkþi st1kþi Þ þ ! Xk i6¼j Cov ðX t1kþi ; stkþj st1kþj Þ: ðA:24Þ ðA:25Þ J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 631 Now consider Cov (Xt1k+i, stk+i). Since ðp 1ÞX t1kþi s2tkþi Pp 2X 3t1kþi 2X t1kþi m¼2 X 2tkþim ¼ þ X t1kþi p p m¼2 Pp Pp Pp 2 4X X tkþim 2X t1kþi m¼2 X tkþim n6¼m X tkþin t1kþi m¼2 p p Pp P 3 2 2 X t1kþi X t1kþi m¼2 X tkþim 2X t1kþi pm¼2 X tkþim þ þ þ p p p Pp 2X t1kþi m<n X tkþim X tkþin ; þ p p X X 3t1kþi X 2tkþim ðA:26Þ by the way similar to Eqs. (A.8) and (A.9), Cov ðX t1kþi ; stkþi Þ ¼ 1 ð3Þ l : 2prX X ðA:27Þ Next consider Cov (Xt1k+i, st1k+i). Since ðp 1ÞX t1kþi s2t1kþi ¼ X t1kþi p X X 2t1kþim 2X t1kþi Pp 2 m¼1 X t1kþim p Pp m¼1 þ þ X t1kþi Pp X t1kþi Pp 2 m¼1 X t1kþim p 2X t1kþi n6¼m X t1kþim X t1kþin p m¼1 X t1kþim Pp n6¼m X t1kþin p ðA:28Þ ; then similar to Eqs. (A.10) and (A.11) Cov ðX t1kþi ; st1kþi Þ ¼ 0: ðA:29Þ Therefore, k X Cov ðX t1kþi ; stkþi st1kþi Þ ¼ i¼1 k ð3Þ l : 2prX X ðA:30Þ Also consider k X Cov ðX t1kþi ; stkþj st1kþj Þ ¼ i6¼j k X Cov ðX t1kþi ; stkþj Þ i6¼j k X Cov ðX t1kþi ; st1kþj Þ: i6¼j From the relationships in Eqs. (A.26)–(A.29), we obtain k X Cov ðX t1kþi ; stkþj Þ ¼ i6¼j k kðk 1Þ ð3Þ X lX ¼ Cov ðX t1kþi ; st1kþj Þ: 4prX i6¼j Therefore, k X i6¼j Cov ðX t1kþi ; stkþj st1kþj Þ ¼ 0: ðA:31Þ 632 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 Hence, by Eqs. (A.30) and (A.31), Cov k X k X X t1kþi ðiÞ; ðstkþi ðiÞ stk1þi ðiÞÞ i¼1 ! i¼1 Also, in the same way, we can show that Cov k X ¼ X t1kpþi ðiÞ; i¼1 k X ðstkþi ðiÞ stk1þi ðiÞÞ k ð3Þ l : 2prX X ! ¼ i¼1 k ð3Þ l : 2prX X Proposition 7. The following holds: k X Cov stkþi ðiÞ stk1þi ðiÞ; stkþj ðjÞ stk1þj ðjÞ i6¼j ðk 1Þ p 2 ð4Þ 2 4 ðk 1Þ p2 3p þ 3 ð4Þ p2 6p þ 3 4 l 3 r : l 2 rX þ ¼ p3 X p p3 ðp 1Þ X p ðp 1Þ X 2r2X 2r2X ðA:32Þ Proof. For notational convenience, let sr+i(i) = sr+i. Then, k X Cov stkþi ðiÞ stk1þi ðiÞ; stkþj ðjÞ stk1þj ðjÞ i6¼j ¼ k X Cov ðstkþi ; stkþj Þ i6¼j k X Cov ðstkþi ; stk1þj Þ þ i6¼j k X k X Cov ðstk1þi ; stk1þj Þ i6¼j Cov ðstkþj ; stk1þi Þ: i6¼j Since Cov (stk+i, stk+i+m) = Cov (st, st+m) = Cov (st, stm) = Cov (stk+i, stk+im), we obtain k X Cov stkþi ðiÞ stk1þi ðiÞ; stkþj ðjÞ stk1þj ðjÞ i6¼j ¼ 2ðk 1Þ½Cov ðst ; st1 Þ Var ðst Þ þ 2½Cov ðst ; st1 Þ Cov ðst ; stk Þ: Now consider Cov (st,stk). From Eqs. (A.3) and (A.5), we obtain " # 1 ðp kÞðp3 4p2 þ 6p 3Þ ð4Þ p4 ðk þ 7Þp3 þ ð7k þ 9Þp2 ð9k þ 3Þp þ 3k 4 lX rX : Cov ðst ; stk Þ ¼ 2 2 2 4rX p3 ðp 1Þ p3 ðp 1Þ ðA:33Þ Using Eqs. (A.2), (A.3), (A.5) and (A.33), we obtain k X Cov stkþi ðiÞ stk1þi ðiÞ; stkþj ðjÞ stk1þj ðjÞ i6¼j ðk 1Þ p 2 ð4Þ 2 4 ðk 1Þ p2 3p þ 3 ð4Þ p2 6p þ 3 4 l 3 r : l 2 rX þ ¼ p3 X p p3 ðp 1Þ X p ðp 1Þ X 2r2X 2r2X J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 633 Proof of Lemma Pk1. By using Propositions Pk1–7, we prove Lemma 1. We consider the terms in Eq. (12). First, consider i¼1 Var ðX t1kþi ðiÞÞ and i¼1 Var ðX t1kpþi ðiÞÞ. We obtain k k k 1 X 1 X k 2 1 X 2 Var ðX ðiÞÞ ¼ r ðiÞ ¼ r ¼ Var ðX t1kpþi ðiÞÞ: t1kþi p2 i¼1 p2 i¼1 X p2 X p2 i¼1 ðA:34Þ By Proposition 5, we obtain k 2 X kðk 1Þ kþ1 2 Cov ðX t1kþi ðiÞ; X t1kþj ðjÞÞ ¼ lL rD p2 i<j p2 3 ¼ k 2 X Cov ðX t1pkþi ðiÞ; X t1pkþj ðjÞÞ: 2 p i<j By definition, Var ðDtk Þ ¼ r2D . By Proposition 1, " # ð4Þ k X kz2 p 2 lX 2 2 2 Var ðstkþi ðiÞ stk1þi ðiÞÞ ¼ 2 rX ; z p3 r2X p 2 i¼1 ðA:35Þ ðA:36Þ and by Proposition 7, z2 k X Cov stkþi ðiÞ stk1þi ðiÞ; stkþj ðjÞ stk1þj ðjÞ i6¼j ¼ ðk 1Þz2 p 2 ð4Þ 2 4 ðk 1Þz2 p2 3p þ 3 ð4Þ p2 6p þ 3 4 l r l r þ : p3 X p2 X p3 ðp 1Þ X p3 ðp 1Þ X 2r2X 2r2X From Eq. (A.20), k k X X 2 Cov X ðiÞ; X t1pkþi ðiÞ t1kþi p2 i¼1 i¼1 ðA:37Þ ! ¼ 0: ðA:38Þ Now consider the covariance between X and D: ! k k X 2 2X Cov X t1kþi ðiÞ; Dtk ¼ Cov ðX t1kþi ; Dtk Þ; p p i¼1 i¼1 where X t1kþi ¼ Li 1 X Dt1kþiþm : m¼0 Hence, when i = 1, Cov ðX t1kþi ; Dtk Þ ¼ r2D . For i = 2, 3, . . ., k, we obtain Cov (Xt1k+i, Dtk) = 0. Hence, ! k X 2 2 Cov X t1kþi ðiÞ; Dtk ¼ r2D : p p i¼1 ðA:39Þ From Eq. (A.24) in Proposition 6, ! k k X X 2z kz ð3Þ Cov X t1kþi ðiÞ; ½stkþi ðiÞ stk1þi ðiÞ ¼ 2 lX : p p rX i¼1 i¼1 ðA:40Þ 634 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 From Eq. (A.16) from Proposition 4, ! k X 2 Cov X t1pkþi ðiÞ; Dtk ¼ 0: p i¼1 ðA:41Þ From Eq. (A.25) in Proposition 6, ! k k X X 2z kz ð3Þ Cov X t1pkþi ðiÞ; ½stkþi ðiÞ stk1þi ðiÞ ¼ 2 lX : p p rX i¼1 i¼1 ðA:42Þ By Proposition 3, k X 2z Cov Dtk ; ½stkþi ðiÞ stk1þi ðiÞ ! i¼1 ¼ kz 2ðlL 1Þ ð3Þ ð3Þ ½lD þ lD r2D : lD prD p1 Hence, from Eqs. (A.35)–(A.44), we prove Lemma 1. ðA:43Þ h Proof of Lemma 2. First consider EðQkt Þ. Since Qkt ¼ S kt S kt1 þ Qk1 t1 , 0 EðQkt Þ ¼ EðQk1 t1 Þ ¼ EðQtk Þ ¼ EðDtk Þ ¼ lD : ðA:44Þ Now consider Var (Yt1(k)) and Var (Yt1p(k)). By Eq. (A.44), k1 2 2 t1 Qk1 Var ðY t1 ðkÞÞ ¼ Var sumLi¼0 t1þi ¼ lL Var ðQt1 Þ þ lD rL ¼ Var ðY t1p ðkÞÞ: Also, " Cov ðQk1 t1 ; Y t1 ðkÞÞ ¼E E Qk1 t1 Lt1 X Qk1 t1þi Lt1 i¼0 Similar to Eq. (A.16), " ( k1 Cov ðQk1 t1 ; Y t1p ðkÞÞ ¼ E E Qt1 L1 X i¼0 Similar to Eq. (A.20), " ( Cov ½ðY t1 ðkÞ; Y t1p ðkÞÞ ¼ E E Lt1 X i¼0 !# 2 k1 lL E½Qk1 t1 ¼ Var ðQt1 Þ: ¼ L: )# E½Qk1 Qk1 t1pþi L t1 E½Y t1p ðkÞ ¼ 0: Qk1 t1þi Ltp X i¼0 ðA:45Þ )# Qk1 t1pþi L 2 l2L E½Qk1 t1 ¼ 0: Therefore, Eq. (22) reduces to 2 2l 2 2 2 2 Var ðQkt Þ ¼ 1 þ þ 2L Var ðQk1 t1 Þ þ 2 lD rL þ z Var ðst ðkÞ st1 ðkÞÞ p p p 2z Cov ½Y t1 ðkÞ; ðst ðkÞ st1 ðkÞÞ þ 2z Cov ½Qk1 t1 ; ðst ðkÞ st1 ðkÞÞ þ p 2z Cov ½Y t1p ðkÞ; ðst ðkÞ st1 ðkÞÞ: p ðA:46Þ ðA:47Þ ðA:48Þ ðA:49Þ Instead of finding an exact expression for Var ðQkt Þ, we shall approximate it. Since we can express Yt(k), the LTD at stage k, in terms of the LTD at the first stage, Xt, as J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 Y t ðkÞ ¼ XX X 635 ðA:50Þ X t; all covariance terms and Var (st(k) st1(k)) can be approximated by using relationships in Eq. (A.50). As a result, we obtain the following iterative approximation of the bullwhip effect when information on customer demand is not shared: When k = 1, Eq. (A.49) becomes Eq. (16). For k P 2, Var ðQkt Þ 2 2lL 2 2 2 2zlk1 z2 lk1 p 2 ð4Þ 2 4 ð3Þ k1 L L ffi 1 þ þ 2 Var ðQt Þ þ 2 lD rL þ 2 l þ l 2 rX p p p3 X p p p rX X 2r2X k1 zl 2ðlL 1Þ ð3Þ ð3Þ ½lD þ lD r2D : þ L lD p1 prD ðA:51Þ Appendix B Let Dt ¼ lD þ qDt1 þ et ; ðB:1Þ lD and Var ðDt Þ ¼ then, EðDt Þ ¼ 1q fined as Xt ¼ L1 X r2D . 1q2 Let the lead time, L, be deterministic, with lead-time demand, Xt, de- ðB:2Þ Dtþi : i¼0 m q 2 From Eq. (15), let us consider Var (Xt1) first. Since Cov ðDt ; Dtþm Þ ¼ 1q 2 rD , Var ðX t1 Þ ¼ L1 X Var ðDt1þi Þ þ 2 X Cov ðDt1þi ; Dt1þj Þ i<j i¼0 L 2q qð1 qL1 Þ ðL 1Þ þ r2D ¼ Var ðX t1p Þ: ¼ 1 q2 ð1 qÞð1 q2 Þ 1q Now consider the covariance terms, Cov ðDt1 ; X t1 Þ ¼ Cov Dt1 ; L1 X ! Dt1þi i¼0 ¼ 1 qL r2 : ð1 q2 Þð1 qÞ D ðB:3Þ ðB:4Þ Similarly, Cov ðDt1 ; X t1p Þ ¼ qpLþ1 ð1 qL Þ 2 r : ð1 q2 Þð1 qÞ D ðB:5Þ Next consider Cov ðX t1 ; X t1p Þ ¼ L1 X L1 X i¼0 Cov ðDt1þi ; Dt1pþj Þ ¼ j¼0 qpLþ1 ð1 qL Þ ð1 q2 Þð1 2 qÞ 2 r2D : ðB:6Þ From Eqs. (A.17) and (A.18), Cov ðX t1 ; ðst ðX Þ st1 ðX ÞÞÞ ¼ Cov ðX t1p ; ðst ðX Þ st1 ðX ÞÞÞ: By putting (B.3)–(B.7) into Eq. (15), we obtain Eq. (19). ðB:7Þ 636 J.G. Kim et al. / European Journal of Operational Research 173 (2006) 617–636 References Bagchi, U., Hayya, J.C., Ord, J.K., 1984. Modeling demand during lead time. Decision Sciences 15 (2), 157–176. Bagchi, U., Hayya, J.C., Chu, C.-H., 1986. The effect of lead-time variability: The case of independent demand. Journal of Operations Management 6 (2), 159–177. Cachon, G.P., Fisher, M., 2000. Supply chain inventory management and the value of shared information. Management Science 46 (8), 1032–1048. Chatfield, D., Kim, J., Harrison, T., Hayya, J., 2004. The bullwhip effect in supply chains—impact of stochastic lead times, information quality, and information sharing: A simulation study. Production and Operations Management 13 (4), 340–353. Chen, F., Drezner, Z., Ryan, J., Simchi-Levi, D., 2000. Quantifying the bullwhip effect in a simple supply chain: The impact of forecasting, lead times, and information. Management Science 46 (3), 436–443. Dejonckheere, J., Disney, S.M., Lambrecht, M.R., Towill, D.R., 2003. Measuring and avoiding the bullwhip effect: A control theoretic approach. European Journal of Operational Research 147 (3), 567–590. Dejonckheere, J., Disney, S.M., Lambrecht, M.R., Towill, D.R., 2004. The impact of information enrichment on the bullwhip effect in supply chains: A control theoretic approach. European Journal of Operational Research 153 (3), 727–750. Güllü, A.R., 1997. A two-echelon allocation model and the value of information under correlated forecasts and demands. European Journal of Operational Research 99 (2), 386–400. Kendall, M., Stuart, A., 1977. The Advanced Theory of Statistics, vol. 1. MacMillan Publishing Co., Inc., London. Lee, H., So, K.C., Tang, C.S., 2000. The value of information sharing in a two level supply chain. Management Science 46 (5), 628– 643. Mason-Jones, R., Towill, D.R., 1997. Information enrichment: Designing the supply chain for competitive advantage. Supply Chain Management 2 (4), 137–149. Merkuryev, Y., Pethova, J., Buikis, M., 2003. Simulation-based analysis of the bullwhip effect in supply chains. Paper presented at the EURO/INFORMS meeting, Istanbul Turkey, July 6–10, 2003. Mitra, S., Chatterjee, A.K., 2004. Leveraging information in multi-echelon inventory systems. European Journal of Operational Research 152 (1), 263–280. Mood, A.M., Graybill, F.A., Boes, D.C., 1974. Introduction to the Theory of Statistics, third ed. McGraw-Hill, New York. Ord, J.K., 1972. Families of Frequency Distributions. Griffin, London.
© Copyright 2026 Paperzz