Decision Sciences Volume 37 Number 2 May 2006 C 2006, The Author C 2006, Decision Sciences Institute Journal compilation TECHNICAL NOTE Recursive Behavior of Safety Stock Reduction: The Effect of Lead-Time Uncertainty Ping Wang Fisher College of Business, The Ohio State University, Columbus, OH 43210, e-mail: wang.607@osu.edu James A. Hill† Fisher College of Business, The Ohio State University, Columbus, OH 43210, e-mail: hill.249@osu.edu ABSTRACT Motivated by a recent paper on the effect of lead-time variability reduction on safety stocks, we provide evidence of the recursive nature of safety stock changes. When lead times follow a gamma distribution we demonstrate that, for cycle service levels between .60 and .70, the reduction of lead-time variability will first increase safety stock and then either recursively decrease safety stock or make it remain constant. We also numerically show the existence of the recursive effect. A two-by-two matrix is introduced to assist managers in making decisions regarding safety stock policy. Subject Areas: Inventory, Lead-Time Variability, Probability Models, and Safety Stock. INTRODUCTION In a recent paper, Chopra, Reinhardt, and Dada (2004; hereafter referred to as CRD) argued that many firms in practice operate at cycle service levels (CSLs) in the 50–70% range. CRD proved that, for CSLs above 50% but below a threshold, reducing the lead-time variability increases the reorder point and safety stock. They concluded, “In this range of CSLs, a manager who wants to decrease inventories should focus on decreasing lead times rather than lead-time variability.” This result contradicts the prescriptions drawn from using the normal approximation of demand during lead time. However, their results did not articulate where the threshold point is in the 50–70% range. We demonstrate in this article, numerically, that CRD’s results most likely hold at CSLs in the 50–60% range. CRD demonstrate that at a CSL of 60%, reducing the standard deviation of lead time by 20% (from 5 to 4) increases the safety stock. What they did not show is what happens at different levels of lead-time variability and, more important, at CSLs between 60% and 70%. † Corresponding author. 285 Recursive Behavior of Safety Stock Reduction 286 Table 1: Safety stocks for gamma lead time at different lead-time variability and service levels (mean lead time = 10). CSL .95 .75 .70 .65 .60 .55 .51 ρ t = .1 ρ t = .2 ρ t = .3 ρ t = .4 ρ t = .5 ρ t = .6 ρ t = .7 ρ t = .8 ρ t = .9 98 45 36 29 22 15 10 119 50 40 31 22 14 8 148 57 44 33 22 12 4 181 65 49 35 22 9 0 218 72 52 35 20 5 –6 256 78 55 35 17 0 −12 296 83 56 33 13 –6 –20 337 86 56 30 7 –14 –29 379 89 55 26 1 –22 –39 After replicating their calculation of safety stocks at different lead-time variability and CSLs (see Table 1), we find that at a CSL of 60%, any further reduction of the standard deviation of lead time (from 4 to 3 and so on) does not increase safety stock and that at a CSL of 65% there exists a recursive effect of the reduction of lead-time variability on safety stock, where safety stock first increases and then decreases recursively. We also find that this threshold value of CSL is dynamic in that it changes with a reduction of lead-time variability. We offer the following proposition when lead time follows a gamma distribution: For a fixed CSL above 60% but below a threshold CSL, the reduction of leadtime variability will first increase the safety stock, until the lead-time variability goes below some threshold lead-time variability; any further reduction will either have no effect on the reduction of safety stock, or recursively decreases the safety stock. In the remainder of this article we refer to this phenomenon as the recursive effect. The proof of the existence of such an effect is given in the Appendix. NUMERICAL ANALYSIS AND BOUNDARY CHARACTERISTICS OF THE RECURSIVE ZONE To further explore the recursive effect, we replicated CRD’s numerical experiment (CRD, 2004, p. 8). Parameters we used for the replication are summarized in Table 2. Our computational results reveal that the recursive effect is more pronounced within the range of CSL from .60 to .70, defined as the recursive zone. See Hill and Wang (2005) for detailed computational results. From our results we observe that CSL = .70 behaves as an upper bound of the recursive zone and CSL = .60 behaves as its lower bound. We call these bounds as CSL bounds of the recursive zone. The lower bound of the recursive zone partitions the CSL range of 50– 70% into two subranges, where in the 50–60% range, hereafter referred to as the counter zone, the effect observed by CRD occurs, and in the 60–70% range the recursive effect occurs. At CSL above 70%, hereafter referred to as the conventional zone, the effect of reductions of lead-time variability on safety stocks follows the conventional prescription of normal approximation. Wang and Hill 287 Table 2: Parameters. Variables Default Value Viable Values Description μd σd μt ρt 20 10 10 .1, .2, . . ..8, .9 10, 30 5, 15 5, 15 Mean demand SD of demand Mean lead time Coefficient of variance of lead time Cycle service level CSL .95, .90, .85, .80, .75, .70, .65, .60, .55, .51 Table 3: Boundary behaviors of recursive zone as reducing the lead-time variability. High Lead-time Variability (ρt ≥ .60) Low Lead-time Variability ( ρt < .60) Upper bound Has no significant impact on Decreases reorder point significantly (CSL close to .70) reorder point (stable zone) (conventional zone) Lower bound Increases reorder point Has no significant impact on (CSL close to .60) significantly (counter zone) reorder point (stable zone) The boundary characteristics of the recursive zone are summarized in Table 3, where the columns represent the lead-time variability. High lead-time variability ranges from .60 to .90 (inclusive), while low lead-time variability represents the region below .60. The entire table constitutes the recursive zone. To simplify, we refer to the two extreme quadrants, “Low ρt /upper bound” and “High ρt /lower bound,” as conventional zone and counter zone, respectively. The other two zones, that is, “High ρt /upper bound” and “Low ρt /lower bound,” are referred to as stable zones. When operating in the conventional zone, the reorderpoint behavior follows the conventional effect. On the other hand, when operating in the counter zone, the reorder point behaves as described by CRD. The stable zones mean that in those areas the reduction of lead-time variability has either no impact or only a slight impact on reducing the reorder point. In-depth observations reveal that the recursive effect is more pronounced at CSL = .65 than at other levels. Figure 1 shows how CSL = .60 partitions the CSL range of 50–70% and how the recursive zone functions as a transition area between the conventional zone and the counter zone. Finding the Recursive Zone We now analytically show the existence of the recursive zone. Figure 2 is a schematic illustration of three CDFs with lead-time variability at different levels: Recursive Behavior of Safety Stock Reduction 288 Figure 1: Recursive effects when lead time is gamma. Figure 2: Finding the recursive zone. CSL 1 α α2 γ α1 β cdf for y cdf for y+2 cdf for y+1 ROP Wang and Hill 289 (y + 2), (y + 1), and y. ᾱ2 is the crossover between cdf for (y + 2) and cdf for (y + 1), while ᾱ1 is the crossover between cdf for (y + 1) and cdf for y. The change from ᾱ2 to ᾱ1 shows the dynamic feature of the threshold CSL, which is neglected in CRD. For a CSL at α, which is higher than both ᾱ2 and ᾱ1 , any reduction of lead-time variability from (y + 2) to (y + 1 ) or from (y +1) to y will decrease the reorder point, as indicated by the conventional effect. For a CSL at β, which is lower than both ᾱ2 and ᾱ1 , any reduction of lead-time variability from (y + 2) to (y + 1) or from (y + 1) to y will increase the reorder point, as indicated by the counter effect. For a CSL at γ , which is in the range between ᾱ1 and ᾱ2 , a reduction of lead-time variability from (y + 2) to (y + 1) will first increase the reorder point, but a further reduction from (y + 1) to y will recursively decrease the reorder point, as indicated by the recursive effect. Thus the CSL range between ᾱ1 and ᾱ2 is the recursive zone. CONCLUSION We demonstrated in this article that the results of CRD did not fully explain the behavior of the reduction of lead-time variability on the direction of reorder-point changes. The reason for that is the neglect of the dynamic feature of the threshold CSL. We prove theoretically and demonstrate numerically that a recursive zone exists in the CSL range of .60–.70. Within the recursive zone a reduction of leadtime variability might first increase the reorder point and then recursively decrease the reorder point or make it remain constant. A more robust model of lead-time variability and demand variability remains an open and challenging problem. [Received: October 2005. Accepted: April 2006.] REFERENCES Chopra, S., Reinhardt, G., & Dada, M. (2004). The effect of lead-time uncertainty on safety stocks. Decision Sciences, 35(1), 1–24. Hill, J. A., & Wang, P. (2005). Demand uncertainty and lead-time uncertainty: Implications on the boundary behavior of the recursive zone. Working Paper. APPENDIX We are going to prove the existence of the recursive effect by following the logic of CRD’s proof of Theorem 3, Part 2. Theorem 1. Given a CSL α, .5 < α < 1, there exists y ∗ > 0 such that (i) R y ∗ (α) < R0 (α) and (ii) R y ∗ (α) < R y ∗ +1 (α), where y ∗ is a threshold lead-time variability. Proof. First, following CRD’s logic in their proof of Theorem 3, Part 2, we can have R y (α) < R0 (α), for a small y > 0. However, we find a condition, , which is weaker than k < (4γ +1 4γ 2 ) , the one identified by CRD. k = 2cY 2 < (4γ4 ++ 3γ 4γ 2 ) To find the weaker condition, let the LHS of (A1) be B(γ , β): Recursive Behavior of Safety Stock Reduction 290 4γβ + 2γ 2 β + 2β 2 B(γ , β) ≡ exp −k 1 − β2 − 2+γ +β ≥ 0. (1 + β)3/2 2+γ −β (1 − β)3/2 (A1) Now, to make sure that B(γ , β) is initially a nonincreasing function at β = 0, we take a partial derivative of B(γ , β) with respect to β and set it bigger than zero, that is, ∂ B(γ , β) = −k(4γ + 4γ 2 ) + 4 + 3γ > 0. ∂β β=0 . This will lead to k = 2cY 2 < (4γ4 ++ 3γ 4γ 2 ) Next, we need to illustrate that there exists a y ∗ , such that for y > y ∗ , such as y = y ∗ + 1, R y ∗ (α) < R y ∗ +1 (α). It is equivalent to showing that after increasing y beyond a certain value y ∗ , any further increase in y will make B(γ , β) negative. Without loss of generality, we choose β → 1− . After simple algebra we have 4γβ + 2γ 2 β + 2β 2 2+γ −β 2+γ +β lim exp −k − β → 1− 1 − β2 (1 − β)3/2 (1 + β)3/2 =0− 2+γ +β 2+γ +β =− < 0. 3/2 2 23/2 By continuity we know that there exists a lead-time variability y∗ , y ∗ ∈ (0, Y ), such that A changes from a positive value to a negative value (refer A to CRD, p.19). Following CRD’s logic we conclude that, when the lead time is uniformly , a small increase in lead-time variability but distributed, for k = 2cY 2 < (4γ4 ++ 3γ 4γ 2 ) below a threshold lead-time variability y∗ will make the CSL initially increase and the associated reorder point decrease, that is, R y ∗ (α) < R0 (α) holds; once the lead-time variability increases above y ∗ , the CSL will recursively decrease and the associated reorder point will increase, that is, R y ∗ (α) < R y ∗ +1 (α) holds. Ping Wang is a PhD candidate at The Ohio State University. He received his Master’s of engineering in logistics from MIT and his Master’s of engineering in computer science from the Chinese Academy of Science. His research interests include lead-time uncertainty in supply chains and supplier relationship management. James A. Hill is an assistant professor at The Ohio State University. He received his PhD in operations management from The Ohio State University. He conducts research on master production scheduling in process industries, lead-time uncertainty in supply chains, and the value of cross-docking in supply chains. His research has appeared in the International Journal of Production Research and Interfaces.