You are on page 1of 5

Operations Research Perspectives 3 (2016) 27–31

Contents lists available at ScienceDirect

Operations Research Perspectives
journal homepage: www.elsevier.com/locate/orp

Safety stock placement in supply chains with demand forecast updates
Youssef Boulaksil∗
Assistant Professor of Operations Management and Logistics, UAE University, College of Business and Economics, P.O. Box 15551, Al Ain, United Arab Emirates

a r t i c l e

i n f o

Keywords:
Safety stocks
Demand uncertainty
Supply chains
Simulation-based optimization

a b s t r a c t
Supply chains are exposed to many types of risks and it may not be obvious where to keep safety stocks
in the supply chain to hedge against those risks, while maintaining a high customer service level. In this
paper, we develop an approach to determine the safety stock levels in supply chain systems that face
demand uncertainty. We model customer demand following the Martingale Model of Forecast Evolution
(MMFE). An extensive body of literature discusses the safety stock placement problem in supply chains,
but most studies assume independent and identically distributed demand. Our approach is based on a
simulation study in which mathematical models are solved in a rolling horizon setting. It allows determining the safety stock levels at each stage of the supply chain. Based on a numerical study, we find that
a big portion of the safety stocks should be placed downstream in the supply chain to achieve a high
customer service level.
© 2016 The Authors. Published by Elsevier Ltd.
This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction
Many firms and supply chains are under the pressure to offer
a high customer service level while operating efficiently with low
inventory levels. At the same time, supply chains are exposed to
different types of risks, such as uncertain customer demand, uncertain supply, uncertain yields, uncertain lead times, and natural and
man-made disasters [17]. Several strategies have been developed
to hedge against these risks, such as safety time, safety stocks or a
combination of both [4].
The objective of this paper is to present an approach that allows determining safety stocks to hedge against demand uncertainty. Demand uncertainty is the risk factor that is supposed to
have the biggest impact on the performance of supply chains [17].
Setting safety stocks along a supply chain has been described by
Graves and Willems [7] as a strategic effort in supply chain planning that allows absorbing demand uncertainties and avoiding lost
sales and backorders.
The main contributions of this paper are twofold. First, we develop an approach that determines the safety stocks in supply
chains by assuming that the demand follows the Martingale Model
of Forecast Evolution process. Based on a case study and an extensive simulation study, Heath and Jackson [9] shows that MMFE
is a better approach to model the demand evolution. It better re-

Corresponding author.
E-mail address: y.boulaksil@uaeu.ac.ae, youssef.boulaksil@gmail.com

flects the demand pattern for many products in a real-life situation, compared to modeling it as independent and identically distributed demand. Our second contribution is that we provide managerial insights about where the safety stocks should be positioned
in the supply chain: at downstream stages close to the customer or
upstream in the supply chain where the inventory holding cost is
lower, but the response time longer.
In our approach, we assume that the supply chain is controlled
by a central authority, which has full visibility on the status of
the supply chain. Nowadays, many companies have implemented
Advanced Planning Systems that allow full visibility of the supply
chain and that assist in the coordination and decision making in
the supply chain [16]. These systems use mathematical programming models to decide on the optimal quantities to be produced,
given several parameters, materials and resources constraints, and
the target service level [3].
The remainder of this paper is organized as follow. In the next
section, we review the most relevant studies from the literature.
The model formulation and the solution approach are presented in
Section 3. In Section 4, we present the results of a numerical study,
and in Section 5, we draw a few conclusions and present the main
managerial insight from this study.
2. Literature review
An extensive amount of literature studies supply chains and inventory systems under uncertain demand. We refer the reader for
good reviews to Axsater [1], Federgruen [6], Van Houtum et al.

http://dx.doi.org/10.1016/j.orp.2016.07.001
2214-7160/© 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

t +2 dt +1. The demand and demand forecasts evolve from one time period to the next according − → to an additive evolution model. given several input parameters. Many more papers use the MMFE as a demand forecasting model for other purposes (see e. Graves and Willems [7] discuss the so-called guaranteed-service model for setting safety stocks in a supply chain under demand uncertainty. Evolving forecasts and demand patterns may be more realistic in a real-life setting. which is the objective of this paper. [3] develop simulation based optimization approach to determine the safety stocks for a multi-stage supply chain.t +1 . Decisions variables for future time periods represent only planned decisions. dt . where T is the planning horizon.t +T ).t +2 = dt . He obtains the system-wide order-up-to level and the expected system cost under the forecast evolution model. Then.t +2 . A large number of studies addresses the safety stock placement problem. the demand forecasts for the coming T time periods become available. Although the objective of this paper is similar to ours. under forecast evolution.t +T +1 = μ + εt +1.t +s denotes the demand forecast made in time period t for the demand in time period t + s. Jung et al. which may not be a realistic assumption in many business contexts [9]. .. as an Advanced Planning System that has full visibility and supports the decision making in the supply chain. The safety stock levels were determined based on a simulation approach in which the planning model was solved in a rolling horizon setting. Model formulation We consider a supply chain with several stages. . the forecasts get updated by: dt +1. the demand forecast is equal to the mean demand μ. Jung et al. that he compares with the order-up-to level and the expected system cost under a standard demand model.t +2 + εt +1. We assume that for all the further periods. the modeling approach and assumptions are fundamentally different. The supply chain is controlled by a centralized planning system.. the results demonstrate that his productiondistribution model yields significantly better results when using MMFE demand forecasts compared to moving average or exponential smoothing. and Inderfurth [10]. Boulaksil et al. such Fig. 1. which allows us to determine the required safety stock levels to achieve the target service level. [12] propose a simulation-based approach to determine the safety stocks in a chemical process supply chain. [20]). However. but customer demand can only be satisfied from the most downstream stage and we assume that all unsatisfied demand is backordered. The solution of the planning model for time period t = 1 are implemented. 3. .28 Y. which is the demand realization in time period t. Heath and Jackson [9] introduce MMFE as a modeling technique for evolving demand forecasts and compare it with traditional forecasting methods. Other relevant studies that used a simulation approach to determine safety stock levels are Jung et al. Using a series of simulations. At the end of period t. At the beginning of time period t. Sitompul et al. Yücer [21] builds on Heath and Jackson’s work to model the evolution of forecasts in a two-stage productiondistribution system by using stationary. dt . After the decisions are made. [12]. the planning model is solved. . The authors find that MMFE outperforms the traditional forecasting methods in terms of forecast accuracy and it results in lower total supply chain cost. We apply the approach of Heath and Jackson [9] who propose a general technique called the Martingale Model of Forecast Evolution (MMFE). Each time when the horizon is shifted. They develop a model for determining safety stock levels in a supply chain where each stage is controlled by a basestock policy and under the assumptions that an upper bound exists for the customer demand and infinite capacity constraints. Güllü [8] studies a twoechelon supply chain that consists of a central depot and multiple retailers. dt . which represent production and inventory decisions for the current period. and Boulaksil et al. [15] extend this stream of papers by considering capacity constraints. and the demand forecasts for each period of the planning horizon. Boulaksil / Operations Research Perspectives 3 (2016) 27–31 [19]. normally distributed demand with an autoregressive order-1 structure (AR-1). Each time period t. Kindly note that we assume that customer demand can only be satisfied from the most downstream stage.t +3 . − → − → Let dt denote the forecast vector dt = (dt . Hence. Given dt . we are able to derive the distribution of the inventory levels and backorders. The updates are due to updated information and circumstances. status information about inventory levels and backorders.t +T +1 . the demand forecasts are updated following the MMFE method. 1. The standard demand model results in higher order-up-to levels and higher system costs. Dt becomes available. They find that safety stocks should be increased by a constant correction factor which is dependent on the capacity limitation. Demand (forecast) generator In this section. the planning horizon is shifted by one period and a new planning problem arises that needs to be solved following the same approach.. dt +1. Jung et al. (2008) extend this work by including capacity constraints. none of the papers determined the safety stock levels by using the MMFE demand forecast evolution. [11]. [5].1. Other studies that also assumed uncapacitated supply chains are Simchi–Levi and Zhao [14] and Ettl et al. [3]. the actual demand gets revealed based on which the actual cost and customer service level are determined for the current time period. but with a strong focus on improving the forecasting method or on inventory planning. In this vector. we describe the model that generates the demand forecasts that are input to the planning model.t +3 = dt .g. A few papers have considered such demand patterns. A schematic overview of the safety stock placement approach. All these studies have in common that they assume that the customer demand is independent and identically distributed. but we will only discuss the most relevant ones. Our solution approach to determine the safety stock levels consists of a number of steps that are shown in Fig.t +3 + εt +1. the demand (forecast) generator only generates demand (forecasts) for the most downstream stage. By repeating this cycle very often. The planning problem is formulated by mathematical programming principles and assumes a planning horizon of T discrete time periods. Similar results have been obtained by Toktay and Wein [18]. 3.

t +k + εt .t +k + εt −1.t +1 . which are production quantities to become available after L time periods.t +1 . while the frozen horizon j contains Ft that remain unchanged. 3. . t ≥ 1} forms a stationary and independent sequence with zero mean and which follows the normal distribution. The safety stock amount that is added is: SS j = ⎧   j ⎨ Btj − 1 − SL j Dt ⎩ t 0 s.t +1 + εt +1. 3.t +1 . .Y. Consequently.2. based on which we can determine the actual inventory level. To find the maximum amount of backorders that satisfies the target service level. We keep the planning model as simple as possible. such as prices of products from the competitors. the planning j horizon is shifted by one period. the j planning model decides on At . Boulaksil / Operations Research Perspectives 3 (2016) 27–31 − → For any t ≥ 1.t +k = dt −2.t +T ) is the vector that represents the evolution of demand and forecasts from t − 1 to t. the actual backorder quantity. The main idea of this approach is that as we get closer to the demand period t. These forecasts are updated from one period to the next period based on new information received by the company. we adjust the numerator of the function SL j by adding an amount to it that is equal to the safety stock SS j . 29 i f SL j < SL∗j t otherwise The idea of the adjustment procedure is shown in Fig. The planning model determines the optimal material and resource quantities to release in the whole supply chain for the planning horizon T with the objective to minimize the total inventory holding and backorder cost in the supply chain. as our objective is to determine the safety stock levels. by updating the demand forecasts for that period successively (through obtaining new information in each period). . j unit inventory holding cost at stage j unit backorder cost at stage j bill-of-materials factor from stage j − 1 to stage j planned inventory level in time period t at stage j j Bˆt planned backorder quantity in time period t at stage j Pt j production quantity in time period t at stage j j frozen production quantity in time period t at stage j At j planned production quantity in time period t at stage j Rt Cj ˆ  j replenishment quantity in time period t at stage j capacity level at stage j planned total supply chain cost Lead time of stage j Lj ˆ = Min  J  j=1  α j T   Iˆtj t=1 + J  j=1 j j Iˆtj − Bˆtj = Iˆt−1 − Bˆt−1 + Rtj − dtj =P j t−L j Pt = (1 − γ )Ft j + γ Atj j Ptj ≤ C j ∀j ≥ 2 if t ≤ L otherwise 0 1 Iˆtj . i. εt = (εt .t +k In many real-life situations.t +k = dt −1. we reduce the standard deviation of the forecast errors. which means that the amount of backorders was too high. 2 shows the inventory development after having have added safety stocks. In that figure. The right-sided figure in Fig. demand of period t + 1 is found by adding an error term to the most recent forecasts dt . In that case. After the planning model is solved. At+L+1 j will become Ft+L after shifting the horizon. like other planning parameters.3. Rtj γ=  We assume a supply chain with J stages.4. the ratio t tj was obviously too Dt high. t + T ].t. j It actual inventory level in time period t at stage j j Bt j Dt actual backorder quantity in time period t at stage j actual demand quantity in time period t actual total supply chain cost actual service level offered at stage j  β j T  t=1  Bˆtj SL j  +  + j j Itj = It−1 − Bt−1 + Rtj − Dtj j j Btj = Bt−1 − It−1 − Rtj + Dtj = J   α j T   Itj t=1 j=1 + J   β j T   Btj t=1 j=1 After many simulation runs. fixed batch sizes. and capacity levels.. This approach is similar to the . We develop the planning model according to the mathematical programming principles to allow the model being implemented in Advanced Planning Systems. the actual total supply chain cost .. Bˆtj . Evaluation model i=1 j Iˆt dtj = δ j−1. We − → assume that { εt . it might be that it is below the tar Bj get service level SL∗j . where stage j = 1 is the most downstream stage. Safety stock calculation After having determined SL j . Dt+1 = dt . we obtain: dt . We refer the reader to Heath and Jackson [9] for a justification of these assumptions.t +k = . and expert judgments [2]. j · Ptj−1 . In the decision horizon. Once the planning problem j is solved and At are determined for [t + L + 1. Planning model αj βj δ j−1.t +k + εt . Atj ≥ 0 In the planning model.. Please note that by iterating the evolution equations sufficiently many times. such as the lead times. = μ + T −k  εt −i. When shifting the horizon. we notice that the target service level may not be achieved due to the large amount of backorders. 2. the planning horizon is divided into the frozen horizon and the decision horizon.e. The demand forecasts are input to the planning model. εt . demand forecasts for a certain product are obtained for a number of periods ahead (the planning horizon) by using statistical methods and by considering several factors. we will be able to determine the actual service level offered. the actual demand Dt gets revealed. The left-sided figure shows the inventory development without any safety stocks. which is defined as the fraction of demand satisfied immediately from stock SL j = 1 −  Bj t t Dtj . Ft  3. .

When increasing L1 .30 Y. The results of all experiments can be found in Appendix A. Fig. . and t + 1 becomes period t. In each time period t. we notice that in all runs. For the first run. more safety stocks are placed at the downstream stage than at the upstream stage.25} Lj β1 β2 These 5 steps are repeated N times to determine the distribution of It1 .t +2 .6 Ghz processor and 2 Gb RAM. as shown in Fig. L2 ) = (1. . the results of the first 10 periods were not taken into consideration. The simulation study was meant to test our approach and to provide some managerial insights about the safety stock placement in supply chains. the average safety stock placement is 78% at the downstream stage and 22% at the upstream stage. . .1.3). Dt is revealed by the demand generator.t +T ). 2.50} {1. 2. B21 ) are determined and stored in the database. We assume that an Advanced Planning and Scheduling system is used to assist the decision making in the supply chain. Hence. α1 = 1. We notice that for all scenarios where (L1 .2. Ft+L . σε {20. The status information of the supply chain (It−1 t−1 1 2 1 1 1 2 2 2 Bt−1 . . The impact of the lead time structure on the safety stock allocation.1. . .5.3} {2. 3. The warm-up period was considered 10 periods. Ft . . Ft+1 . . Bt−1 . Numerical study In this section. 4. we present the results of a numerical study that we conducted by simulating the approach that is presented in the previous section. The demand forecast generator generates dt = (dt . 4. The simulation procedure is as follows. we assume μ = 100. . Go to step 1. 1 . From the results. and the evaluation model is solved. I12 . B11 . We assume that the supply of raw materials is sufficient and we are interested in determining the safety stock levels of the semi-finished products and the finished products. we assume that − → dt = μ. the following steps are conducted: − → 1. 4 shows how the lead time structure affects the safety stock allocation in the supply chain. Stages 1 and 2 produce the finished goods and semi-finished goods respectively. Bt2 . safety stock adjustment procedure as presented by Kohler–Gudum and De Kok [13]. and T = L1 + L2 + 1 The other parameters were varying according to Table 1. I2 . and the decisions are implemented.10 0. Table 1 Values used in the simulation study. Bt1 and It2 . Ft . The simulation study was designed in a way that it reflects many real-life situations much as possible [3].t +1 . . Simulated two-stage supply chain. For the simulation study. The left-sided figure shows the inventory development over time. and. 5. and the number of replications 3. it gets updated based on the equations presented in Section 3. Fig.20 0} σ ε /μ {0. 3. The planning horizon is shifted by one period. dt . 3. Ft+1 . based on which (I11 .2} {1. it results in a shift of almost all safety stocks to downstream in the supply chain. . the right-sided shows how the safety stocks shift the inventory function such that the amount of backorders is decreased. 4. some safety .5. we consider a two-stage supply chain. Ft+T ) is extracted from the database. Fig. We find that the two parameters that determine the safety stock allocation in the supply chain are: lead time structure and the level of demand uncertainty. dt . This means that if the most downstream stage has a short lead time. For the numerical study. All simulation runs were conducted on a desktop computer with a Pentium 4. based on which SS1 and SS2 are determined that allow to achieve SL∗1 and SL∗2 .2. In next runs.10. The planning model is solved. In the simulation study. 3. α2 = 0. N = 300 runs. Boulaksil / Operations Research Perspectives 3 (2016) 27–31 Fig.

Thesis. When the level of demand uncertainty increases.99 0. Vol 4. Dullaert W.84 0 0 0. Rinnooy Kan AH. Yao DD.783 7237 4572 75.80 0.85 0. Materials coordination in stochastic multi-echelon systems.99:386–400. Rinnooy Kan AH. A supply network model with base-stock control and service requirements. A simulation based optimization approach to supply chain management under demand uncertainty. Inderfurth K.679 4750 7077 71. Modelling the evolution of demand forecasts in a production-distribution system. 5. [12] Jung JY. Wein LM. Van Landeghem H.Y.103:451–88.99 0.332 135. Eversdyk D. 2006. Zhao Y.95:1–23.146 10. Int J Prod Econ 1994.79 0.47(9):1268–81. Feigin GE.32:2570–81. 2005. Technische Universiteit Eindhoven.26 0.888 776 1468 1254 3805 12. Graves SC. Optimizing strategic safety stock placement in supply chains. Amsterdam. which may better reflect the customer demand evolvement for many products compared to when assuming it being independently and identically distributed. The Netherlands: North-Holland Publishing Company. [4] Buzacott JA.79 0 SS1 30 31 278 Stage 2 2 2 B¯ I¯ 27 29 35 157 964 1150 S L2 SS2 0. [19] Van Houtum GJ. Conclusions In this paper. A two-echelon allocation model and the value of information under correlated forecasts and demands.97 0. we presented an approach that optimizes the safety stock placement in a supply chain under demand uncertainty. BETA Working Paper 71. Safety stock versus safety time in MRP controlled production systems. 3rd edition.31 0. Safety stock placement problem in capacitated supply chains. Comput Chem Eng 2004.7(4):295–318.28:2087–106. [17] Tang CS. Zipkin PH. Oper Res 20 0 0.99 0. Boulaksil / Operations Research Perspectives 3 (2016) 27–31 Input parameters Fig.97 0. [7] Graves SC. Eur J Oper Res 1997. [14] Simchi-Levi D. logistics of production and inventory. Handbooks in operations research and management science. Safety stocks in multi-stage. Reklaitis GV. [8] Güllü R.39(2):168–71. These insights might be counter intuitive to many managers who believe that safety stocks should be placed upstream in the supply chain. Zipkin PH. Int J Prod Res 2007:1–19. Zijm WHJ.35:321–9.66 128 543 698 30 102 12 1088 881 134 1941 1720 2617 2622 75 1463 807 149 1167 5988 5768 2227 4823 0 1625 5 3 6 3 5 1 98 84 76 4 8 3 11 5 4 251 45 207 5 9 3 143 27 15 0. especially when the demand uncertainty is high.153 14. [20] Wang T. Integrated safety stock management for multi-stage supply chains under production capacity constraints.302 4933 18. Analysis of a forecasting-production-inventory system with stationary demand. Modeling the evolution of demand forecasts with application to safety stock analysis in production/distribution systems. which allows the determination of the safety stock levels such that target customer service levels are achieved.2(1):68–83.81 0 0 20 28 34  474 1268 16. [21] Yücer CT. Experts’ stated behavior. 5 shows how the level of demand uncertainty affects the safety stock allocation in the supply chain.76 0 0.12 0. OR Spectrum 2009. Van Halm ENG. Vol 4. Reklaitis GV. Ankara.99 0. Fig. Centralized planning models for multi-echelon inventory systems under uncertainty. Our approach is a simulation study in which the supply chain planning model is frequently solved in a rolling horizon setting. Shanthikumar JG. 1993.60 0. Appendix A. Berlin: Springer.99 0. [16] Stadlter H. 2002. Handbooks in operations research and management science. 5. [11] Jung JY.97 0.30 0. Kurtulus M. We find in our experiments that almost all safety stocks are placed downstream in the supply chain.873 483 180 63 684 357 425 649 249 1001 677 241 524 1839 3589 284 1349 2160 699 1950 621 3157 1531 6856 8782 10 145 73 23 15 17 1582 4041 3206 480 515 2156 20 13 34 2818 5516 5297 55 1077 1821 59 73 532 References [1] Axsater S. When demand uncertainty is low (σε = 20). [18] Toktay BL. Manufact Serv Oper Manage 2005. Eversdyk D. editors.59 0 0 0 0 0. Eur J Oper Res 1996. close to the final customer. Comput Chem Eng 2008. [10] Inderfurth K. Turkey: Middle East Technical University. Otherwise.99 0. We also conducted a numerical study.78 0 0. Kilger C. Manufact Serv Oper Manage 20 0 0.78 0. [9] Heath DC.98 0.99 0.112 35. By doing so. on average. Pekny JF. editors. [5] Ettl M.577 7315 104. [6] Federgruen A. Summary of the simulation results Input parameters σε 1 2 3 20 20 20 L1 L2 β1 β2 1 1 1 3 3 3 2 10 50 1 5 25 Stage 1 1 1 B¯ I¯ 37 41 291 294 231 114 S L1 0. The Netherlands: North-Holland Publishing Company. Atasu A. Blau G. about 10% of the safety stock is placed upstream in the supply chain.99 0.99 0. as storing the products at upstream stages is cheaper than at downstream stages. Lin GY. Safety Stock positioning in supply chains with stochastic lead times. stocks can be placed upstream in the supply chain. Supply chain management and advanced planning. this decreases to an average of about 5%.40(12):1678–89. IIE Trans 1994. We assume that all unsatisfied demand is backordered.25 0. De Kok AG. [2] Boulaksil Y. [15] Sitompul C. Int J Prod Econ 2006.72 0.26(3):17–30. Graves SC.97 0. Fransoo JC.241 60. [3] Boulaksil Y.07 0. logistics of production and inventory.514 11.146 (continued on next page) 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 31 σε L1 L2 β1 β2 Stage 1 1 1 B¯ I¯ S L1 SS1 Stage 2 2 2 B¯ I¯ S L2 SS2  20 20 20 20 20 20 100 100 100 100 100 100 100 100 100 200 200 200 200 200 200 200 200 200 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 2 2 2 1 1 1 3 3 3 2 2 2 1 1 1 3 3 3 2 2 2 1 1 1 2 10 50 2 10 50 2 10 50 2 10 50 2 10 50 2 10 50 2 10 50 2 10 50 1 5 25 1 5 25 1 5 25 1 5 25 1 5 25 1 5 25 1 5 25 1 5 25 186 557 713 34 108 15 1133 945 152 2006 1779 2682 2689 95 1505 870 193 1253 6129 5908 2307 4936 29 1903 0. Blau G. . Setting safety stocks in multi-stage inventory systems under rolling horizon mathematical programming models. the distribution of the actual inventory and backorder levels can be derived. Amsterdam. Manage Sci 1994. Perspectives in supply chain risk management.99 0.94 0.99 0 0 0 0 0 0 65 78 71 0 0 0 0 0 0 192 34 196 0 0 0 30 0 0 866 28. The impact of the level of demand uncertainty on the safety stock allocation. Chapter 3. Chapter 4.27 0 0 0 0 0. Manufact Serv Oper Manage 2012. Jackson PL.48(2):216–32. A multi-ordering newsvendor model with dynamic forecast evolution. Safety stocks are only shifted to upstream stages in the supply chain if the downstream stage has a relatively short lead time and customer demand is less uncertain. A safety stock adjustment procedure to enable target service levels in simulation of generic inventory systems.14(3):472–84.96 0. Aghezzaf E-H.286 119. Manage Sci 2001. all safety stocks should be placed at the most downstream stage in the supply chain. [13] Kohler-Gudum CK. Franses PH. Pekny JF. We modeled the customer demand following the Martingale Model of Forecast Evolution (MMFE) method. Willems SP.99 0. Interfaces 2009.31(1):121–40. . Continuous review policies for multi-level inventory systems with stochastic demand demand. divergent inventory systems: a survey. 1993.