MANAGEMENT SCIENCE ‘Vol. 54, No. 8, August 2008, pp. 1453-1466 19x 0025-1908 | issn 1526-550] |08 | 54081453 ints, 101 10.1287 /mnse:1080.0871 (©2008 INFORMS, Inventory Models for Substitutable Products: Optimal Policies and Heuristics Mahesh Nagarajan Operations and Logistics, Sauder School of Business, University of British Columbia, Vancouver, Bagh Columbia V6T 122, Canada, mahesh nagarjandenuderuibccn S. Rajagopalan Information and Operations Management, Marshall choo! of Busines, University of Southern California, Los Angeles, California 0089, srajagop@marsalluscect his paper we examine the nature of optimal inventory policies in a system where a reser manages substitutable products. We first consider a system with two products 1 and 2 whose total demand is D and. individual demands are negatively correlated. A fixed proportion of the unsatisfied customers for an item will purchase the other item if itis available in inventory. For the single-period case, we show that the optimal inventory levels of the two items can be computed easily and follow what we refer to as “partially decoupled” policies, ie., base stock policies that are not state dependent, in certain critical regions of interest both when D is known and random. Furthermore, we show that such a partially decoupled base-stock policy is optimal even in a multiperiod version of the problem for known D for a wide range of parameter values and in an N-product single-period model under some restrictive conditions. Using a numerical study, we show that heuristics based on the decoupled inventory policy perform well in conditions more general than the ones assumed to obtain the analytical results. The analytical and numerical results suggest that the approach presented here is most valuable in retail settings for product categories where the level of substitution between items in a category is not high, demand variation at the aggregate level is not high, and service levels or newsvendor ratios are high. Key words: inventory / production; stochastic multiperiod inventory models; stockout-based substitution History: Accepted by William S. Lovejoy, operations and supply chain management; received June 12, 2006. This paper was with the authors 54 months for3 revisions. Published online in Articles in Advance July 1, 2008. 1. Introduction ‘This paper explores the impact of consumerdriven substitution due to stockouts on the inventory deci- sions of a product category at a retailer. Consider a retailer who manages the inventory of two items that are partial substitutes, for example Duracell and Ener- gizer size C batteries. Because the demand for these items is random, the retailer may stock out of one or both of them and some unsatisfied customers for an item will purchase the other one if it is available in stock. The retailer has to consider these substi- tution effects in addition to the traditional trade-off between overstocking and understocking in deter- mining stocking levels. Several papers in the inven- tory literature have considered substitution effects, as will be clear later in the literature review. We consider a stylized model that yields new results and insights, which are used to develop a simple but effective heuristic to make inventory decisions for substitutable items within a product category. This approach can be useful to retailers of staple products such as batteries, grocery items, and office products. 1453, Consider a scenario with two substitutable prod- ucts, say 1 and 2, whose total demand in a period is D (D may be random) where D is, say, traffic at a store or demand for a particular product category. The individual demand proportions of the two products are p and (1—p) where p is random and the aver- age value of p is the market share of product 1 at the retailer. The retailer has to decide the stocking level for the two items. A fixed proportion y of unsatis- fied customers for a product will purchase the other one if it is available in stock. We analyze this model and its variants in both single-period and multiperiod scenarios. The main contributions of the paper are as follows. First, we show that, for reasonable parame- ter values, specifically for y values that are not large, the optimal base stock level of product 1 is indepen- dent of the inventory of 2, when the latter is above its base stock level—we call these “partially decou- pled” policies. Thus, even if product 2's inventory is very high, it is optimal to raise the inventory of prod- uct 1 to its base stock level. Furthermore, the opti- ‘mal inventory levels can be computed easily using closed-form expressions for any parameter value if D is known. Second, we identify conditions when such a[Nagarajan and Rajagopal 1454 result holds even if D is random and for an N-product model, Third, we show that the “partially decoupled policies” are optimal in the multiperiod two-product model under reasonable conditions on y. Finally, we show numerically that a heuristic based on the ana- lytical results performs well under conditions more _general than the ones assumed to obtain the analytical results and when there are more than two products. In particular, we use data from numerous academic and industry studies on service levels, substitution behavior, and demand characteristics to evaluate the performance of the heuristic and to point out when ur approach works well and when it is necessary to consider such substitution effects, 11. Literature Review ‘The inventory literature with substitution effects and a single decision maker can be broadly classi- fied into two categories, depending on whether the substitution is driven by the consumer or by the decision maker. The decision-maker-driven substitu- tion literature is also referred to as the transship- ment literature because substitution in such a setting may involve transshipment or transfer of products between different locations. Examples of this litera- ture include Robinson (1990), who examines a mul- tiproduct, multiperiod transshipment problem and characterizes the structure of the optimal policy, and Bassok et al. (1999), who consider a model with full downward substitution, show that the optimal pol- icy is a base stock policy and provide an algorithm to compute optimal inventory levels. Recent examples of papers in this literature are Rudi et al. 2001), Van Mieghem and Rudi (2002), Axsater (2003), and Deniz et al. (2008). ‘There is an important difference between the above setting and consumer-driven substitution in retail set~ tings, which is the focus of this paper, because when the decision maker makes substitution decisions, the level of substitution + referred to eatlier is nat exoge- nous (Mahajan and van Ryzin 2001). Also, unlike in consumer-driven substitution settings, itis typically ‘optimal to substitute between two products (or loca- tions) fully or not at all, assuming that there is suffi- cient inventory. However, the models formulated for the two contexts are similar, and so insights devel- ‘oped in one context have some relevance for the other, and we point out these linkages between our model and the transshipment model later in the paper. In any case, neither the transshipment literature nor the consumer-driven substitution literature has identified the type of partially decoupled policies discussed in this paper. Models with substitution effects tend to be inher- cently hard. As Mahajan and van Ryzin (1998) point out, dynamic substitution models, where the substi- tution behavior depends on the inventory status at Inventory Models or Subtitle Pract? Optima! Poles and Heuristics Management ene S48), pp. 145-168, 2008 INFORMS the time a customer makes his choice (as in our case), are more realistic but are typically less tractable, and the profit functions are in general quite complex and not easily amenable to analysis. McGillivray and Silver (1978) develop heuristic policies for the two- product problem when there is partial substitution. Parlar and Goyal (1984) also study the partial sub- stitution problem and show that the expected profit function is concave. Parlar (1985) considers a perish- able inventory model with substitution between new and old items when there is a stockout. Pasternack and Drezner (1991) show some interesting character- istics of optimal inventory levels in a two-product problem with full substitution. Gerchak and Mossman (1992) show that, contrary to intuition, substitution effects may result in an increase in optimal inventory levels under certain cost and demand conditions. The above papers consider single-period models where demands of the products are not correlated, unlike in our work. There are a few recent papers that study inven- tory policies with substitution effects when demands are correlated. Rajaram and Tang (2001) explore the impact of partial substitution and corcelation in demand in a single-period model, point out that char- acterizing exact optimal policies analytically is com- plex, and develop heuristics to compute inventory levels when demand is a bivariate normal distribu- tion. More recently, Yang and Schrage (2002) show that increased substitution may result in higher inven tory levels under certain conditions and discuss other interesting anomalies related to risk pooling. Ernst and Kouvelis (1999) study a problem of packaged goods with two substitutable products that may be sold singly or as a bundle and perform a numerical analysis on the effects of correlation on the optimal policy. Netessine and Rudi (2003) study an N-product ppartial substitution problem with random, correlated demands. They show that in general this problem need not be concave and perhaps not even unimodal, They compare models with one centralized decision maker versus N competing decision makers and show that optimal inventory levels are higher in the com- petitive scenario. The above papers are restricted to single-period scenarios, unlike our work, and illus- trate the difficulty in solving. substitution problems with or without demand correlation. There is another recent stream of literature on assortment planning, representative early works being Smith and Agrawal (2000) and Mahajan and van Ryzin (2001), which study inventory. systems ‘where consumer choice and substitution effects are modeled in great detail and substitution in out-of stock situations is probabilistic. The focus ofthis work. is on determining both optimal product assortmentsManagement Secce 548), p. 5-4, £2008 INFORMS jan and Rajagopalan: Inventory Mls for Sustutable Products: Optinal Poles and Herises 1455 and inventory levels ofthe items as a function of con- ‘sumer purchase behavior. In recent work, KSk and Fisher (2004) consider an assortment planning prob- Jem taking into account substitution effects and pro- vide structural results and an iterative heuristic that performs very well. They also discuss the implemen- tation of the model at a supermarket chain. Gaur and ‘Honhon (2005) study a single-period assortment plan- rning and inventory management problem that uses a locational choice model of consumer choice. They obtain bounds on profit when customers dynami- cally substitute and identify conditions under which static substitution serves a3 a good approximation. Kok et al. (2006) provide a comprehensive review of the literature on assortment planning, ‘The above papers present numerous insights— some of them consider models that are sicher than ‘ours in terms of the modeling of substitution and ‘demand, and some consider assortment decisions that we do not consider—but our work yields insights not found in these works. Specifically, the prior works do not show the optimality of “decoupled” inventory policies in single-period or multiperiod models in any type of scenario. Furthermore, we show numerically that the decoupled policies perform well even when some of the underlying assumptions are relaxed. The organization of this paper is as follows. In $2, we present the singlepperiod, two-product model, characterize the optimal inventory levels providing closed-form expressions, and extend the result to a special N-product model. In §3, we present an anal- ysis of the finite- and infinite-horizon two-product model and characterize the optimal policy. In §4, we present the results of a computational study, and finally, in §5, we summarize and present poten tial extensions of this research. Proofs of all the results are in the online appendix (provided in the e-companion).! 2. The Single-Period Model We begin with a single-period model where a retailer stocks tro products, say 1 and 2, which are par- tial substitutes. For ease of exposition and to focus attention on the demand substitution effects, we first study a simple model where total demand for the two products, D, is known but demand for each prod- uct is uncertain. Demand for products 1 and 2 are =p xD and d, = (1~p) «D, respectively, where is a random variable with a continuous distribu- tion function F that has a finite support [a, b], 0 Oand 5> c. Also, we assume zero starting inventories in our initial exposi- tion but relax it thereafter. The retailer decides how much of each product to stock to maximize expected profits. Let the order quantities be aD and BD, where 0 1, stockout-generated substi- tution demand can be fully satisfied from available inventory, which simplifies the formulation. We then have the following derivation for II, (a, B): ifa+p21, T(a, B) = ifa+p<1 Ma 1(a,A)=f {sp+8-+71-p—B)] = Map) 70 =p-B)] —a(1= (1p B)} aE) + ["s-Ma-p+8—a-pp} arte +f'ttera =P) +P) HB - (=p) ~y(p)] —a(1— (pa) Fp) — e(a +B) The three integral terms correspond respectively to the following three scenarios: (1) Inventory exceeds demand for product 1, demand exceeds inventory for product 2, and substitution demand of 2 is fully satisfied from remaining inventory of 1; (2) inven- tory exceeds demand for both products; (3) inven- tory exceeds demand for product 2, demand exceeds inventory for product 1, and substitution demand of 1 is fully satisfied from remaining inventory of 2. Note that the expression for Il, (a, 8) can be separated into sets of terms involving only « or only B. To facilitate the derivation of Tl,(a,B), we de- fine @ =(a— y(1 ~ f))/1— y and B=(1-B-ya)/ (17). Intuitively, @ is the p value at which excess inventory of product 1 is exactly equal to the sub- stitution demand from product 2. Similarly, A is the value of p at which excess inventory of product 2 is exactly equal to substitution demand from product 1. Note that @ 1~8. Furthermore, @ = a and 6 =«=1~8 when a +8 =1. We can then write Ti,(a,) as na, 6)= [(sp+8-+70-p-B)] = Hla —p)- 90 -p-B)) —a(1 yr) -p~B)} dF(P) Mangement Selence 548), pp. 1459-1466, 02008 INFORMS A + [/Islo+B)— a(t (a+ A))laF(p) + floor re- ai) —h(B-(1—p)-y(p—a)) ~ #(1—9)(P— «)} dF(p) ~ cle B) ‘The first (third) integral term represents the scenario where inventory exceeds demand for product 1 (2) but not product 2 (1) and all the substitution demand of product 2 (1) is satisfied. The second integral term collects all of the situations when there is no excess inventory of either product and substitution demand is not filly satisfied. Note that, unlike TI,(a, 8), the expression for I1,(«, B) carmot be separated into terms involving only a or only B. ‘The function M(a, B) was shown to be concave in (a). Itis easy to show that I1,(a, 8) is strictly con- cave in (a, B). Thus, the global maximum of TI(a, B) is uniquely attained. The optimal values of (a, 8) and (@,B), denoted respectively by (a*, 8°) and (@, B*), can be obtained by solving the corresponding first- order conditions of TI,(a, 8). So, we have: ‘Taeorem 2.2. Let y*=max{1—2(h+0)/(s-+7 +h), 1. If ¥ <7", the optimal inventory level is determined using a) _Stm—c—ye4a +h) reer er, Te Ate C= nG+a+hy’ * Sa hte FU-F)=Gayeeeth) a If y>y*, the optimal inventory level is determined using a au hte MO = T55" Tne TAH) strc ~ ane’ i) ey hte MP) =T45* Caner _ustath+hte ” (4 yist +h) where & and B* are functions of both a and . When y < 7", we have a “partially decoupled” inventory policy wherein the optimal base stock level of a product is independent of the inventory level of the other product. Thus, even if product 2's inven- tory (after ordering) is very high, it is optimal to raise the inventory of product 1 to its base stock level evenNagarajan and Rajagopalan: Imentory Modal for Substitute Products: Optimal Policies end Heiss Management Since 8), pp 145-1464, ©2008 INFORMS. 1457 though the products are substitutable. When y > y* this is not true because there is so much substitution that high inventory levels of one product significantly diminish the attractiveness of holding an incremental ‘unit of inventory of the other product. Recall that the expression for TI,(a, 8) can be separated into terms involving only a or only fin the case @-+B > 1, cor responding to the case y = 7° (as is clear from the proof of Theorem 2.2), and so the optimal decisions for the two products are decoupled. We will exploit this idea in later subsections to show that the results can be generalized to the case where total demand D is random and the product costs are different and for an N-product model. Observe that the optimal inventory expressions in (1) anc (2) are similar to the newsvendor formula, except for the term (1 ~ ). This implies that the higher the substitution fraction, the lower the opti mal inventory level. This effec i also true in the case where > y*, even though this is not transparent in ) and (4). This is consistent with prior results a3 well as the intuitive reasoning that, as items become more substitutable, the retailer needs to carry lower inventories. Even the early papers on both consumer- driven and retailer-driven substitution (McGillivray and Silver 1978, Robinson 1990) have pointed out that increased levels of substitution result in lower inventories. However, prior literature has not identi- fied an optimal policy for substitutable items wherein cone could decouple the inventory decisions without ‘ignoring substitution. Finally, the closed-form nature of the optimal inventory expressions (1)-(4) is attrac- tive, because it suggests a simple approach to adjust for substitution. 1m reality, consumers who have to frequently sub- stitute due to stockouts may stop visiting the retailer, and so the retailer may incur a cost due to such substi- tution. We can easily incorporate such a substitution cost, say b per unit substituted (or by more generally to represent cost of substitution from i to j). In this case, we add ~by(1 ~ p — f) within the frst integral term and —by(p—a) within the third integral term in the equations for 1,(a, ). The optimality equations will change as follows: the denominator in (1) will include a term yb, and in (3) (1+) in the denomi- nator will be replaced by (1+ ~by/(s-+7+i)). The impact is as one would expeet—optimal inventories increase with substitution cost b. Our formulation can also be seen as equivalent to a two-location transshipment problem wherein a frac- tion 7 of the customers at a location that is stocked cout will be willing to wait for the product (ie, a partial backorder) and the manufacturer or distribu- tor will transship the product from the other loca- tion to meet this backlogged demand. Note that the transshipment cost incurred by the manufacturer is ‘equivalent to the switching cost b above. Hence, the results we obtain here are valuable in solving special instances of the transshipment problem with partial backordering. The condition y < 7" is not restrictive, and the higher the newsvendor ratio (NR), the higher the value of 7°. If NR=(s+7—-o/(s + +h), then / = max(2«NR—1, 0) For instance, suppose $= 100, 20, 3 =5, and h = —8 (implying a significant sal- vage value). In this case, 7" =0.75, where the NR for each product without substitution is 0.87. So, we find that for a wide range of 7 values between 0 and 0.75 the decoupled policy is optimal. We discuss the range of 7 values typically observed in practice in $4 y 24. Asymmetric Costs and Random D ‘We now consider the case where the costs, price, and substitution parameter for the two producis are dif- ferent and total demand is random. We begin by con- sidering a general case where the individual demands for the two products d, are random and may be corre- lated. We then provide some specific insights for the case where demand for the two products is given by = pD and d, = (1~p)D, where D is random and p and D are not correlated. The cumulative dist tion and probability density functions ford, are given by F() and f(), respectively, and we assume that they have a compact support given by [L;, Hi]. So, the ‘maximum total demand for the two produets is given by H=(H, +H,). The cost parameters are now 5, 6, 1, h,>0 (I=1,2). A fixed proportion y, (0< 7, <1) of the customers who come in for item { and do not find it in stock will switeh and buy the other product if itis available, and a proportion (1 — 7.) will walk away. Multiproduct substitution inventory problems with asymmetric costs are difficult to analyze because the objective function may not be concave and sometimes not even unimodal (see for instance Netessine and Rui 2003). In addition to the assumptions made ear- lier, one needs to make certain assumptions on the cost parameters to guarantee concavity (see Parlar and Goyal 1984, Eenst and Kouvelis 1995). Specfi- cally, we assume that (5+ 7, +h) > 165, + hy + 7), i,j =1,2 (¢ j). This condition ensures that it is worthwhile to stock product i because, if it were not true, the retailer can earn higher reveniie by not stock- ing item i and having all potential consumers for item i substitute to item j. For similar reasons, we also require that (5, +) —<,) > (5) +h, +m), 1,j=1,2 (iF). Let Q, be the order quantity for product i (=1,2): because D is random, we do not use the variables a and B representing fractions of total demand D in this sub-section. Let the remaining inventory or unsold units of product i, after substitution demand from[Nagarajan and Rajagopalan: hnvertory Model for Substtutable Products: Optinal Policies aud Hewristios 1458 Managemen Sclence54(), pp 1453-1466, ©2008 INFORMS j (#i) is satisfied, be denoted by U, = [(Q,- 4))* — ‘y(d; — Q,)"F* where (x)* =max(x, 0). The total profit function is TH(Qh, Qs) X {(Qi-E[Uj)) QW ELUj] ~m1Q—yE4,—O)"] FEL((d-Q)* -(Q)-4)"], @) where E[.] denotes expectation. The first term is the revenue, the second is the purchase cost, the third is the inventory cost for unsold units, and the final term is the penalty cost due to both initial lost sales (of product i) and lost sales of substitution demand (of product j). We then have: Lema 2.1. If (Q, + Q:) = H, then M(Q,,Q2) = 11, (Qi) +11, (Q,), where TQ) = (5, — 6) Q; (5, + MELO; - 4)" (a Wl5, +h) + MEL — QI (6) Note that the profit function (6) is separable in the two products and the profit for each product is sim- ilar to a standard newsvendor profit function, with the overstocking cost equal to (¢;+h,) and the under stocking cost equal to [(s,—¢) +) = (5 +) + ™)] We then have the following result, which we state without proof, because it is similar to the proof of ‘Theorem 2.2. TuoreM 23. Let yf = max{(s,-+7)—(t,+2c)/ (+h +m) Ui f=1,2 147 UE HS yf, then G+m-o)- HS +47) GF mh) H+ hi +m)" Lj=LZi#}, FQ) Thus, when y, < yf, the “partially decoupled” inventory policy is again optimal. The expression for £(Q°) in (7) confirms the intuition that if product 2 is the higher-priced product (with, say, higher 5, ca, and h,), then the optimal base stock level of product 1 will be lower, and in fact if s: and h, are high enough, the optimal base stock level can be below the mean demand and even close to zero (also, as 5 increases relative to s,, yj decreases). Thus, the risk-pooling effect due to substitution depresses the inventories of both products as in the symmetric case. However, this impact is more significant on the lower-priced prod- uuct because the retailer finds it more profitable to hold. lower inventory of the lower-priced item and “allow” original consumers of this item to substitute to the higher-priced one. Another implication of asymmetric costs is that yj is lower for the lower-priced product relative to the 7" value in the symmetric case. Using the parameters in the symmetric case for product 1 (= 100, etc.) and one-half of the corresponding val- ues for product 2 so that, the newsvendor ratio is 0.87 for both products, we find that yj =0.38, yj = 1. Thus, the maximum substitution fraction allowed from the lower-priced product 2 is only 0.38 for the decoupled policy to be optimal. The expression in (7) also clari- fies the conditions imposed earlier to ensure concav- ity; note that, if the conditions are not satisfied, the numerator or denominator in the expression in (7) can. become negative. The condition (Q} + Q;) 2 H, which drives our results, can be quite restrictive in the general case. For instance, consider a case where demand for the two products is iid. with bounded support, say uniform [0, 1]. Then, the requirement (Q; + Q;) = H implies that we need (Qj + Q3) > 2. Clearly, this cannot hap- pen unless the newsvendor ratio is 1 when y, =0. When 7; is greater than zero, note that risk pooling kicks in, and thus this makes the requirement impos- sible for any value of the substitution parameter. This, illustrates the limitation of the value of this result in the general case. However, if demand for the two products is negatively correlated, then the results are far more valuable. To illustrate this, we consider the demand structure used earlier, where total demand is D (but random now) and demand for the two products is pD and (1 p)D, with p and D being lncorrelated. We assume that y; is the same for both products, equal to y. We derive the value of y* for some specific distributions for p and D and show that is indeed high enough. The calculations are tedious and can be found in the online appendix. Corortary 2.1. (1) Let p be uniform [a,b] and D be uniform [L, H]. We then have a G-xHEra+h’ Y where if H > 2bL then 05H (1 +1n(2b)) — aH ~ L(b a) =e =a} and X= if H <2bL then O.5H(In(H/L)) — a(H ~L) X= DOA) (2) When p follows the theta distribution, iv, F(x) = x, 02 and D is uniform [L, H) with H > aL, c+h =Y)e+a+h)’ _ (05H) "(HY = Le where Y= Fr 1 y@ = iy(HILy To get an idea of what the values of + are in these scenarios, we considered various possible values of[Nagarajan and Rajagopalan: Inventory Moss for Substiutable Products: Optimal Policies and Heuristics [Management Science 548), pp. 1453-1465, 2008 INFORMS 1459 the parameters. We found that in all cases, as one would expect, y° decreases with increased variabil- ity in total demand D and ° increases when the newsvendor ratios increase. For instance, when D varies between (500, 1,000] and p is uniformly dis- tributed between (0, 1) for newsvendor values of 0.91, the value of y* = 0.72. Note that the uniform dis- tribution represents much higher variability than the family of theta distributions. In the above scenario, the value of 7" is higher and equals 0.83 when theta equals 2. We also note that as theta increases the value of 7° also increases. For instance, when theta equals 5, y* = 0.88. Overall, our analysis indicates that, for reasonable values of newsvendor ratios that ‘one may observe in practice, the decoupled policy is, optimal even if D is random but within some reason- able ranges. 22. N-Product Model Using an approach similar to the one in the previous subsection, we consider a single-period N-product substitution problem. We assume that all of the N products have identical values of s, c, 7, and h. We can relax this assumption and have nonidentical parameters, except for the holding cost /. We assume for ease of exposition that the total demand for the N products is D, which is random, and the demand for each product is p,D where p, is a random vari- able independent of D. As before, the cumulative dis- tribution and probability density functions for d, are given by F(.) and f;(.), respectively, and we assume they have a compact support given by [L;,H,]. Let ‘y; denote the fraction of customers for product i that will substitute to another product if product 7 is out of stock and (1 — ¥;) will walk away. We assume that the customers who substitute will be willing to pur- chase any of the other products available in stock. That is, all customers have a strong first choice; if the first choice is not available a proportion will walk away, and the others are indifferent about their sec- ond choice. Let the profit function be denoted, as before, by I1(Q) where Q denotes the vector of Q; val- ues. As in the previous subsection, it can be shown that when Sy,Q; = H, where Dy, H; =H, 1(Q) is separable by product, and this allows us to write a newsvendor-like expression for the order quantity; ie, if y 1), it is opti- ‘mal to follow a base stock policy for each product that is independent of the inventory level of the other product; ie., the partially decoupled structure is opti- mal. Moreover, we show that the optimal order-up-to level is monotonic nondecreasing in the number of periods until the end of the horizon. Without loss of generality, we assume that the dis- count factor 8 =1 and as before the total demand D=1. Let G,(I}, 2) represent the expected profit, when there are n periods until the end of the hori- zon, the starting inventory before ordering of the two products is I} and 12, and an optimal policy is used in all of the remaining periods. The corresponding, dynamic program? is as follows: GRE, Max. Tile B)—elay+B.——2 and Gull, : eames = MO, [ae B.) e048) ~ +f Gesle,—p—1(1—p-B,), 0)AE +] Gusle.~p.B.—(—p) ak 1 +f 6,.10,8,-(1=9)-1p-a,) AF]. Note that IT,(a, B) in the expression below is the single-period expected profit excluding the purchasing cost and is different from, the one used in §2Nagarajan and Rajagopalan: Inoentory Models for Substitutable Products: Optimal Policies ad Heuristics 1460 “Management Science (8, pp. 1453-1456, ©2006 INFORMS Let Gy(I}, 12) =Maxe, 90, 1)(%n(@nr Ba) +60 +20), where X,(a,, 8,) is the expected profit with n periods until the end of the horizon, the starting inventories of the two products is «, and 8, and an optimal policy is used in all the remaining periods. Our first result, which we show using ideas used in the proof of The- orem 2.1 and induction, is as follows. ‘TuoreM 3.1. X, is jointly concave in (a, B,). The concavity of X, immediately implies that if (ai,, Bz) is the maximizer of X,(a B), Gh Re) =e +h) +Xlae, By; TS Mh SBiy that is, when the initial inventory of both products is below the optimal target levels, it is optimal to raise the inventory of both products to the optimal level (aj, B;). Note that, when there is more than one decision variable, in general, concavity does not guar- antee the optimality of a simple base stock policy, wherein the order-up-to level of a product is inde- pendent of the starting inventory level of the other product. Thus, for instance, even if [! < a}, it may not be optimal to raise the inventory of product 1 to a; if B> Br. ‘However, our next result shows that when y=", where y* is defined as in Theorem 22, the op! ‘mal policy in the multiperiod setting also exhibits, the partially decoupled structure. The proof of this result uses two key observations. The first observation shows that, in a multiperiod setting, when y < 9" it is sufficient to restrict our attention to the case when in every period the optimal total inventory level after ordering is at least equal to D. In the single-period model, as we saw, this is the case. To show that this is true requires the use of joint concavity and the calcu- lation of the value of a marginal unit of inventory. The second observation works out the technical details of how the dynamic program preserves separability, ie,, the decoupled structure that we observed in the single-period problem. We also show in the theorem below that the optimal base stock levels are mono- tonic nondecreasing in 1, a property that is known to be true in certain single-product dynamic inventory control problems without substitution. ‘TwEonEM 32. When y < 7° we have Gy te) CU +H) + Xu, Bed CU + Fe) + Xa Fa)s Nea, B By CAR) AR Bide > eS By e(L+ +X); B> ay, B> By (541, Bist) = (ey By) } The above result is quite surprising because, when products are substitutable, one would expect some type of a state-dependent base stock policy to be optimal, where the state contains information about the initial inventory of the products. We note that, because we have not made assumptions about the end of the horizon whereby a myopic policy is not optimal (as can be seen from (b) of the above theo- rem), one cannot simply conclude the property in a dynamic setting from its validity in a single-period setting. The intuition for the above result is as fol- lows. In period n, the cost structure is such that it is, ‘optimal to have a total inventory level of at least D after ordering and before demand realization. This immediately implies that every customer who is will- ing to substitute is satisfied from inventory. Thus, any incremental inventory more than, say, «, can be used to satisfy only the original demand of prod- uct 1. This partially drives the structure of the policy. ‘This intuition can be seen somewhat clearly in the single-period case. In the multiperiod problem, the intuition carries through, with several technicalities needed that can be seen in the proof. The optimality of such a decoupled policy has not been shown in single-period or multiperiod models in the consumer- driven or retailer-driven inventory substitution literature ‘To extend these results to the asymmetric cost case, observe that a key result we used in the proof is that a, +B, > 1 Yn if y <7". Equivalently, in the asym- metric cost case, we have shown that (Qj +Q3) =H, or equivalently a” +8" =1 when D is constant in the single-period model, if 7; = 7. In addition, assuming, the same conditions as in the single-period model to ensure concavity, we can use an approach similar to that used in the symmetric case to obiain an optimal solution to the finite-horizon problem in the asym- metric case. 3.2, Infinite-Horizon Case ‘We continue to restrict our attention to the case when the values of the parameters are such that it is optimal to order a+ > 1 in every period and focus again on the symmetric case for ease of exposition. Denote by G({,,) the total discounted expected infinite profit obtained by starting with an initial inventory I, of product i and using an optimal policy. Then, G satis- fies a functional equation of the form: Cth h) Max (a.Bath, [Bee )-cterB-4—1) He +8) Gle-p—y1—p-B),0)dFNagarajan and Rajagopalan: Inventory Models for Substitutable Products: Optimal Policies and Heuristics Management Science 54(), pp 1453-1466, ©2008 INFORMS 1461 +8)" Clap, B-(1-p) dF +3 600,p-0-p)-rlp-a) 7] Oy We now characterize the solution of the infinite- horizon problem. To do so, we rewrite (9) as Max Sth) ~@ PZ bY [V(a, B) +e +h)* In the proof of the theorem that follows, we demon- strate that G is a bounded continuous function. This is established by standard convergence techniques using fixed point operators. The proof is illustrative in that several desirable properties of the finite-horizon problem hold for the infinite-horizon problem. In par- ticular, concavity and the decoupled policy structure continue to hold in the infinite-horizon case. This allows us to analyze V(a, 8), which leads to the opti- mal policy of the infinite-horizon problem. Let 1218) stath—ée and a’, and :, be such that ot (+2-d-(st-7+h—5e) Fla) (stm+h—de)(l—y) = hc +8) F.-6,) =) Ped ep Be) We then have the following result: ‘TueoReM 33. Let y < y°. Then the optimal policy for the infinite horizon problem is to follow a stationary base- stock policy with order-up-to levels a, and Br, We find that our earlier result for the single- petiod problem extends to the infinite-horizon prob- em. That is, for certain + values, we have a partially decoupled inventory policy similar to the one in the single-period model. In fact, the range of y values is larger in this case (because y* < y®) and can even extend to a value close to 1. For instance, if ¢ = 0.25 a = 0.15 and h =0.12c, and 6 = 0.95, then y° = 0.93, The decoupling effect is sharpened in an infinite- horizon setting, where the optimal inventory level is typically higher than in a single-period setting for the same set of cost parameters. Based on our results, one can perhaps make a more general conjecture. We find that, if the prob- Jem parameters are such that the optimal inventory level of an item is sufficiently high, then all substi- tution demand is fully satisfied, and so the marginal “We note that V(.,.) is the infinite-horizon discounted version of the function X used in earlier discussions on the fnite-horizon, multiperiod model, benefit of additional inventory does not come from. meeting substitution demand and in turn the opti- mal inventory level of the item is not impacted by the inventory level of a substitute. One may then expect a similar phenomenon to occur even in more general inventory models with multiple products provided the substitution fractions are not too high, products are not too dissimilar in terms of the cost param- eters, and total demand D is not highly variable. Then, one can compute inventory levels using sim- ple newsvendor-type equations such as in (1), and these may be near-optimal in the original problem. This conjecture, if validated, can be valuable because substitution problems are in general very complex to analyze. However, such a simple inventory policy is, likely to perform poorly when costs for the various products are substantially different—for the reasons outlined in §2.2. 4. Computational Study In earlier sections, we have pointed out various scenarios where an optimal policy involves a sim- ple modification to the newsvendor policy to adjust for substitution. These scenarios might be seen as restrictive because we have conditions on the value of the substitution parameter and on the demands of the products. Furthermore, much of our analysis was restricted to the two-product case. To explore whether the results obtained are reasonably robust under more general conditions, we conducted a com- putational study. In particular, we wanted to under- stand whether the approach performs well when the problem parameters take on values typically found in one or more retail sectors. So, we performed two numerical studies with two and six products. 4.1. Problem Parameters To explore realistic problem parameters for the study, wwe researched both the academic and trade literature as will be clear shortly’ The manner in which data generated and the parameters used in the study are described next, focusing first on the two-product case. 4.1.1. Demand. In the two-product experiments, the demands for the two products were generated as follows. First, the demand for each product was picked from a uniform distribution so as to test the performance under high levels of variation in demand. The demand values for each product were picked from the range [0, 100] or a subset of it such as [10,30], [20, 80], etc., with two ranges considered * In addition, we talked to several managers in the grocery industry ‘who partiipate in a semester-long management training program, called the Food Industry Management Program at the University of Southern California.[Nagarajan and Rajagopalan: Inventory Models for Substitutable Producis: Optimal Policies an Heuristics 1462 ‘Management Selence 58), pp. 1483-1466, ©2008 INFORMS to reflect high and low demand variance. The range of demand for each product was picked so that the ratio of the mean demand for the two products varied from 1:1 to 1:4, and three different ratios were con- sidered. Furthermore, the demands for the two prod- ucts were generated so that their correlation varied over a wide range; eight correlations were considered: +0.53, +0.36, +0.21, 0, -0.14, -0.27, -0.41, and -0.56. Specifically, we ran the simulations for 500 periods, and demand for the two products across these 500 periods had the above correlations. To identify whether we were considering realistic data points for the demand values in the numer- ical study, we accessed sales and gross margin data from the Dominick’s grocery chain database (http://research.chicagogsb.edu/marketing/databases/ dominicks/index.aspx), used widely by researchers in the marketing literature. Dominick's is a Chicago- based grocery chain with about 81 stores in 2007, and the data were collected in the 1990s for 28 product categories. We found that for almost all of the product categories, excluding items that had sporadic demand (ie, many periods with zero demand), raw demand correlation values between items in a product category varied between 05 and -0.4. If we controlled for seasonality effects in total store demand (due to holidays, store promotions, etc.), these correlation numbers are lower. Furthermore, the demand ratio between items in a product category for items such as canned tuna, bath tissue, paper towels, etc., at most stores was well within the range described above. For instance, the average weekly demand at one of the stores for two brands of canned. tuna was 19 and 27 with correlation in their demands of 0.05, and for paper towels they were 28 and 36 with a correlation of —0.29. Thus, we ensured that the demand variance, correlation, and demand ratio values considered in the study subsumed the values observed in the Dominick's database. 4.1.2. Revenue and Cost Parameters. Data from industry and academic studies (Andersen Consulting 1996, Smith and Agrawal 2000, Gruen et al. 2002) indi- cate that out-of-stock percentages range between 1% and 10% for various product categories in grocery stores.® Hence, the revenue and cost parameters were generated so as to achieve newsvendor ratios for the items in the range of 0.83-0.98. The rationale for this parameter choice is as follows. Assuming probability of no stockout as the service level measure, which is appropriate for retailers because they cannot observe lost sales, this range of newsvendor ratios directly “The grocery managers we talked to confirmed this and indicated that they aim for around 95% in-stock service levels for the items stocked but the actual service levels varied from 90% to 98% for most items. translates into a range of service level measures con- sistent with those found in the studies referenced above. In particular, we used the following values: 0.83, 0.88, 0.93, and 0.98. These parameter choices included the type of revenue and cost data, appropri- ately scaled, observed in the Dominick's data. 4.1.3. Substitution Parameter (7). Several studies (Emmelhainz et al. 1991, Campo et al. 2000, Gruen et al. 2002) have looked at consumer responses to stockouts and substitution levels, primarily in gro- cery or discount stores. The substitution percentages, within a store vary depending on the product char- acteristics and the study, with typical values between 30% and 60%, Furthermore, the studies suggest that customers coming in for products with higher brand, equity are less likely to substitute. Also, substitution rates are higher for perishable products and “necessi- ties” such as bread. Hence, we considered the follow- ing three 7 values in the study: 03, 05, and 0.6. We explored both identical and different values of y for the two products. 4.1.4. Heuristic Approach. ‘The heuristic approach tested was to use the formula in (1) and (2) for the two- product problems and (8) for the six-procuct problems as the order quantity. Note that these order quanti- ties, which are optimal in the stylized settings stud- ied earlier, need not be optimal for the problems tested in the study. We compared the heuristic solution to the average profit for the optimal solution. We are able to obtain the optimal solution because we dis- cretized the state space and computed the profit val- ues for every possible set of inventory values. This was feasible because the demand values were generated, so that they were in the range of 0-100 in the two- productcase, and so we have to evaluate at most 10,000 possible combinations of inventory values because we consider only integer values. In the six-product prob- lems, we restricted the study to products with identi- cal parameters so that we could restrict the search to optimal solutions with identical inventories for all the products, which substantially reduces the state space. We also compared the heuristic with a simple base stock policy that ignores substitution; ie, we simply compute the target base stock level for each item and raise the inventory in each period to this level. Such a base stock policy may be representative of current best practice in the retail grocery sector. For each set of demand, revenue, cost, and substitu- tion parameters, the average profit over 500 demand realizations was computed for the heuristic, optimal solution and the simple base stock policy. 4.1.5. Six Product Experiments. The procedure for generating parameters was largely the same as inNagarajan and Rajagopalan: Inventory Models for Substtutale Products: Optimal Policies and Heuristics Management Science 548), pp. 1853-1466, ©2008 INFORMS 1463 the two-product case except for the following differ- ences. The average demand values for each product was taken to be 25, and two cases were consid- ered: high variance where demand is generated uni- formly from the range [0, 50] and low variance where the range is [15,35]. The demands for the six prod- ucts were generated so that demand for product i was correlated with sum of the demand for prod- ucts 1 through (i— 1). Five different correlation values were used: 0.55, 0.27, 0.05, -0.17, and —0.48. We also considered three possible scenarios for substitution behavior: customers who do not find their first-choice item in stock will make one, two, or three substitution attempts to satisfy their demand. The revenue param- eters and substitution parameter (y) values consid- ered were the same as in the two product experiments except that we considered identical values for all of the products. This allowed us to restrict the search for the optimum to solutions where inventory is identical for all of the six products. 4.2. Performance of the Solution Approach ‘The main results of the two-product experiments are presented in Table 1, which presents the percent- age gap relative to the optimal solution for both the heuristic and the base stock policy ignoring substitu- tion. The results for the six-product experiments are in Table 2. Based on the results, the heuristic seems to perform very well for a wide range of parameter values. The average optimality gap of the heuristic is 0.22% in the two-product problems and 0.25% in the six-product problems. We have the following observa- tions on the performance of the heuristic as a function of the problem parameters. * The performance of the heuristic is better at higher values of the newsvendor ratio. If the newsvendor ratios of both products are 0.88 or higher, the average optimality gap in the two-product (six- product) problems is 0.14% (0.19%) and below 0.8% (1.39%). + In the two-product problems, if the newsvendor ratios are low, the performance of the heuristic is better at low substitution parameters. At lower newsvendor ratios and high y values, the heuristic tends to hold too, little inventory, and this hurts its performance when demand correlation is highly positive. If the newsven- dor ratio is high, the performance is a bit worse when. is lower and demand correlation is highly negative. This is clear from Table 3, which provides detailed results for various demand correlation values for two different newsvendor ratios and y values. * In the six-product problems, the impact of substi- tution parameters is moderated by the number of sub- stitution attempts and the correlation values. So, the worst performance of the heuristic is when is high (y= 0.6) and the maximum number of substitution attempts is 1. When the maximum number of substi- tution attempts is 3, the heuristic performs very well and the worst optimality gap is 1.18%. This is because, when consumers are willing to make more attempts to search and substitute, carrying lower inventory is more desirable. * In two-product problems where newsvendor ratios are different for the two products, it becomes attractive to hold less inventory of the product with the lower newsvendor ratio and deliberately stock out of that item to move customers to the higher- priced product. This is not necessarily a reflection of the poor performance of the heuristic but is instead an indication of the relative unattractiveness of meet- ing a product’s demand from that product's inventory relative to stocking out of the product and meeting that demand from a substitute. But we are not trying to model this type of scenario. Table 1 Results of the Computational Study: Two Products Substitution fraction forthe two items (7, =) Newsvendor 08 05 06 rats of items tand2 Heuristic _—«Base-stock —_—Hourstic—=—«Basestock Heuristic ——_—ase stock 0.83, 083, 009,033 182,285 038,094 = 2.26,340 168,389 1.76,4.14 0.83, 088 013,037 189,259 019,051 «= 210.311 088,11 288,374 0.89, 093 042,052 154,234 007,029 «192,278 «03,087 24,333 0.83, 098 041049 = 139,207 = att,049 =—1.76,247 «008,082 «2.27, 297 0.98, 0.88 023.061 134,203 «= at,097 = 1.64.245 = 0.25,078 2.00, 293 0.88, 093 02,05 «117,175 = otto4t «= 142.208 «009,028 1.71, 248, 0.88, 098 025,056 101,151 014,098 4.23.1.77 008,023 182,211 0.93, 093, 022,045 074,105 = at4,041 = 088,125 © 0.09,0383 1.06, 152 0.98, 098, 021042 «058,088 = 016,098 =r? §=— 02.081 09,113, 0.98, 0.98 005.011 013,018 005,012 01.021 005.011 0.18, 025 ‘Notes. The numbers (x,y) in each col represent, respectively, the average and maximum percentage optimality gap over the 48 problems (tree demand ratios, eight demand correlation values, and two demand valance valves) foreach combination of nevsvendor ratios and substitution fraction. The base stock policy ignores substitution, unite the heurstic approach,[Nagarajan and Rajagopalan: Incenfory Models for Substitutable Products: Optimal Policies and Heuristics 1464 Management Slence 548), pp 1453-1466, ©2008 INFORMS. Table 2 —_ Results of te Computational Study: Sx Products Newsvendor ratio. Masso. of 0.83 0.88 093 098 Substitution substitution fraction atterots Policy Mean Max Mean Mac Mean Max Mean’ Max 03 1 Heurstc =—«021 088 0291.13 02376108 Base stock 094 285 «O71 203.052 1.28 0281.88 03 2 Howistic «0.2772 03S 181025? Basestock 1.10 «313079224 «058141028204 03 a Heuristic «O31 «089038 «139028 «OM 2s Basestock «1.18 «3380812330585 028208 045 1 Heurste 010089013083 02108001024 Basestock 1.19 908089217 088150 037-278 045 2 Houistic «0.040.501 073 OF? Base stock 1693.41.07 281A 16904388 045 3 Heuristic —«.072t 02809) 028 08s ats OT Basestock 169 «3.981.192.7176 169 Od 3.88 06 1 Houistc «158452030128 008 0G m1 88 Base stock 148 «333113282 OBA 224 OMB 8.86 06 2 Heuristic «0.82250 012 0015S 008025 Basesstock 207 «426148277103, 288084 5.18 06 3 Houristic = «0.381.487 02519088028 Base stock 244 «487163 3.15 108292088588 ‘oie. The Mean and Max values represent, respectively, the average and maximum percentage optimality gap over 10 problems (tive demand correlation values and two demand variance values foreach combination of nwsvenor rato, subsiution racton, nd maximum numberof substitution attempts The detailed results (not reported due to space lim- itations) indicate that the optimality gap is larger when the demand variance is higher. This is to be expected because higher variance implies that the heuristic inventory may be less than optimal espe- cially if demand is positively correlated. However, we do not find the optimality gap to be sensitive to the ratio of mean demand of the two products. We solved several problems with different y values for the two products but did not find the results to be very sensi- tive to differences in y, values. Instead, the results are primarily sensitive to the level of both 7, values. The heuristic policy carries lower inventory than the optimum in most scenarios but sometimes does carry higher inventory. The simple base stock policy that ignores substitution always carries higher inventory than the optimum. In the six-product problems, the tal 3 Optimality Gap of Heuristic for Dtlerent Demand Correlation Values (Two Products) ‘Sibson Demand carlton Newsvendor rattan sate (y=)_055 096 021-003-014 —027 -042 ~096 093,085 06 O11 GtT O06 aoe om 009 024 O38 03 038 032 028 om 012 0a OMS O45 og3.o8 = 08191 17 157 153 11 O77 0M oor 03 029 018 005 092 0 037 03 o2t ‘Note. n these problems, the demand variance is high and te rato of mean demands is 1 inventory determined by the heuristic is on average 3% above the optimum inventory whereas the sim- ple base stock policy carries on average 11% higher inventory. The heuristic inventory varies between 0.9 and 1.14 times the optimum inventory, and the simple base stock policy inventory varies between 1 and 1.4 times the optimum inventory. (We have not provided detailed results on inventories due to space limita- tions.) When the substitution fraction y is high but the maximum number of substitution attempts is low, the heuristic typically carries insufficient inventory rel- ative to the optimum. However, if the newsvendor ratio is high, 7 is low (=0.3), and the number of sub- stitution attempts is high, the heuristic policy may carry higher inventory than the optimum. Comparing the performance of the heuristic with the base stock policy ignoring substitution, we find that the profit of the heuristic policy is on average 1.5% (0.7%) better than the policy that ignores sub- stitution in the two-product (six-product) problems. Note that these are increases in gross profits and may translate into a significant increase in net prof- its. (Kroger, the largest grocery chain in the United States, had gross profits of $15 billion and net prof- its of $1 billion in 2005.) Moreover, the heuristic is far more consistent in its performance than the base stock policy. For instance, in the six-product prob- ems, the heuristic outperforms the simple base stock policy in more than 97% of the problem instances.Nagarajan and Rajagopalan: Inventory Models for Substtutable Products: Optimal Policies and Heuristics Management Science 54(8), pp 1453-146, ©2008 INFORMS 1465 Furthermore, the maximum optimality gaps are sig- nificantly larger for the base stock policy as compared to the simple heuristic. However, in the two-product, problems, there is little difference between the two approaches when newsvendor ratios are very high (0.98). Indeed, the optimality gap is very small for either policy, and in such scenarios the tetailer can ignore stockout-induced substitution effects in mak- ing inventory decisions. Thus, we find that the heuristic performs well under realistic conditions. For most branded prod ucts, retailers and manufacturers aim for high service levels, generally around 90%-95%. In these cases, the optimality gap is likely to be small (less than 1%) as is clear from Tables 1 and 2. Thus, the heuristic approach is likely to be ideal for nonperishable items that have high brand equity, high newsvendor ratios or service levels, and low substitution fractions and when the consumer who is willing to substitute is, likely to make two or three substitution attempts. If service levels are very high (98% or more), then our results suggest that the retailer can essentially ignore substitution in making inventory decisions. Clearly, ‘our numerical study is not comprehensive, and so the results have to be interpreted and used with some caution. ‘An important input to using the model and heuris- tic outlined here is the demand for each item, and our approach implicitly assumes that the original demand for a product is known. In reality, we observe only sales, and this comprises both original and sub- stitution demand. However, Anupindi et al. (1998) and Kék and Fisher (2004) have proposed and tested empirical methods for estimating original demand and substitution fractions from sales data when there is stockout-based substitution. These methods can be used together with our heuristic to determine stock- ing levels. 5. Conclusions and Future Research We have derived the optimal inventory policy for certain reasonable problem parameters for a single decision maker in a model with two substitutable products whose demands are negatively correlated and that are partial substitutes in stockout situations in both single-period and multiperiod scenarios. We show that the optimal inventory level for a product is partially decoupled from the inventory level of the other product in such scenarios. The numerical study shows that the approach and results developed in the paper are robust and perform well under a variety of conditions not explicitly modeled, and in particu- lar under scenarios observed in reality. The approach developed in the paper is most valuable to a retailer, making inventory decisions of items within a cate- gory having the following characteristics: substitution across the items is moderate, demand of the category is not highly variable, and desired service levels are high (over 85%) but not too high (>98%). Numerous branded product categories and items in the retail sec- tor have these characteristics. ‘There are natural extensions of this work. First, we would like to extend the analytical results to a more general N-product case. Second, we could explore experimentally how the decoupled policies perform in multiproduct, multiperiod problems with replen- ishment lead times, random demand D, and gen- eral substitution behavior. Finally, we would like to investigate the contractual aspects of a single decision maker managing the inventories of the two products versus two independent players managing them. The closed-form nature of inventory policies we have derived may facilitate this analysis. 6. Electronic Companion An electronic companion to this paper is available as part of the online version that can be found at http:// ‘mansci journal informs.org/. Acknowledgments The authors are grateful to the department editor for his valuable suggestions and guidance and to the associate edi- tor and referees for their detailed comments, which have improved the paper significantly. The authors are espe- ally grateful to one of the referees for pointing out the ‘equivalence of their model tothe transshipment model with backordering and suggesting ideas for streamlining and generalizing the analysis. References Andersen Consulting, 1996. Where to look for incremental sales gains: The retail problem of out-of-stock merchandise. The Coca-Cola Retailing Research Council ‘Anupindi, R, M. Dada, 5. Gupta. 1998. Estimation of consumer ‘demand with stockout based substitution: An application to vending machine products. Marketing Sci. 17(4) 406-423. Axsater, 8.2003, A new decision rule for lateral transshipment in inventory systems. Menagement Sc. 49(9) 1168-1179. Bassok, ¥. R, Anupindi, R. Akella, 1999. Single period multi product inventory model with substitution. Oper. Res. 47(8) 632-642. ‘Campo, K, E, Gijsbrechts, P. Nisol. 2000. Towards understar consumer response to stock-outs, J. Retailing 76 219-22. Deniz, B, A. Scheller-Wolf, I. Karaesmen. 2004. Managing invento- ries of perishable goods: The effect of substitution. Graduate ‘School of Industrial Administration Working Paper 2004-E55, ‘Carnegie Mellon University, Pittsburgh. Emmelhainz, M,,].Stock, L. Brumelhainz. 1991. Consumer response to retail stock-outs. J Retailing 67 138-147, Exnst,R, P. Kouvelis. 1999. The effect of selling packaged goods on inventory decisions. Management Sci 45(8) 1142-1156. Gaur, V,, D. Honhon. 2005. Assortment planning and inventory decisions under locational choice model. Working, paper, Stern School of Business, New York University, New York. Gerchak, ¥, D. Mossman. 1992. On the effect of demand random- ‘ness on inventory and costs. Oper. Res. (4) 804-807.Nagarajan and Rajagopalan: Inventory Models for Substtutable Products: Optimal Policies and Heuristics 1466 ‘Management Science 5(8, pp. 1453-1466, 02008 INFORMS Gruen, TW, D. Corsten, S. Bharadwaj, 2002. Retail Out of Stocks: A ‘Wordcoide Examination of Causes, Rates, and Consumer Responses. Grocery Manufacturers of America, Washington, D.C. Kok, G, M. L, Fisher. 2004. Demand estimation and assort- ‘ment optimization under substitution: Methodology and appli- ‘ation. Working paper, Fuqua School of Business, Duke University, Durham, NC. Kok, G,, ML. Fisher, R. Vaidyanathan. 2006. Assortment planning: Review of literature and industry practice, Working paper Fuqua School of Business, Duke University, Durham, NC. ‘Mahajan, S,,G. van Ryzin. 1998. Retail inventories and consumer choice.'S. Tayur, R. Ganeshan, M. Magazine, eds. Quantita- tive Models for Supply Chain Management, Chap. 17. Kluwer, ‘Amsterdam: ‘Mahajan, S, G. van Ryzin. 2001. Stocking retail assortments under dynamic consumer substitution. Oper. Res. 49(3) 334-351. McGillivray, A, E Silver. 1978, Some concepts for inventory control ‘under substitutable demand. INFOR 16 47-63. [etessine, S,,.N. Rudi, 2003. Centralized and competitive inventory models with demand substitution. Oper. Res. 51(2) 329-385. Parlar, M. 1985. Optimal ordering policies for a perishable and substitutable product—A Markov decision model. INFOR 23, 182-195, Parlar, M, S. Goyal. 1984. Optimal ordering decisions for two sub- ‘stitutable products with stochastic demand. OPSEARCH 21 1.5. Pasternack, B, Z. Drezner. 1991. Optimal inventory policies for ‘substitutable commodities with Stochastic demand. Naval Res Logis. 38 221-240. Rajaram, K, C. S. Tang. 2001. The impact of product substitution, ‘on retail merchandising, Eur J. Oper. Res. 138 582-601 Robinson, L. W. 1990. Optimal and approximate policies in multi- period, mult-location inventory medels with transshipments (Oper. Res. 38(2) 178-295, Rudi, N., 8. Kapur, D. F. Pyke. 2001. A tworlocation inventory model with transshipment and local decision making, Manage- ‘ment Sei. 47(12) 1668-1680. Smith, S, A.,.N. Agrawal. 2000. Management of multi-item retail inventory systems with demand substitution. Oper Res, 48(1) 50-64, Van Mieghem, J. A., N. Rudi. 2002. Newsvendor networks: lnven- tory management and capacity investments with discretionary activities. Manufacturing Service Oper. Management 4(4) 13-33. Yang, H.,, L. Schrage. 2002. An inventory anomaly: Risk pooling ‘may increase inventory. Working paper, Graduate School of, Business, University of Chicago, Chicago.