Production, Manufacturing and Logistics

Accepted Manuscript

Optimal inventory policy for two substitutable products with customer

service objectives

Xu Chen, Youyi Feng, Matthew F. Keblis, Jianjun Xu

PII: S0377-2217(15)00326-4
DOI: 10.1016/j.ejor.2015.04.033
Reference: EOR 12905

To appear in: European Journal of Operational Research

Received date: 29 March 2014

Revised date: 10 April 2015
Accepted date: 17 April 2015

Please cite this article as: Xu Chen, Youyi Feng, Matthew F. Keblis, Jianjun Xu , Optimal inventory
policy for two substitutable products with customer service objectives, European Journal of Operational
Research (2015), doi: 10.1016/j.ejor.2015.04.033

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Optimal Inventory Policy for Two Substitutable Products with Customer Service
Objectives

Xu Chen
School of Management and Economics,
University of Electronic Science and Technology of China, Chengdu, P.R.China

E-mail: xchenxchen@263.net

T
Youyi Feng

IP
Zaragoza Logistics Center, Zaragoza, Spain

E-mail: yyfeng@live.com

CR
Matthew F. Keblis

US
Macquarie Graduate School of Management, Macquarie University, Macquarie Park, NSW, Australia

E-mail: matthew.keblis@mgsm.edu.au
AN
Jianjun Xu1
Zaragoza Logistics Center, Zaragoza, Spain
M

Phone: +34 976019260, Fax: +34 976077601

E-mail: jxu@zlc.edu.es
ED

Abstract:
PT

We consider a firm facing stochastic demand for two products with downward, supplier-driven substitution and
customer service objectives. We assume both products are perishable or prone to obsolescence, hence the firm
faces a single period problem. The fundamental challenge facing the firm is to determine in advance of observing
CE

demand the profit maximizing inventory levels of both products that will meet given service level objectives.
Note that while we speak of inventory levels, the products may be either goods or services. We characterize
AC

the firm’s optimal inventory policy with and without customer service objectives. Results of a numerical study
reveal the benefits obtained from substitution and show how optimal inventory levels are impacted by customer
service objectives.

Key words: Inventory management, capacity management, substitution, perishability, customer service ob-
jective

1 Corresponding author
 ACCEPTED MANUSCRIPT

1 Introduction

Product substitution is important to many firms. For example, manufacturers of components used

in consumer electronics (CE) routinely design new generations of their devices in a way that allows

more than one model of a generation to work in the end-product of a CE manufacturer. If one such

model being supplied to a CE manufacturer for use in an end-product stocks out, then the component

manufacturer is often able to still meet the CE manufacturer’s demand by providing, in lieu of the

model that has stocked out, a superior model. This is an example of downward, supplier-driven

T
substitution. Downward refers to a supplier meeting customer demand for a product j by providing a

IP
product i that has better quality and/or more functionality than product j. Supplier-driven indicates

CR
it is the supplier of the product that reacts to the stockout and takes action. This is unlike the

situation in a restaurant for instance where after a customer learns that his first choice is unavailable,

he picks another item from the menu, which is an example of customer-driven substitution.

US
The earliest study of substitution was undertaken several decades ago by McGillivray and Silver
AN
(1978). There now exists a stream of research on this topic. Interestingly, none of it considers

customer service objectives, e.g. fill rates or in-stock probabilities (see Cachon and Terwiesch (2012)

for a thorough discussion of these measures). Juxtaposed with this reality are 1) surveys such as the
M

one by Aberdeen (Kay (2005)), which reported that 70% of companies feel that providing a high level

of service is critical to their business operations, and 2) the fact that in the more general operations
ED

and supply chain management literature, there is a steady flow of research with a service level aspect

(e.g. see Bensoussan et al. (2011), Chen and Shen (2012), Alptekinoglu et al. (2013), and Wang et
PT

al. (2014)). The lacuna in the substitution literature with respect to service levels, combined with

the increasing importance of service levels to business, motivates our study of a substitution problem
CE

with customer service objectives.

More specifically, we consider a firm that faces stochastic demand for two products where an in-stock
AC

probability service level objective has been set for each product. The in-stock probability of a product

is the probability the firm has inventory to meet every customer demand for the product. We designate

the two products “1” and “2” with product 1 as good as or superior to product 2 in every respect, hence

product 1 can substitute for product 2. As is common practice in the CE industry, we assume that the

substitution is supplier-driven. Furthermore, we also assume both products are perishable or prone to

obsolescence, i.e. both products have short life cycles. The firm under consideration therefore faces a

single period problem, which is to determine in advance of observing demand the profit maximizing

1
 ACCEPTED MANUSCRIPT

inventory/stocking levels of both products that will meet given customer service objectives. Note that

throughout the paper we speak of inventory or stocking levels, however the products may be either

goods or services and in the latter case we are interested in determining capacities to put in place.

Also, we use increasing to mean non-decreasing and decreasing to mean non-increasing.

The rest of the paper is organized as follows. In the next section, we review the relevant substitution

literature. In the third section, we begin our study of the problem setting without customer service

objectives. We develop a general model of inventory policies for the two products with downward,

supplier-driven substitution. In section 4, we introduce into the analysis service level constraints of

T
the in-stock variety and identify the optimal inventory policy for the firm. In Section 5, we present the

IP
results of a numerical study that reveal the benefits obtained from substitution and show the impact

CR
of customer service objectives on optimal inventory levels. Section 6 summarizes our findings and

highlights future research opportunities.

2 Literature Review US
AN
As already noted, the earliest investigation of inventory management involving substitution was un-

dertaken by McGillivray and Silver (1978). In their study inventory was managed using an order-up-to
M

policy with a fixed review period. Some analytical results were established for limiting cases including

complete-substitutability and no-substitutability. For cases in-between these two extremes, a heuristic
ED

approach was developed and tested for establishing order-up-to levels.

Research undertaken subsequent to McGillivray and Silver (1978) can be divided into whether it is
PT

consumer or supplier -driven and whether it involves deterministic or stochastic demand. We refer the

reader to the survey by Pentico (2008) for research that has been done on consumer-driven substitution
CE

and to the papers of Hsu et al. (2005) and Bardhan et al. (2013) for reviews of the research on

substitution that has been done in the deterministic context. In the rest of this literature review we
AC

focus on the previous research that involved stochastic demand and supplier-driven substitution.

Pasternack and Drezner (1995) investigated a single-period two-product setting where both products

can substitute for each other. They showed that the optimal profit function is concave and they

derived formulas for the optimal order quantities. The authors also derived formulas for the optimal

stocking levels under specific conditions in the case where only one product can serve as a substitute.

Van Mieghem (1998) studied the setting of a two-product firm that has the option to invest in product-

dedicated resources and/or in a flexible resource that can produce either product. He showed that

2
 ACCEPTED MANUSCRIPT

investment in a flexible resource can be advantageous even with perfectly correlated product demands.

Bassok, et al. (1999) considered a single-period multi-product setting with downward substitution.

Under various assumptions, they showed the concavity and submodularity of their expected profit

function. The authors also developed explicit expressions for the first partial derivatives of the profit

function and derived the optimal order policy.

Van Mieghem and Rudi (2002) extended the multidimensional newsvendor model of Van Mieghem

(1998) to a newsvendor network that allows for multiple products and multiple processing and storage

T
points. Using this framework, they studied various problems of capacity and inventory management,

IP
including assembly, commonality, distribution, flexibility, substitution and transhipment.

CR
Axsater (2003a, 2003b) examined a single-echelon inventory system consisting of a number of parallel

warehouses. While the warehouses are normally replenished from an outside supplier, lateral tran-

shipment between the warehouses is allowed. When demand for a low-quality item cannot be met, a

US
high-quality item can be substituted, which is referred to as a unidirectional transhipment.
AN
Karaesmen and van Ryzin (2004) studied an overbooking problem with multiple reservation and

inventory classes and a general substitution structure. They formulated a two period model, where

reservations are accepted in the first period and then in the second period after cancelations are
M

realized the remaining customers are assigned to the various inventory classes to maximize the net

benefit.
ED

Rao, et al. (2004) also considered a single-period, multi-product setting with downward substitution

but with setup costs. They formulated a two-stage integer stochastic program with recourse where
PT

the first stage decision variables determined which products to produce and in what quantity, while
CE

the second stage decision variables determined how the products were allocated to meet demand.

Shumsky and Zhang (2009) studied a multi-period, multi-product (each product corresponds to a
AC

class of demand) capacity allocation problem with downward substitution and a single replenishment

opportunity at the beginning of the first period. They showed that the optimal solution is to first

use any available inventory to satisfy same-class demand and then upgrade customers until inventory

reaches a protection limit.

Most recently, Deflem and Van Nieuwenhuyse (2013) examined the problem of stocking two products

with downward substitution in both single period and infinite horizon settings. They assumed a base

stock model for the management of each inventory, with full backordering in the infinite horizon case

and lost sales for the single period model, and then proceeded to show the cost conditions under which

3
 ACCEPTED MANUSCRIPT

substitution outperforms having separate distinct inventories.

Interestingly, none of the existing literature on product/capacity substitution where demand is stochas-

tic concerns itself with the impact of service level constraints on stocking levels, despite the fact that

the study of service-level constraints has received much attention in the more general inventory man-

agement literature. Our paper appears to be the first to characterize the optimal inventory policy of

a system with product substitution and service level requirements. The numerical study contained

herein provides insight into how stocking levels are impacted by service level requirements and how

profitability is affected by substitution.

T
IP
3 Substitution without Service Commitments

CR
3.1 Model Description

US
We consider a firm that faces stochastic demand for two products, designated “1” and “2”, where

product 1 can downward substitute for product 2. Furthermore, we assume that the substitution is
AN
supplier-driven. In addition to the substitutability aspect, we assume that both products are perishable

or prone to obsolescence, hence the firm faces a single period problem. The objective of the firm is to
M

determine in advance of observing demand the profit maximizing inventory levels of both products.

In this section customer service objectives are ignored. Table 1 defines the parameters and variables
ED

used throughout the paper; other notation is introduced as needed.

PT

Table 1: Parameter and Variable Definitions

CE
AC

4
 ACCEPTED MANUSCRIPT

Notation Description
Di Random demand for product i, i = 1, 2.
f (ξ1 , ξ2 ) Joint probability density function of demand for products 1 and 2.
F (ξ1 , ξ2 ) Joint distribution function of demand for products 1 and 2.
fi (ξi ) Marginal density function of demand for product i, i = 1, 2, e.g., f1 (ξ1 ) =
R∞
0 f (ξ1 , ξ2 )dξ2 .
Fi (ξi ) Probability distribution function of demand for product i, i = 1, 2, e.g.,
Rξ Rξ R∞
F1 (ξ1 ) = 0 1 f1 (ζ)dζ = 0 1 0 f (ζ, η)dηdζ.

T
wi Unit procurement/production cost of product i, i = 1, 2.

IP
pi Unit sales price of product i, i = 1, 2.
vi Unit salvage value of product i, i = 1, 2.

CR
αi Service level objective for product i, i = 1, 2.
Qi Inventory level after procuring/producing product i, i = 1, 2.
T (Q1 , Q2 ) Expected profit of firm with inventories Q1 and Q2 but without any

service commitments. US
AN
In order for the model we develop in this paper to “make sense”, the parameters defined in Table 1

must satisfy certain conditions. We next specify these conditions as assumptions:

M

Assumptions: A1) pi > wi > vi > 0 for i = 1, 2; A2) p1 > p2 ; A3) v1 > v2 ; A4) p2 − v1 > 0.

A1 simply states there is a positive profit margin on each unit of product stocked and then sold to a
ED

customer. On the other hand, the salvage value is less than the procurement/production cost, hence

there is a loss if an item goes unsold. This assumption incentivizes the firm to stock product, but in
PT

reasonable amounts. A2 implies that the gain from meeting a demand for product 1 is always greater

than the gain from meeting a demand for product 2 using product 1. A3 indicates that surplus product
CE

1 can be salvaged for more than surplus product 2, hence the firm should always use product 2 to

meet demand for product 2 when product 2 is available. The implication of A4 is that more is gained
AC

from using product 1 to meet a product 2 demand than from salvaging a unit of product 1. Note that,

p2 − v1 is the unit gain from substitution. Without this assumption, substitution is not profitable.

Using the above notation and assumptions, we next derive the expected profit function when immedi-

ately prior to the realization of demand the inventory levels of products 1 and 2 are Q1 and Q2 units

respectively. The initial inventory of products 1 and 2 is denoted by (x1 , x2 ) and the inventory levels

after procurement/production (but before demand is observed) by (Q1 , Q2 ). To start, we assume that

x1 = 0 and x2 = 0. The objective is to maximize the expected profit (net revenue) after stocking,

5
 ACCEPTED MANUSCRIPT

which is given by:

T (Q1 , Q2 ) = E{p1 (D1 ∧ Q1 ) + p2 [D2 ∧ (Q2 + (Q1 − D1 )+ )]

+ v1 [(Q1 − D1 )+ − (D2 − Q2 )+ ]+ + v2 (Q2 − D2 )+ } − w1 Q1 − w2 Q2 , (3.1)

where for x, y ∈ R, x∧y = min(x, y). The problem of optimizing expected profit when (x1 , x2 ) = (0, 0)

is generalized to the problem of optimizing expected profit for arbitrary (x1 , x2 ). In the first instance

the goal is to find

T
max T (Q1 , Q2 )
Qi ≥0,i=1,2

IP
and in the second instance to find

CR
max [T (Q1 , Q2 ) + w1 x1 + w2 x2 ]
Qi ≥xi ,i=1,2

which is equivalent to finding:

max
US
Qi ≥xi ,i=1,2
T (Q1 , Q2 ) (3.2)
AN
since w1 x1 + w2 x2 is constant.
M

3.2 Monotone Optimal Ordering Policies with Threshold Curves

ED

In this subsection, we first establish properties of the profit function T (Q1 , Q2 ). We then use these

properties to identify three monotone curves that are helpful in characterizing the optimal inventory
PT

policy.

We begin with some definitions. For x, y ∈ R, x ∨ y = max(x, y). A sublattice of RK , say V, is a set
CE

such that for all x0 = (x01 , · · · , x0K ) and x00 = (x001 , · · · , x00K ) belonging to V, the meet x0 ∧ x00 = (x01 ∧

x001 , · · · , x0K ∧ x00K ) and the join x0 ∨ x00 = (x01 ∨ x001 , · · · , x0K ∨ x00K ) also belong to V. Let ψ(x1 , · · · , xK ) be
AC

defined on V. Then ψ is called submodular if for any x0 , x00 ∈ V, ψ(x0 )+ψ(x00 ) ≥ ψ(x0 ∧x00 )+ψ(x0 ∨x00 ).

Lemma 1. T (Q1 , Q2 ) is a jointly concave and submodular function of Q1 and Q2 .

Proof. See the Appendix.

This lemma establishes the existence of global maximizers for T (Q1 , Q2 ), which in the sequel are fun-

damental in characterizing the optimal inventory policy. Throughout the paper the partial derivative

of the profit function T (Q1 , Q2 ) with respect to Qi is denoted by T i (Q1 , Q2 ), i = 1, 2. From (3.1) and

6
 ACCEPTED MANUSCRIPT

Leibniz’s rule for differentiating integrals we obtain:

Z Q1 Z Q1 +Q2 −x
1
T (Q1 , Q2 ) = (p2 − p1 ) F1 (Q1 ) + (v1 − p2 ) f (x, y) dydx + p1 − w1 (3.3)
0 0

and

Z Q1 Z Q1 +Q2 −x 
2
T (Q1 , Q2 ) = (v1 − p2 ) f (x, y) dydx − F (Q1 , Q2 )
0 0
+ (v2 − p2 ) F2 (Q2 ) + p2 − w2 . (3.4)

T
The partial derivative T i (Q1 , Q2 ) is the marginal profit of product i. It gives the rate at which the

IP
expected profit changes per additional unit of product i stocked for interior feasible stocking quantities.

CR
A comparison of the partial derivatives reveals which product should be stocked a priori.

In addition, we establish as an identity the difference of the two marginal profits:

1 2 US
T (Q1 , Q2 ) − T (Q1 , Q2 ) = −(p1 − p2 )F1 (Q1 ) + (p2 − v1 )
Z

0
Q2 Z

Q1
∞
f (x, y)dxdy (3.5)
AN
+(v1 − v2 )F2 (Q2 ) + (p1 − p2 ) − (w1 − w2 ).

It follows from this identity that the next lemma holds:

M

Lemma 2. T 1 (Q1 , Q2 )−T 2 (Q1 , Q2 ) is decreasing (non-increasing) in Q1 and increasing (non-decreasing)

ED

in Q2 .

Proof. Omitted.
PT

This lemma establishes that T (Q1 + z, Q2 − z) is supermodular in (z, −Q1 ) and (z, Q2 ), which in our
CE

two product case is equivalent to diagonal dominance. The ensuing corollary follows from this lemma.

Corollary 1. T (Q1 + z, Q2 − z) is supermodular in (−Q1 , Q2 , z).

AC

An implication of this corollary is that when adding to the stock of product 1, it is optimal to

simultaneously decrease the inventory of product 2 by the same number of units. Corollary 1 also

reveals that T (Q1 , Q2 ) is anti-multimodular in (Q1 , Q2 ) - see Li and Yu (2014) for an in depth discussion

of multimodularity/anti-multimodularity defined in real space, the space assumed in this paper, as

opposed to on the integers as multimodularity has traditionally been defined (see Hajek (1985), Murota

(2005)). This implies that 1) T (Q1 + z, Q2 − z) is supermodular in (z, Q2 ), i.e. the quantity z and the

inventory of product 2 are complements, and 2) T (Q1 + z, Q2 − z) is submodular in (Q1 , z), i.e. the

quantity z and the inventory of product 1 are substitutes.

7
 ACCEPTED MANUSCRIPT

We now define threshold curves that are used to guide the firm’s stocking decisions:

S1 (x2 ) = sup{x1 ≥ 0 : T 1 (x1 , x2 ) > 0},

S2 (x1 ) = sup{x2 ≥ 0 : T 2 (x1 , x2 ) > 0},

S0 (x1 ) = inf{x2 ≥ 0 : T 1 (x1 , x2 ) − T 2 (x1 , x2 ) > 0}.

These curves can be obtained respectively by setting the right-hand sides of (3.3), (3.4) and (3.5) to

zero. By Lemma 1 T (x1 , x2 ) is concave in xi , i = 1, 2 respectively so that T 1 (x1 , x2 )(T 2 (x1 , x2 )) is

T
decreasing in x1 (x2 ). Thus S1 (x2 ) is the largest x1 for a given x2 for which augmenting product 1 is

IP
profitable, and S2 (x1 ) is the largest x2 for a given x1 for which augmenting product 2 is profitable. By

Corollary 1, T (x1 +z, x2 −z) is supermodular in (z, x2 ) and thus T 1 (x1 , x2 )−T 2 (x1 , x2 ) is increasing in

CR
x2 . Thus S0 (x1 ) is the smallest x2 for a given x1 above which replenishing product 1 is more profitable

than replenishing product 2.

US
To illustrate the sort of information provided by these threshold curves, consider the curve S1 (x2 ). For

a fixed x2 , by Lemma 1, we have that the derivative T 1 (x1 , x2 ) is decreasing in x1 . We also have that
AN
if x1 < S1 (x2 ), then T 1 (x1 , x2 ) > 0 and it is optimal to raise the inventory of product 1 (up to S1 (x2 ));

otherwise, T 1 (x1 , x2 ) ≤ 0 and it is better to do nothing. In addition, as T (x1 , x2 ) is continuous in

M

(x1 , x2 ), at x1 = S1 (x2 ), T 1 (x1 , x2 ) = 0. In other words, S1 (x2 ) indicates the optimal inventory level

for product 1 given the amount of product 2 is fixed at x2 .

ED

Based on the properties established by Lemmas 1 and 2 as well as Corollary 1, we can obtain monotone

properties of the three threshold curves and use the curves to characterize the optimal inventory policy.
PT

Lemma 3. (1) S1 (x2 ) and S2 (x1 ) are decreasing and S0 (x1 ) is increasing.
CE

(2) The slope of S1 (x2 ) is less than or equal to −1 and the slope of S2 (x1 ) is greater than or equal to

−1.
AC

Proof. See the Appendix.

The decreasing nature of S1 (x2 ) can be explained as follows. As already noted, for a fixed x2 , S1 (x2 )

indicates the optimal inventory level for product 1. If the inventory of product 2 is greater than x2 ,

say x02 , then the potential need to substitute product 1 for product 2 is reduced, and hence the optimal

amount of product 1 to stock will be no greater than the optimal inventory level when the stock of

product 2 is x2 , i.e. S1 (x02 ) ≤ S1 (x2 ). A similar story holds for S2 (x1 ). The more product 1 available,

the less need there is for product 2 as product 1 can substitute for product 2, if necessary. Hence, the

8
 ACCEPTED MANUSCRIPT

optimal inventory level for product 2 is decreasing in the inventory level of product 1. In addition,

there is a lower bound of −1 on the rate at which S1 (x2 ) and S2 (x1 ) decrease. This fits with intuition

as one expects the optimal inventory level of a product to be more sensitive to its own inventory level

than to that of another.

x2

S1
S0

T
IP
(Q1* , Q2* )
S2

CR
x1

Figure 1: Optimal policy without customer service objectives.

US
We next discuss how the optimal stocking levels can be determined using the above defined threshold
AN
curves {Si : i = 0, 1, 2}. The first quadrant of the inventory plane (See Figure 1) is divided into two

regions by the curve S0 (x1 ). In the region to the left of S0 (x1 ) we have T 1 (x1 , x2 ) − T 2 (x1 , x2 ) > 0, i.e.

the marginal profit of stocking product 1 is higher than the marginal profit of stocking product 2, hence
M

if any inventory adjustment is made, it will only be for product 1. On the other hand, in the region to

the right of S0 (x1 ) we have T 1 (x1 , x2 ) − T 2 (x1 , x2 ) < 0, i.e. the marginal profit of stocking product 1
ED

is lower than the marginal profit of stocking product 2, hence if any inventory adjustment is made, it

will only be for product 2. At the point where the three threshold curves intersect, which we denote
PT

by (QI1 , QI2 ), we have T 1 (Q1 , Q2 ) = T 2 (Q1 , Q2 ) = 0. If (x1 , x2 ) the initial inventory of each product is

such that x1 ≤ QI1 and x2 ≤ QI2 , then the point of intersection is a global maximizer of T (Q1 , Q2 ). In
CE

the case where T (Q1 , Q2 ) is simply concave, there may be multiple global maximizers. When there are

multiple global maximizers, the curves S1 and S2 overlap/intersect at more than one point (at each
AC

such point T 1 (Q1 , Q2 ) = T 2 (Q1 , Q2 ) = 0) and we designate the maximizer with the smallest Q1 as

(Q∗1 , Q∗2 ). In the case where T (Q1 , Q2 ) is strictly concave, (QI1 , QI2 ) the point of intersection is the only

global maximizer and we denote it by (Q∗1 , Q∗2 ). For x1 ≤ QI1 and x2 ≤ QI2 , it is optimal to increase

the initial inventory of product 1 by Q∗1 − x1 and the initial inventory of product 2 by Q∗2 − x2 . If the

initial inventory (x1 , x2 ) is to the right of S0 and x1 > QI1 , when x2 < S2 (x1 ) it is optimal to increase

the inventory of product 2 by S2 (x1 ) − x2 units, whereas when x2 ≥ S2 (x1 ) it is optimal to do nothing.

Analogously, if the initial inventory (x1 , x2 ) is to the left of S0 and x2 > QI2 , when x1 < S1 (x2 ) it is

optimal to increase the inventory of product 1 by S1 (x2 ) − x1 units, whereas when x1 ≥ S1 (x2 ) it is

9
 ACCEPTED MANUSCRIPT

optimal to do nothing. The following theorem formalizes the previous discussion.

Theorem 1. For any (x1 , x2 ) ∈ R2 , let x̄1 = S1 (x2 ) ∧ Q∗1 and let x̄2 = S2 (x1 ) ∧ Q∗2 . Then the optimal

order-up-to level is (x1 ∨ x̄1 , x2 ∨ x̄2 ).

Proof. See the Appendix.

4 Optimal Ordering Policies with a Service Commitment

T
In this section, we consider the situation where the firm commits to providing a minimum level

IP
of service, which is defined in terms of an in-stock probability. For product i, the target in-stock

CR
probability is given by αi , which is a number that lies between 0 and 1. The selection of a value αi

indicates that the firm wants to stock-out of product i with probability no greater than 1 − αi . Hence,

US
when the stocking level for product i is Qi , for i = 1, 2, the service commitment can be characterized

mathematically as
AN
Pr{D1 ≤ Q1 } ≥ α1 and Pr{D2 ≤ Q2 + [Q1 − D1 ]+ } ≥ α2 . (4.6)
M

The latter of the foregoing inequalities takes into account any substitution of product 1 for product

2, i.e., where product 1 inventory is used to meet demand for product 2.

ED

For an initial inventory (x1 , x2 ), the firm’s decision-making problem is to choose inventory levels

(Q1 , Q2 ) that maximize (3.2) while satisfying (4.6). In order for an inventory level Q1 to be feasible
PT

for product 1, we must have Q1 ≥ F1−1 (α1 ), where F1 (Q1 ) = P r{D1 ≤ Q1 } denotes the in-stock

probability of product 1. In order to determine an inventory level Q2 that is feasible for product 2
CE

given a specific Q1 , let H(Q1 , Q2 ) = Pr{D2 ≤ Q2 + (Q1 − D1 )+ } denote the in-stock probability of

product 2. Explicitly, by conditioning on D1 < Q1 and D1 ≥ Q1 respectively, we have

AC

Z Q1 Z Q1 +Q2 −ξ1 Z ∞ Z Q2
H(Q1 , Q2 ) = f (ξ1 , ξ2 )dξ2 dξ1 + f (ξ1 , ξ2 )dξ2 dξ1
0 0 Q1 0
Z Q1 Z Q1 +Q2 −ξ1 Z Q1 Z Q2 Z ∞ Z Q2
= f (ξ1 , ξ2 )dξ2 dξ1 − f (ξ1 , ξ2 )dξ2 dξ1 + f (ξ1 , ξ2 )dξ2 dξ1
0 0 0 0 0 0
Z Q1 Z Q1 +Q2 −ξ1 Z Q2 Z ∞
= f (ξ1 , ξ2 )dξ2 dξ1 + f (ξ1 , ξ2 )dξ1 dξ2
0 Q2 0 0
Z Q1 Z Q1 +Q2 −ξ1
= f (ξ1 , ξ2 )dξ2 dξ1 + F2 (Q2 ).
0 Q2

10
 ACCEPTED MANUSCRIPT
∂H(Q1 ,Q2 )
Also let H i (Q1 , Q2 ) denote ∂Qi , the partial derivative of H(Q1 , Q2 ), i = 1, 2. As

Z Q2 +Q1 −Q1 Z Q1 Z Q1
H 1 (Q1 , Q2 ) = f (Q1 , ξ2 )dξ2 + f (ξ1 , Q1 + Q2 − ξ1 )dξ1 = f (ξ1 , Q1 + Q2 − ξ1 )dξ1
Q2 0 0

and

Z Q1 Z Q1 Z ∞
2
H (Q1 , Q2 ) = f (ξ1 , Q1 + Q2 − ξ1 )dξ1 − f (ξ1 , Q2 )dξ1 + f (ξ1 , Q2 )dξ1
0 0 0
Z Q1 Z ∞
= f (ξ1 , Q1 + Q2 − ξ1 )dξ1 + f (ξ1 , Q2 )dξ1 ,

T
0 Q1

IP
H(Q1 , Q2 ) is increasing in both Q1 and Q2 . Assuming that Pr{Di < ∞} = 1 for i = 1, 2,

limQ2 →∞ H(Q1 , Q2 ) = 1. So for any Q1 ≥ F1−1 (α1 ), there exists a smallest Q2 such that H(Q1 , Q2 ) ≥

CR
α2 . We let q(Q1 , α2 ) denote this minimum inventory level for product 2. In order for an inventory level

Q2 to be feasible for product 2, we must have Q2 ≥ q(Q1 , α2 ). In particular, when Q1 is sufficiently

US
large, H(Q1 , 0) ≥ α2 and q(Q1 , α2 ) = 0. In general, q(Q1 , α2 ) is a decreasing function of Q1 for a fixed

0 ≤ α2 ≤ 1 and an increasing function of α2 for a fixed Q1 ≥ F1−1 (α1 ).

AN
To achieve a target service level for product 2, the inventory level of product 1 or of product 2 or of

both may be increased. As H 1 (Q1 , Q2 ) ≤ H 2 (Q1 , Q2 ), this implies that if the objective is to achieve
M

a specific service level for product 2 by increasing the inventory of only one product, then this can

be accomplished with fewer units of product 2 than product 1. This also implies that the slope of
ED

the curve q(x1 , α2 ) is greater than or equal to −1. Considering that S0 (x1 ) is increasing and the

slope of S1 (x2 ) is less than or equal to −1, q(x1 , α2 ) must intersect with S0 (x1 ) and with S1 (x2 ). In
PT

addition, once q(x1 , α2 ) separates from S0 (x1 ) or from S1 (x2 ), they do not intersect again. Let q0 be

the smallest inventory level of product 1 that satisfies the service requirement of product 2 completely
CE

by substitution, i.e., without the use of any product 2. Then for any Q1 ≥ q0 , q(Q1 , α2 ) = 0. When α2

is large, the curve q(x1 , α2 ) may have no intersection with S2 (x1 ). In the case there is an intersection
AC

at (Q1 , q(Q1 , α2 )) and Q1 ≤ q0 , then by (3.4), (v1 − p2 )α2 − (v1 − v2 )F2 (q(Q1 , α2 )) + p2 − w2 = 0.

Because q(x1 , α2 ) is decreasing in x1 , it is easy to check that once the two curves separate, they will

not intersect again.

The constraint curve q(x1 , α2 ) can intersect with S0 (x1 ) and effective parts of S1 (x2 ) (i.e., the part

where x2 ≥ Q∗2 ) and S2 (x1 ) (i.e., the part where x1 ≥ Q∗1 ) at one point or three points (if q(x1 , α2 ) has

no intersection with S2 (x1 ), we just assume that they intersect at x1 = ∞). The one point case arises

when α2 is small. As Case 2 of Figure 2a shows, in the one point case the constraint curve intersects

with S0 and stays below the controlling parts of curves S1 and S2 . At any inventory level below

11
 ACCEPTED MANUSCRIPT

the constraint curve the effective parts of the controlling curves S1 and S2 still remain in the feasible

decision domain. In this case the optimal inventory levels of the unconstrained problem will satisfy the

in-stock commitment for product 2. In the three point case we label the three intersecting points from

left to right along the constraint curve q(x1 , α2 ) as (Q11 , Q12 ), (Q01 , Q02 ) and (Q21 , Q22 ), respectively (see

Figure 2b). The next lemma shows that in the three point case there exists at least a local maximizer

along the curve between (Q11 , Q12 ) and (Q01 , Q02 ).

Lemma 4. Suppose the constraint curve q(x1 , α2 ) intersects with S1 (x2 ), S0 (x1 ) and S2 (x1 ) at

(Q11 , Q12 ), (Q01 , Q02 ) and (Q21 , Q22 ), respectively. Then there exists at least one local maximizer of

T
T (Q1 , Q2 ) along q(x1 , α2 ) between (Q11 , Q12 ) and (Q01 , Q02 ), and there is no local maximizer of T (Q1 , Q2 )

IP
along q(x1 , α2 ) between (Q01 , Q02 ) and (Q21 , Q22 ).

CR
Proof. See the Appendix.

US
When the assumption of the preceding lemma holds and the initial inventory (x1 , x2 ) = (0, 0), as

will be assumed henceforth, it will be shown below that one of the local maximizers mentioned in the
AN
lemma optimizes T (Q1 , Q2 ). The following theorem shows that the optimal solution to the constrained

problem can be found using the Karush-Kuhn-Tucker (KKT) conditions.

M

Theorem 2. The optimal inventory levels can be found amongst the following points: (i) the global

maximizer of T (Q1 , Q2 ), i.e. (Q∗1 , Q∗2 ); (ii) Q1 = F1−1 (α1 ), Q2 ≥ q(F1−1 (α1 ), α2 ), where T 1 (F1−1 (α1 ), Q2 ) <
ED

0 and T 2 (F1−1 (α1 ), Q2 ) = 0; (iii) Q1 = F1−1 (α1 ), Q2 = q(F1−1 (α1 ), α2 ), where T 1 (F1−1 (α1 ), q(F1−1 (α1 ), α2 ))) <

0, T 2 (F1−1 (α1 ), q(F1−1 (α1 ), α2 ))) < 0 and

PT

T 1 (F1−1 (α1 ), q(F1−1 (α1 ), α2 ))) H 1 (F1−1 (α1 ), q(F1−1 (α1 ), α2 )))
> , (4.7)
T 2 (F1−1 (α1 ), q(F1−1 (α1 ), α2 ))) H 2 (F1−1 (α1 ), q(F1−1 (α1 ), α2 )))
CE

and (iv) Q1 ≥ F1−1 (α1 ), Q2 = q(Q1 , α2 ), where T 1 (Q1 , q(Q1 , α2 )) < 0, T 2 (Q1 , q(Q1 , α2 )) < 0 and
AC

T 1 (Q1 , q(Q1 , α2 )) H 1 (Q1 , q(Q1 , α2 ))

= . (4.8)
T 2 (Q1 , q(Q1 , α2 )) H 2 (Q1 , q(Q1 , α2 ))

Proof. See the Appendix.

It should be noted that case (ii) represents the situation where S1 could not satisfy the service con-

straint. Cases (iii) and (iv) together are related to the local maximizers of T (Q1 , Q2 ) along the bound-

ary x2 = q(x1 , α2 ) where x1 ≥ F1−1 (α1 ). Denoting the derivative of T (x1 , q(x1 , α2 )) with respect to

12
 ACCEPTED MANUSCRIPT

x1 for x1 ≥ F1−1 (α1 ) by T̃ 1 (x1 , α2 ) we have by the chain rule that

T̃ 1 (x1 , α2 ) = T 1 (x1 , q(x1 , α2 )) + T 2 (x1 , q(x1 , α2 ))q 1 (x1 , α2 ). (4.9)

As H(x1 , q(x1 , α2 )) = α2 , taking the derivative with respect to x1 on both sides of this identity leads

to

H 1 (x1 , q(x1 , α2 )) + H 2 (x1 , q(x1 , α2 ))q 1 (x1 , α2 ) = 0,

and thus

T
H 1 (x1 , q(x1 , α2 ))
q 1 (x1 , α2 ) = − . (4.10)
H 2 (x1 , q(x1 , α2 ))

IP
Substituting this expression for q 1 (x1 , α2 ) into (4.9) establishes the first-order condition for the max-

CR
imizer of T along the boundary of x2 = q(x1 , α2 ), i.e.,

H 1 (x1 , q(x1 , α2 ))
T 1 (x1 , q(x1 , α2 )) − T 2 (x1 , q(x1 , α2 ))
US H 2 (x1 , q(x1 , α2 ))

Thus, cases (iii) and (iv) actually enumerate all of the possible maximizers along the boundary. With
= 0. (4.11)
AN
them and the values of T for them available, we are able to determine the maximizer of the objective T

that not only optimizes T along the boundary but also maximizes T in the admissible decision region.
M

x2 x2
x1  F (1 )
1
ED

x1  F11 (1 )
1

S1 S1
S0 S0
IV
IV
PT

III III
(Q10 , Q20 )
( Q11 , Q 21 ) q( x1 ,  2 ) * *
(Q , Q )
1 2
I I
*
(Q , Q )* S2
CE

1 2
S2 II
II (Q10 , Q20 ) q( x1 ,  2 )
x1 x1
Case 1. The corner point in area I Case 2. The corner point in area II
AC

Figure 2a: Optimal policy with customer service objectives where the corner point (where the two

constraint curves meet) lies in regions I (Case 1) and II (Case 2) respectively.

13
 ACCEPTED MANUSCRIPT

x2 x2

x1  F11 (1 ) x1  F11 (1 )

S1 S1
S0 S0
IV IV
III III
( Q 10 , Q 20 )
1 1 ( Q 11 , Q 21 ) ( Q 10 , Q 20 )
(Q , Q )
1 2
I (Q1* , Q2* ) I ( Q1* , Q 2* ) S2
( Q 12 , Q 22 ) S2 ( Q12 , Q 22 )
II
II q( x1 ,  2 ) q( x1 ,  2 )
x1 x1
Case 3. The corner point in area III Case 4. The corner point in area IV

T
Figure 2b: Optimal policy with customer service objectives where the corner point (where the two

IP
constraint curves meet) lies in regions III (Case 3) and IV (Case 4) respectively.

CR
The next theorem helps us visualize the optimal stocking decision. In the unconstrained problem the

feasible decision space, the first quadrant of the two-dimensional plane, is divided into four regions by

US
the controlling curves Si , i = 0, 1, 2 (see Figure 1). Curve S0 divides the quadrant into two regions and

each of these regions is then further divided into two regions either by S1 or S2 . We label these four

regions in a counterclockwise fashion starting from the lower-left (the region closest to the origin) as
AN
I, II, III and IV (see Figures 2a and 2b). In the problem with service level constraints this region is

also delimited by the line/curve x1 = F1−1 (α1 ) and the curve x2 = q(x1 , α2 ). The corner point formed
M

by where these two constraint curves meet, i.e. (F1−1 (α1 ), q(F1−1 (α1 ), α2 )), serves as a reference point

in the following.
ED

Theorem 3. The optimal inventory level is given by the upper envelope constructed from the effective

parts of curves S1 and S2 , and the constraint curves x1 = F1−1 (α1 ) and x2 = q(x1 , α2 ).
PT

Proof. See the Appendix.

CE

It is informative to examine Figures 2a and 2b in light of Theorem 3. We begin with Case 1 of Figure

2a where the corner point falls in region I. In this case the optimal stocking levels are determined
AC

by the part of curve x1 = F1−1 (α1 ) above S1 , the curve S1 from where it intersects with the curve

x1 = F1−1 (α1 ) to where it intersects with the curve q(x1 , α2 ), and the curve q(x1 , α2 ) above S1 and S2 .

Next consider Case 2 of Figure 2a where the corner point falls in region II. In this case the optimal

stocking levels are determined by the part of curve x1 = F1−1 (α1 ) above S1 , the curve S1 from where

it intersects with the curve x1 = F1−1 (α1 ) to where it intersects with the curve S0 , and the curve S2 .

Now consider Case 3 of Figure 2b where the corner point falls in region III. In this case the optimal

stocking levels are determined by the whole curve x1 = F1−1 (α1 ), the curve q(x1 , α2 ) from where it

meets the curve x1 = F1−1 (α1 ) until it intersects the curve S2 , and the curve S2 where it lies above the

14
 ACCEPTED MANUSCRIPT

curve q(x, α2 ). Lastly consider Case 4 of Figure 2b where the corner point falls in region IV. In this

case the optimal stocking levels are determined by the whole curve x1 = F1−1 (α1 ), the curve q(x1 , α2 )

from where it meets the curve x1 = F1−1 (α1 ) until it intersects the curve S2 , and the curve S2 where

it lies above the curve q(x, α2 ).

5 Numerical Study

In this section, we present the results of a numerical study in which we investigated 1) how optimal

T
inventory levels are affected by service level requirements specified as target in-stock probabilities, and

IP
2) how profitability is affected by substitution. In the following we let q1 (α1 , α2 ) and q2 (α1 , α2 ) denote

CR
the optimal inventory levels of products 1 and 2, respectively, when their respective target in-stock

probabilities are α1 and α2 .

US
Lemma 5. For fixed α2 , as α1 increases, q1 (α1 , α2 ), the optimal inventory level of product 1 is

increasing, while q2 (α1 , α2 ), the optimal inventory level of product 2 is decreasing.

AN
Proof. See the Appendix.

In order to gain insight into how service level requirements affect the optimal inventory levels, we
M

investigated four cases. For all the cases we let w1 = 6, w2 = 5, p1 = 10, p2 = 8, v1 = 4 and v2 = 3.
ED

With respect to demand, for each case a different pair of demand distributions was considered: (Case

1) both D1 and D2 uniformly distributed between 0 and 10, denoted by Di ∼ U [0, 10], i = 1, 2; (Case

2) D1 ∼ U [0, 5] and D2 ∼ U [0, 10]; (Case 3) both D1 and D2 exponentially distributed with mean
PT

1, denoted by Di ∼ Exp(1), i = 1, 2; (Case 4) D1 ∼ Exp(2) and D2 ∼ Exp(0.5). For the required

service levels we first defined the set A = {0.50, 0.60, 0.70, 0.80, 0.90}. We then generated a set of
CE

target in-stock probability (TIP) pairs from the Cartesian product of the set A with itself, denoted

by A × A = {(α1 , α2 ) : α1 ∈ A, α2 ∈ A}. This exercise gave us twenty-five TIP pairs to assess for each
AC

case. The results of our study are reported in Tables 2 and 3. Each pair of numbers in the interior of

a table gives the optimal inventory levels for products 1 and 2 for a member of A × A for one of the

four cases.

Table 2: Optimal Inventory Levels of Products 1 and 2 for Cases 1 and 2

15
 ACCEPTED MANUSCRIPT

Case 1 Case 2

α1 \α2 0.5 0.6 0.7 0.8 0.9 0.5 0.6 0.7 0.8 0.9

0.5 8.1 3.5 8.1 3.5 8.1 3.9 8.2 5.2 8.5 6.8 4.3 4.5 4.3 4.5 4.2 5.2 4.6 5.9 4.5 7.3

0.6 8.1 3.5 8.1 3.5 8.1 3.9 8.2 5.2 8.5 6.8 4.3 4.5 4.3 4.5 4.2 5.2 4.6 5.9 4.5 7.3

0.7 8.1 3.5 8.1 3.5 8.1 3.9 8.2 5.2 8.5 6.8 4.3 4.5 4.3 4.5 4.2 5.2 4.6 5.9 4.5 7.3

0.8 8.1 3.5 8.1 3.5 8.1 3.9 8.2 5.2 8.5 6.8 4.3 4.5 4.3 4.5 4.2 5.2 4.6 5.9 4.5 7.3

0.9 9.0 2.9 9.0 2.9 9.0 3.2 9.0 4.6 9.2 6.3 4.5 4.4 4.5 4.4 4.5 5.0 4.6 5.9 4.5 7.3

Table 3: Optimal Inventory Levels of Products 1 and 2 for Cases 3 and 4

T
Case 3 Case 4

IP
α1 \α2 0.5 0.6 0.7 0.8 0.9 0.5 0.6 0.7 0.8 0.9

0.5 1.6 0.4 1.6 0.4 1.7 0.4 1.7 0.8 2.0 1.3 1.0 1.3 1.0 1.3 1.1 1.7 1.3 2.3 1.6 3.4

CR
0.6 1.6 0.4 1.6 0.4 1.7 0.4 1.7 0.8 2.0 1.3 1.0 1.3 1.0 1.3 1.1 1.7 1.3 2.3 1.6 3.4

0.7 1.6 0.4 1.6 0.4 1.7 0.4 1.7 0.8 2.0 1.3 1.0 1.3 1.0 1.3 1.1 1.7 1.3 2.3 1.6 3.4

0.8

0.9
1.6

2.3
0.4

0.1
1.6

2.3
0.4

0.1
1.7

2.3
0.4

0.1
1.7

2.3
0.8

0.4 US
2.0

2.4
1.3

1.0
1.0

1.1
1.3

1.3
1.0

1.1
1.3

1.3
1.1

1.1
1.7

1.7
1.3

1.3
2.3

2.3
1.6

1.6
3.4

3.4
AN
We now discuss the results. We found that for the TIP pair α1 = α2 = 0.50, in every case the optimal

inventory level is equal to the global maximizer (Q∗1 , Q∗2 ). In fact in all four cases the global maximizer

gives the optimal inventory level for TIP pairs with α1 < 0.90 and α2 < 0.7. Hence for these cases
M

service level requirements that are low to middling do not impact inventory levels.
ED

It can also be seen in all cases that for fixed α2 , as α1 increases, the optimal inventory level of product

1 is increasing while the optimal inventory level of product 2 is decreasing, which is consistent with
PT

Lemma 5 (e.g., in Case 3, as α1 → 1, q2 (α1 , α2 ) → 0 and hence all demand for product 2 will be

met through substitution). On the other hand, for a fixed α1 , as α2 increases we observe a more
CE

complex dynamic. In every case we found both inventory levels are higher when α2 = 0.9 compared

to when α2 = 0.5 (the higher inventory of product 1 reflects that it can substitute for product 2),
AC

but the optimal inventory level of product 1 does not always change in a monotone fashion as α2

increases from 0.5 to 0.9. For example in Case 2 it can be observed that as α2 increases, there are

instances where the inventory level of product 1 decreases (a reverse substitution effect reflecting that

some product 1 formerly stocked to potentially meet product 2 demand is no longer needed because

of increased product 2 holdings).

From our review of the results, first with α2 fixed and then with α1 , we can conclude that for a

fixed α2 , the optimal inventory level of product 1 is relatively insensitive to changes in the value of

α1 , when α1 is not larger than 0.9. For example, in Case 1 with α2 = 0.5, it can be seen that the

16
 ACCEPTED MANUSCRIPT

optimal inventory level of product 1 increases from 8.1 to 9.0 as α1 increases from 0.5 to 0.9, which

is a percentage gain in inventory of 11.1%. We found the average percentage change in the optimal

inventory level of product 1 to be 13.3% with the extremes being a 43.8% increase and no change in

inventory at all. By contrast the optimal inventory level of product 2 is quite sensitive to changes

in the value of α2 for a fixed α1 . We found the average percentage change in the optimal inventory

level of product 2 to be 170.8% with the minimum being an increase of 62.2% and the maximum an

increase of 900%. Of course it is important to note that the large percentage increases in the optimal

inventory level of product 2 are attributable at least in part to starting from a low base.

T
Lastly, we observe that when both products face the same demand distribution (see Cases 1 and 3)

IP
and both products have the same required service levels, it can be seen that the inventory level of

CR
product 1 is always much higher than the inventory level of product 2, reflecting that the additional

product 1 can be used to meet unexpectedly high demand for product 1 or used to meet demand for

US
product 2 when demand for product 1 is unexpectedly low.

In the remainder of this section we study the financial benefits of substitution by comparing the optimal
AN
inventory policies with and without substitution. When there is no substitution, the two products are

independent of each other, so we can calculate the optimal inventory level for each product separately.
M

Making use of the newsvendor formula, it is easy to determine that the optimal inventory level for

product i after taking into account the required service level is

ED

p i − wi
Qi = Fi−1 ( ) ∨ Fi−1 (αi ), i = 1, 2.
p i − vi
PT

To understand the impact of substitution on the bottom line, we compute the percentage increase in

expected profit between the optimal policies with and without substitution. The results for Cases 1-4
CE

are shown in Figure 4.

AC

17
 ACCEPTED MANUSCRIPT

%
%
0.25 0.2

0.2 Case 1 Case 2

0.15

0.15 α1=0.5
α1=0.6 α1=0.5
0.1
α1=0.7 α1=0.6
0.1 α1=0.7
α1=0.8
α1=0.9 α1=0.8
0.05 α1=0.9
0.05

0 0
α2 α2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

%
%

1 1.4

T
1.2
0.8 Case 3 Case 4
1

IP
0.6
α1=0.5 0.8 α1=0.5
α1=0.6
α1=0.6
α1=0.7 0.6
0.4 α1=0.7
α1=0.8

CR
α1=0.8
α1=0.9 0.4 α1=0.9
0.2
0.2

0 α2 0 α2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

US
Figure 4: Percentage increase in profit from substitution.
AN
In Case 1, the percentage increase in expected profit that comes from the use of substitution is never

less than 10%. It can be seen that for a fixed α1 and α2 ≤ 0.6, the percentage increase remains at a

constant level. This reflects that for all these required service levels, the global maximizer is feasible
M

and optimal. Above α2 = 0.6, higher inventory levels become necessary, and as noted earlier in this

section, this is sometimes accomplished by stocking more of product 1 which is then substituted for
ED

product 2 as appropriate, hence, the higher percentage increase in the expected profit from substitution

as α2 rises above 0.6. Even more noteworthy is the percentage increase in expected profit observed
PT

as α1 increases from middling values (α1 ≤ 0.7) towards 1 for a fixed α2 (e.g. in Case 1 for α2 = 0.7,

as α1 increases from 0.7 to 0.9, the percentage increase goes from 12.37% to 18.14%). Hence, we find
CE

that the financial gain from substitution is greatest when the required service level for product 1 is

high. Similar remarks apply to the other cases.

AC

6 Conclusions and Suggestions for Further Research

In this paper, we considered a single period problem in which a firm seeks to determine the profit

maximizing inventory levels for two products in advance of observing random demand. In making its

stocking decisions, the firm wants to take into account downward, supplier-orientated substitution that

it can employ and the firm needs to take into account customer service commitments that it has made.

For the firm with this problem setting, we identified the optimal inventory policy with and without

18
 ACCEPTED MANUSCRIPT

service commitments. In order to develop a keener understanding of how optimal inventory levels are

affected by service commitments and how firm profitability is affected by substitution, we undertook

a numerical study. By conducting this study, we learned that middling service level objectives have

no impact on the optimal inventory levels. On the other hand, we found that when the service

level objective for product 1 (the superior product) is fixed at some level, then as the service level

objective for product 2 increases, the optimal inventory level of product 1 may increase, remain the

same, or interestingly even decrease. Through the study we also learned that while the availability of

substitution as an option always increases a firm’s profitability, the impact can be especially dramatic

T
when the firm faces high service level requirements.

IP
Further research is needed to generalize the two-product results of this paper to the multiple product

CR
setting. In addition, research that incorporates price-setting decisions would be valuable and an

interesting topic for future study.

Acknowledgements
US
This work was funded by a Marie Curie International Incoming Fellowship within the 7th Framework
AN
Programme of the European Commission (PIIF-GA-2009-253720), National Natural Science Founda-

tion of China (No.71272128, 71432003), Program for New Century Excellent Talents in University
M

(No.NCET-12-0087), and Youth Foundation for Humanities and Social Sciences of Ministry of Educa-

tion of China (No.11YJC630022). The authors are grateful to the editor and the two referees, whose
ED

comments led to a much improved version of this article.

PT

References

[1] Alptekinoglu, A., Banerjee, A., Paul, A. and Jain, N. 2013. Inventory pooling to deliver differen-
CE

tiated service. Manufacturing and Service Operations Management, 15, 33-44.

AC

[2] Axsäter, S. 2003a. A new decision rule for lateral transshipments in inventory systems. Manage-

ment Science, 49, 1168-1179.

[3] Axsäter, S. 2003b. Evaluation of unidirectional lateral transshipments and substitutions in inven-

tory systems. European Journal of Operational Research, 149(2), 438-447.

[4] Bardhan, A., Dawande, M., Gavirneni, S., Mu, Y. and Sethi, S. 2013. Forecast and rolling horizons

under demand substitution and production changeovers: analysis and insights. IIE Transactions,

45, 323-340.

19
 ACCEPTED MANUSCRIPT

[5] Bassok, Y., Anupindi, R. and Akella, R. 1999. Single-period multiproduct inventory models with

substitution. Operations Research, 47, 635-642.

[6] Bensoussan, A., Feng, Q. and Sethi, S. 2011. Achieving a long-term service target with periodic

demand signals: a newsvendor signals. Manufacturing and Service Operations Management, 13,

73-88.

[7] Cachon, G. and Terwiesch, C. Matching Supply with Demand: An Introduction to Operations

Management. McGraw-Hill/Irwin, New York, 2012.

T
[8] Chen, X. and Shen, Z.-J. 2012. An analysis of a supply chain with options contracts and service

IP
requirements. IIE Transactions, 44, 805-819.

CR
[9] Deflem, Y. and Van Nieuwenhuyse, I. 2013. Managing inventories with one-way substitution: a

newsvendor analysis. European Journal of Operational Research 228, 484-493.

US
[10] Hajek, B. 1985. Extremal splitting of point processes. Mathematics of Operations Research 10,

543-556.
AN
[11] Hsu, V.N., Li, C.L. and Xiao, W.-Q. 2005. Dynamic lot size problems with one-way product

substitution. IIE Transactions, 37, 201-215.

M

[12] Karaesmen, I. and van Ryzin, G. 2004. Overbooking with substitutable inventory classes. Oper-

ations Research 52, 83-104.

ED

[13] Kay, E. 2005. Ways to measure supplier performance. Purchasing.com.

PT

http://www.purchasing.com/article/CA508544.html.

[14] Li, Q., and Yu, P. 2014. Multimodularity and its applications in three stochastic dynamic inven-
CE

tory problems. Manufacturing and Service Operations Management, 16, 455-463.

[15] McGillivray, A. R. and Silver, E. A.. 1978. Some concepts for inventory control under substitutable
AC

demand. INFOR, 16, 47-63.

[16] Murota, K. 2005. Note on multimodularity and L-convexity. Mathematics of Operations Research

30, 658-661.

[17] Pasternack, B. A. and Drezner, Z. S. 1991. Optimal inventory policies for substitutable commodi-

ties with stochastic demand. Naval Research Logistics, 38, 221-240.

[18] Pentico, D. 2008. The assortment problem: a survey. European Journal of Operational Research

190, 295-309.

20
 ACCEPTED MANUSCRIPT

[19] Rao, U. S., Swaminathan, J. M. and Zhang, J. 2004. Multi-product inventory planning with

downward subsitution, stochastic demand and setup costs. IIE Transactions, 36, 59-71.

[20] Shumsky, R. A. and Zhang, F. 2009. Dynamic capacity management with substitution. Operations

Research 57, 671-684.

[21] Van Mieghem J. A. 1998. Investment strategies for flexible resources. Management Science, 44,

1071-1078.

[22] Van Mieghem J. A. and Rudi, N. 2002. Newsvendor networks: inventory management and capac-

T
ity investment with discretionary activities. Manufacturing & Service Operations Management,

IP
4, 313-335.

CR
[23] Wang, Y., Xiao, Y., and Yang, N. 2014. Improving reliability of a shared supplier with competition

and spillovers. European Journal of Operational Research 236, 499-510.

Appendix
US
AN

Proof of Lemma 1
M

First, we observe that the objective function T (Q1 , Q2 ), which, given

ED

T (Q1 , Q2 ) = E{−p1 (Q1 − D1 )+ + p2 [(D2 − Q2 ) ∧ (Q1 − D1 )+ ] + v1 [(Q1 − D1 )+

−(D2 − Q2 )+ ]+ + v2 (Q2 − D2 )+ } + (p1 − w1 )Q1 + (p2 − w2 )Q2 ,

PT

can be expressed as the expectation of

CE

−p1 (Q1 − d1 )+ + p2 [(d2 − Q2 ) ∧ (Q1 − d1 )+ ] + v1 [(Q1 − d1 )+ − (d2 − Q2 )+ ]+

AC

+ v2 (Q2 − d2 )+ + (p1 − w1 )Q1 + (p2 − w2 )Q2 ,

where d1 and d2 can be thought of as any realized demand values. As

(d2 − Q2 ) ∧ (Q1 − d1 )+ = −(Q2 − d2 )+ + (d2 − Q2 )+ ∧ (Q1 − d1 )+

= −(Q2 − d2 )+ + (Q1 − d1 )+ − [(Q1 − d1 )+ − (d2 − Q2 )+ ]+ ,

21
 ACCEPTED MANUSCRIPT

the foregoing identity can be transformed into

−p1 (Q1 − d1 )+ − p2 (Q2 − d2 )+ + p2 (Q1 − d1 )+ − (p2 − v1 )[(Q1 − d1 )+ − (d2 − Q2 )+ ]+

+ v2 (Q2 − d2 )+ + (p1 − w1 )Q1 + (p2 − w2 )Q2

= −(p1 − p2 )(Q1 − d1 )+ − (p2 − v2 )(Q2 − d2 )+ − (p2 − v1 )[(Q1 − d1 )+ − (d2 − Q2 )+ ]+

+ (p1 − w1 )Q1 + (p2 − w2 )Q2 ,

∂[(Q1 −d1 )+ −(d2 −Q2 )+ ]+

which is jointly concave in (Q1 , Q2 ) for any fixed pair d1 and d2 . In addition, as ∂Q1

T
is increasing in Q2 , the right-hand side of the last identity is a submodular function of (Q1 , Q2 ). Since

IP
concavity and submodularity are preserved under expectation, the proof is complete.

CR
Proof of Lemma 3

00 0 0 00 00
US
(1) It is sufficient to prove the first assertion for S1 (x2 ) since the proof for S2 is similar. To this end, let
0 0
0 ≤ x2 ≤ x2 , x1 = S1 (x2 ) and x1 = S1 (x2 ). As T 1 (x1 , x2 ) = 0 and for 0 ≤ x2 < x2 , T 1 (x1 , x2 ) > 0 and
0 0 0
AN
0 0 0 00 00
for x2 > x2 , T 1 (x1 , x2 ) ≤ 0, we have T 1 (x1 , x2 ) ≤ 0. Furthermore, because T 1 (x1 , x2 ) is decreasing
00 00 00 0 0 0
in x1 , and by definition, T 1 (x1 , x2 ) = 0, x1 ≤ x1 . For the last assertion, we start with (x1 , x2 ) and
00 00 0 0 00 00 0 00 0 0 0 0
M

(x1 , x2 ) such that S0 (x1 ) = x2 and S0 (x1 ) = x2 for x1 ≤ x1 . Because T 1 (x1 , x2 ) − T 2 (x1 , x2 ) = 0

and since it follows from Lemma 2 that T 1 (x1 , x2 ) − T 2 (x1 , x2 ) is decreasing in x1 and increasing x2 ,
ED

00 0 00 0 00 00 00 00
T 1 (x1 , x2 ) − T 2 (x1 , x2 ) ≤ 0. Making use of Lemma 2 again, as T 1 (x1 , x2 ) − T 2 (x1 , x2 ) = 0, it turns
00 0
out that x2 ≥ x2 .

(2) It is sufficient to prove this assertion for S1 (x2 ) since the proof for S2 is similar. Because T 1 (x1 , x2 )−
PT

T 2 (x1 , x2 ) is decreasing in x1 and increasing x2 , it follows that for any ε > 0, T (x1 , x2 )−T (x1 −ε, x2 +ε)
0 00
is decreasing in x1 , in other words T 1 (x1 , x2 ) − T 1 (x1 − ε, x2 + ε) ≤ 0. Hence, for x1 ≤ x1 ,
CE

00 00 0 0 0 0 00 00
0 = T 1 (S1 (x2 ), x2 ) = T 1 (S1 (x2 ), x2 ) ≤ T 1 (S1 (x2 ) + x2 − x2 , x2 ).
AC

0 0 00
This implies that S1 (x2 ) + x2 − x2 is on the left side of the curve S1 (x2 ), and consequently, the slope

of S1 (x2 ) is less than or equal to −1.

Proof of Theorem 1

For xi ≤ Q∗i , i = 1, 2, since by Lemma 3 Si (x3−i ) is decreasing, Si (x3−i ) ≥ Si (Q∗3−i ) = Q∗i and

thus x̄i = Q∗i . In this case it is optimal to order up to (Q∗1 , Q∗2 ) = (x1 ∨ x̄1 , x2 ∨ x̄2 ) and thus

22
 ACCEPTED MANUSCRIPT

(Q∗1 , Q∗2 ) is the optimal joint base stock level. Now consider xi > Q∗i , i = 1, 2. By Lemma 3 again

Si (x3−i ) ≤ Si (Q∗3−i ) = Q∗i and so x̄i = Si (x3−i ) ≤ Q∗i ≤ xi . Thus it is optimal not to order product

i for i = 1, 2, implying that the optimal order-up-to level is (x1 , x2 ) = (x1 ∨ x̄1 , x2 ∨ x̄2 ). Finally, in

view of symmetry, it is sufficient to consider x1 ≤ Q∗1 and x2 ≥ Q∗2 . In this case x̄1 = S1 (x2 ) and

x̄2 = Q∗2 . Based on the above discussion it is optimal to order up to x1 ∨ S1 (x2 ) for product 1 and

nothing for product 2. As x2 ∨ Q∗2 = x2 , it also shows that in this case (x̄1 , x̄2 ) = (S1 (x2 ), Q∗2 ) is the

optimal base-stock order level, which completes the proof.

T
Proof of Lemma 4

IP
Note that on the constraint curve, H(Q1 , q(Q1 , α2 )) = α2 . Thus, by the chain rule, we have

CR
H 1 (Q1 , q(Q1 , α2 )) + H 2 (Q1 , q(Q1 , α2 ))q 1 (Q1 , α2 ) = 0,

where q 1 (Q1 , α2 ) stands for ∂q(Q1 ,α2 )

∂Q1 . US
Because for Q1 > 0,
AN
H 2 (Q1 , q(Q1 , α2 )) > H 1 (Q1 , q(Q1 , α2 )) > 0,
M

whence we may divide by H 2 (Q1 , q(Q1 , α2 )) thus obtaining

H 1 (Q1 , q(Q1 , α2 ))
ED

q 1 (Q1 , α2 ) = − .
H 2 (Q1 , q(Q1 , α2 ))

Using the preceding expression for q 1 (Q1 , α2 ), we can write the derivative of T̃ (Q1 , α2 ) = T (Q1 , q(Q1 , α2 )),
PT

with respect to Q1 as:

CE

T̃ 1 (Q1 , α2 ) = T 1 (Q1 , q(Q1 , α2 )) + T 2 (Q1 , q(Q1 , α2 ))q 1 (Q1 , α2 )

H 1 (Q1 , q(Q1 , α2 ))
= T 1 (Q1 , q(Q1 , α2 )) − T 2 (Q1 , q(Q1 , α2 )) .
H 2 (Q1 , q(Q1 , α2 ))
AC

Considering (Q11 , Q12 ) = (Q11 , q(Q11 , α2 )), we use the fact that T 1 (Q11 , q(Q11 , α2 )) = 0 and T 2 (Q11 , q(Q11 , α2 )) <

0, to conclude that T 1 (Q11 , Q12 ) = T̃ 1 (Q11 , α2 ) > 0. Considering (Q01 , Q02 ) = (Q01 , q(Q01 , α2 )), we use the

fact that T 1 (Q01 , q(Q01 , α2 )) = T 2 (Q01 , q(Q01 , α2 )) < 0, to conclude that

 
H 1 (Q0 , q(Q01 , α2 ))
T̃ 1 (Q01 , α2 ) = T 1 (Q01 , q(Q01 , α2 )) 1 − 2 01 < 0.
H (Q1 , q(Q01 , α2 ))

23
 ACCEPTED MANUSCRIPT

Finally, consider Q01 ≤ x ≤ Q21 . Since T 1 (x, q(x, α2 )) ≤ T 2 (x, q(x, α2 )) < 0,

 
1 2 H 1 (x1, q(x, α2 ))
T̃ (x, α2 ) ≤ T (x, q(x, α2 )) 1 − 2 < 0,
H (x, q(x, α2 ))

which implies that there is no local maximizer for the curve between (Q01 , Q02 ) and (Q21 , Q22 ). This

completes the proof.

Proof of Theorem 2

T
By KKT the optimal solution to our problem needs to satisfy the following conditions:

IP
CR
0 = T 1 (Q1 , Q2 ) + µ1 f1 (Q1 ) + µ2 H 1 (Q1 , Q2 ),

0 = T 2 (Q1 , Q2 ) + µ2 H 2 (Q1 , Q2 ), (6.12)

US
0 = µ1 [F1 (Q1 ) − α1 ] = µ2 [H(Q1 , Q2 ) − α2 ] ,

where µi ≥ 0 is the KKT multiplier corresponding to constraint i, i = 1, 2. Obviously µ1 = 0 if

AN
F1 (Q1 ) > α1 , and µ2 = 0 if H(Q1 , Q2 ) > α2 . We examine each pair (Q1 , Q2 ) with respect to the

foregoing conditions. It should be clear that pairs with positive partial derivative are not optimal
M

for the problem, as they make the right-hand side of at least one of the first two equations in (6.12)

positive. On the other hand, the global maximizer is a possible candidate for the optimal solution.
ED

Whether this pair satisfies the service constraints is unknown, but it is attractive to check this solution

first. By simply letting µ1 = µ2 = 0, we find that it meets the KKT conditions.

PT

Continuing with our examination of pairs, we consider (Q1 , Q2 ) with T 2 (Q1 , Q2 ) = 0 and T 1 (Q1 , Q2 ) <

0. As H 2 (Q1 , Q2 ) > 0, the µ2 associated with this pair must be zero and thus the first equation of (6.12)
CE

1
is T 1 (Q1 , Q2 ) + µ1 f1 (Q1 ) = 0. For this pair, as T 1 (Q1 , Q2 ) < 0 and f1 (Q1 ) > 0, µ1 = − T f(Q 1 ,Q2 )
1 (Q1 )
>0

indicating that F1 (Q1 ) = α1 , i.e., this pair is on the boundary and satisfies Q1 = F1−1 (α1 ). It follows
AC

that in this case, at the corner point (F1−1 (α1 ), q(F1−1 (α1 ), α2 )), T 2 (F1−1 (α1 ), q(F1−1 (α1 ), α2 )) ≥ 0,

since otherwise the concavity of T implies that T 2 (Q1 , Q2 ) < 0 for all Q1 = F1−1 (α1 ) and Q2 >

q(F1−1 (α1 ), α2 ), a contradiction. The remaining quantities (Q1 , Q2 ) must satisfy T 2 (Q1 , Q2 ) < 0 and

T 1 (Q1 , Q2 ) ≤ 0. It follows that µ2 > 0 and (Q1 , Q2 ) is on the boundary satisfying H(Q1 , Q2 ) = α2 or
2
T (Q1 ,q(Q1 ,α2 ))
Q2 = q(Q1 , α2 ). In particular, µ2 = − H 2 (Q ,q(Q ,α )) . For a pair to satisfy the first equation of (6.12),
1 1 2

it is necessary that T 1 (Q1 , Q2 ) + µ2 H 1 (Q1 , Q2 ) ≤ 0, or

T 1 (Q1 , q(Q1 , α2 ))) H 1 (Q1 , q(Q1 , α2 )))

≥ .
T 2 (Q1 , q(Q1 , α2 ))) H 2 (Q1 , q(Q1 , α2 )))

24
 ACCEPTED MANUSCRIPT

If the above relation is an equality, then we choose µ1 = 0 to meet the first equation of (6.12);

otherwise, choose µ1 > 0 and Q1 = F1−1 (α1 ). This completes the proof of the theorem.

Proof of Theorem 3

We begin by showing that the theorem holds when the corner point falls in the region labeled I (see

Figure 2a). There are two cases to consider. In the first case the curve x2 = q(x1 , α2 ) intersects

with curve S0 at or below (Q∗1 , Q∗2 ). Two points merit special attention: the first is the intersection

T
of the line x1 = F1−1 (α1 ) with S1 at (F1−1 (α1 ), Q02 ), where S1 (Q02 ) = F1−1 (α1 ), and the second is the

IP
intersection of the curve q(x1 , α2 ) with S0 at (Q01 , q(Q01 , α2 )) where S0 (Q01 ) = q(Q01 , α2 ). When the

initial stock (x1 , x2 ) lies within the constraint feasible region and above the controlling curves, it is

CR
optimal to do nothing just as in the unconstrained problem. For an initial stock (x1 , x2 ) outside this

space, if x1 < F1−1 (α1 ), x2 ≥ Q02 , and if there is no service level commitment, then it will be optimal

US
to raise the stock of product 1 to (S1 (x2 ), x2 ). But, as S1 (x2 ) < F1−1 (α1 ), and it is necessary to satisfy

the service level commitment, the stock must be raised to (F1−1 (α1 ), x2 ) by adding F1−1 (α1 ) − S1 (x2 )
AN
units of product 1 to the stock that would be established if there was no service level commitment.

Finally, for an initial stock (x1 , x2 ), where x2 < Q02 , the decision rule of the unconstrained problem
M

remains feasible and optimal.

In the second case, the curve x2 = q(x1 , α2 ) intersects with curve S0 above (Q∗1 , Q∗2 ). Four points
ED

are to be noted: the first is the intersection of the line x1 = F1−1 (α1 ) with S1 at (F1−1 (α1 ), Q02 ),

where S1 (Q02 ) = F1−1 (α1 ); the second, third and fourth points involve the intersection of the curve
PT

x2 = q(x1 , α2 ) with S1 at (Q11 , Q12 ) = (Q11 , q(Q11 , α2 )), where Q11 = S1 (q(Q11 , α2 )), with S0 at (Q01 , Q02 ) =

(Q01 , q(Q01 , α2 )), where S0 (Q01 ) = q(Q01 , α2 ), and with S2 at (Q21 , Q22 ), where Q22 = q(Q21 , α2 ) = S2 (Q21 ).
CE

Once again, when the initial stock (x1 , x2 ) lies within the constraint feasible region and above the

controlling curves, it is optimal to do nothing just as in the unconstrained problem. For an initial
AC

stock (x1 , x2 ) outside this space, (1) if x1 < F1−1 (α1 ) and x2 ≥ Q02 , it follows from similar reasoning

to that employed above that it is optimal to raise the stock of product 1 to (F1−1 (α1 ), x2 ); (2) if

Q12 ≤ x2 < Q02 , it is optimal to raise the stock of product 1 to (S1 (x2 ), x2 ); (3) if Q02 ≤ x2 < Q12 , it

is optimal to increase the stock of product 1 or both products to some level in the lower boundary

Q2 = q(Q1 , α2 ), say (Q̄∗1 (x1 , x2 ), Q̄∗2 (x1 , x2 )) with Q̄∗2 (x1 , x2 ) = q(Q̄∗1 (x1 , x2 ), α2 ), which is a local

maximizer of T (Q1 , q(Q1 , α2 )) for (x1 , x2 ) ≤ (Q1 , q(Q1 , α2 )); (4) if Q02 > x2 ≥ Q22 , when x1 ≤ Q01 ,

it is optimal to raise, either product 1 or 2 or both products, to some level in the lower boundary

Q2 = q(Q1 , α2 ), say (Q̄∗1 (x1 , x2 ), Q̄∗2 (x1 , x2 )) with Q̄∗2 (x1 , x2 ) = q(Q̄∗1 (x1 , x2 ), α2 ), which is a local

25
 ACCEPTED MANUSCRIPT

maximizer of T (Q1 , q(Q1 , α2 )) for (x1 , x2 ) ≤ (Q1 , q(Q1 , α2 )), and when x1 > Q01 , it is optimal to raise

the stock of product 2 to (x1 , S2 (x1 )); (5) finally, if x2 < Q22 , when x1 ≤ Q21 , it is optimal to stock

to some level in the lower boundary Q2 = q(Q1 , α2 ), which is a local maximizer of T (Q1 , q(Q1 , α2 ))

for x1 ≤ Q01 , and is (x1 , q(x1 , α2 )) for Q21 ≥ x1 > Q01 , and when x1 > Q21 , it is optimal to raise

the stock of product 2 to (x1 , S2 (x1 )). Note that those local maximizers of T along the boundary

Q2 = q(Q1 , α2 ) only appear in between (Q11 , q(Q11 , α2 )) and (Q01 , S0 (Q01 )), and there may exist multiple

local maximizers in this part of the curve. In our selection of optimal stocking levels, we always choose

the best among the feasible local maximizers for an initial stock. The proofs for regions II, III, and

T
IV are similar and therefore omitted for the sake of brevity.

IP
CR
Proof of Lemma 5

For a fixed α2 , as α1 increases, q1 (α1 , α2 ) either stays the same (if (q1 (α1 , α2 ), q2 (α1 , α2 )) remains

US
feasible), or increases to satisfy the higher required service level. From Theorem 3, we know that the

optimal stocking level is either (Q∗1 , Q∗2 ) if this global maximizer is a feasible solution, or is on one of
AN
the following curves: x1 = F1−1 (α1 ), S1 (x2 ), S2 (x1 ) or q(x1 , α2 ). Note that the last three curves are

decreasing in x1 , hence when α1 increases, q2 (α1 , α2 ) which is the optimal stocking level of product 2
M

must be decreasing.
ED
PT
CE
AC

26