Corsten2018 Article AssortmentPlanningForMultipleC

OR Spectrum (2018) 40:875–912
https://doi.org/10.1007/s00291-017-0496-9
REGULAR ARTICLE
Assortment planning for multiple chain stores
Hans Corsten2 · Michael Hopf1 · Benedikt Kasper2 ·

Clemens Thielen1
Received: 30 December 2016 / Accepted: 16 November 2017 / Published online: 21 November 2017
© Springer-Verlag GmbH Germany, part of Springer Nature 2017
Abstract In retail, assortment planning refers to selecting a subset of products to offer

that maximizes profit. Assortments can be planned for a single store or a retailer with
multiple chain stores where demand varies between stores. In this paper, we assume
that a retailer with a multitude of stores wants to specify her offered assortment. To
suit all local preferences, regionalization and store-level assortment optimization are
widely used in practice and lead to competitive advantages. When selecting region-
alized assortments, a trade-off between expensive, customized assortments in every
store and inexpensive, identical assortments in all stores that neglect demand variation
is preferable. We formulate a stylized model for the regionalized assortment plan-
ning problem (APP) with capacity constraints and given demand. In our approach, a
common assortment that is supplemented by regionalized products is selected. While
products in the common assortment are offered in all stores, products in the local
assortments are customized and vary from store to store. The model is both applica-
ble for optimizing a total assortment as well as a particular category of an assortment.
Concerning the computational complexity, we show that the APP is strongly NP-hard.
B Michael Hopf
hopf@mathematik.uni-kl.de
Hans Corsten
corsten@wiwi.uni-kl.de
Benedikt Kasper
benedikt.kasper@wiwi.uni-kl.de
Clemens Thielen
thielen@mathematik.uni-kl.de
1 Department of Mathematics, University of Kaiserslautern, Paul-Ehrlich-Str. 14, 67663

Kaiserslautern, Germany
2 Department of Business Studies and Economics, University of Kaiserslautern,
Gottlieb-Daimler-Str. 42, 67663 Kaiserslautern, Germany
123
876 H. Corsten et al.
We formulate the APP as an integer program and provide algorithms and methods
for obtaining approximate solutions and solving large-scale instances. Moreover, we
extend our model to include substitution effects. Lastly, we perform computational
experiments to analyze the benefits of regionalized assortment planning depending on
the variation in customer demands between stores.
Keywords Retail management · Computational complexity · Approximation

algorithms · Substitution
1 Introduction
In retail, assortment planning refers to selecting a subset of products to offer in each

store such that profit is maximized. Assortments can be planned for a single store
or for a retailer with multiple chain stores. Usually, central planning approaches are
described in the literature: either a central supervisor chooses an assortment that is
offered in all stores or, for each store, there is a local store manager who decides
upon the assortment in that particular store. For multiple stores, this can lead to poor
solutions. Consider a retailer with a multitude of stores where demand varies between
stores. She has two basic strategies for specifying her offered assortment:
– Every store has a customized assortment. Here, all demand differences can be
considered, but customized assortments are expensive to maintain.
– Assortments in all stores are the same. Here, the assortments may not be optimized
to suit all local preferences, but with a single assortment, economies of scale and
a recognition value can be generated.
So far, these two basic strategies have mainly been considered separately in the
assortment planning literature. For example, Kök et al. (2009) stated in their survey
that “while a retailer might have a different assortment at each store, the academic
literature has focused on determining a single assortment for a retailer, which could
be viewed either as a common assortment to be carried at all stores or the solution to
the assortment planning problem for a single store.”
In this paper, we analyze the benefit of mixed strategies, i.e., the selection of a
common assortment that is supplemented by regionalized products. Thus, products in
the common assortment are offered in all stores, while products in the local assortments
are customized and vary from store to store. Our model is both applicable for a retailer
who wants to optimize her total assortment as well as for a retailer who wants to
optimize the product selection for a predefined category.
1.1 Previous work
The assortment planning problem (APP) is considered in both operations research

and retail literature in various settings. For extensive literature reviews, see Hübner
and Kuhn (2012), Kök et al. (2009), Mantrala et al. (2009), Shin et al. (2015).
123
Assortment planning for multiple chain stores 877
1.1.1 Demand estimation
Modeling approaches for assortment planning are often based on demand forecasts.
Usually, choice models are used to estimate demand based on actual customer behav-
ior. A review of different choice modeling approaches can be found in Kök et al.
(2009). Common approaches are multinomial logit (MNL) models (Grewal et al.
1999; Rusmevichientong et al. 2014), nested logit models (Li and Rusmevichientong
2014), locational choice models (Gaur and Honhon 2006), exogenous demand models
(Hübner et al. 2016; Kök and Fisher 2007), and generalizations thereof (Farias et al.
2013; Vulcano et al. 2012). Other approaches for estimating demand are described,
for example, in Borin and Farris (1995) and Borin et al. (1994), where four factors
that influence demand are identified, in Urban (1998), where demand depends on the
displayed inventory, and in Fisher and Vaidyanathan (2014) and Rooderkerk et al.
(2013), where attribute-based approaches are proposed. For example, these models
are applied in Chong et al. (2001), Kök (2003), and Kök and Fisher (2007), who
provide methods to calculate demand from scanner data using maximum likelihood
estimation. Vulcano et al. (2012) used the expectation-maximization method to esti-
mate demand in a simple, yet efficient way. With these approaches, only observable
data like sales volumes and shelf availability are necessary to forecast demand for any
product.
1.1.2 Single store assortment planning
Researchers usually only consider one of the basic strategies for assortment planning
described above. Frequently, shelf space (Chen and Lin 2007), inventory (Gaur and
Honhon 2006; Honhon et al. 2010), or pricing decisions Wang (2012) are taken into
account. Talluri and Ryzin (2004) showed that a greedy algorithm is optimal for an
unconstrained assortment planning problem under an MNL model. Rusmevichientong
et al. (2014) gave an example where a greedy heuristic is suboptimal for a capaci-
tated problem and develop a polynomial-time algorithm. Rajaram (2001) proposed an
assortment selection heuristic for a problem with given demand. Chen and Lin (2007)
formulated a sequential assortment and shelf space allocation procedure when gross
margins are given. Rooderkerk and Heerde (2016) proposed a robust knapsack model
for the assortment planning problem with given profit.
1.1.3 Regionalization and store-level assortment optimization
Regionalization and store-level assortment optimization lead to competitive advan-

tages and are widely used in practice (see, for example, Fisher and Vaidyanathan
2014; Grewal et al. 1999; Hwang et al. 2010). Here, chain store management dictates
a portion of the assortment that is carried in all stores, while the remainder is chosen
to satisfy local customer preferences. Surprisingly, very little research is done in this
context.
Fisher and Vaidyanathan (2014) considered a similar problem: They characterize
an algorithm for assortment planning that allows a maximum number L of individual
assortments (where L is chosen between 1 and the number of stores I ). To achieve
123
this, they construct L store clusters and identify optimal assortments. While this leads
to a reduced number of assortments, these assortments are independent of each other.
With independent assortments, economies of scale cannot be generated. Syring (2004)
allocates products to assortment modules. Then, these modules are assigned to stores.
Hereby, a trade-off between standardization and individualization is possible. But,
as in Fisher and Vaidyanathan, a recognition value and economies of scale cannot
be generated. Rooderkerk et al. (2013) developed a model where only a fraction of
products can change between two periods. If they use a chain-wide assortment as a
starting point and apply their model once at every individual store, custom assortments
that differ in a fixed fraction from the original assortment can be generated.
1.1.4 Substitution
An important issue in assortment planning is substitution. If a product is not available,

customers may substitute it by an offered one (or choose not to buy any other prod-
uct). Empirical results show that substitution is the predominant strategy in case of a
stockout (see Aastrup and Kotzab 2010 for a review on stockout-based substitution,
and also Emmelhainz et al. 1991; Sloot et al. 2005; Walter and Grabner 1975; Zinn
and Liu 2001). In this context, Campo et al. (2004) showed that customer reaction to a
stockout is similar to that of a permanent assortment reduction. Thus, it seems benefi-
cial to model substitution effects when considering the assortment planning problem.
For a closer examination, see, for example, Kök et al. (2009), Li and Rusmevichien-
tong (2014). Usually, only one round of substitution is assumed, which means that,
if a customer cannot find her first choice as well as her substitute choice, the sale is
lost. Kök (2003) and Smith and Agrawal (2000) showed that this assumption is not
too restrictive. In this context, for example, Hübner et al. (2016) and Kök and Fisher
(2007) modeled the assortment planning problem for a single store under the exoge-
nous demand model. Additionally, it is often assumed that substitution only takes
place between products of the same category. Sloot et al. (2005) examined switching
categories and found this behavior negligible, while Huh et al. (2016) found that cus-
tomers are likely to stay within a category even though they perceive the substitution
product inferior compared to another one outside of that category.
1.2 Our contribution
We propose an alternative solution method for the APP that reflects industry practice.
Items for the common assortment and items for the local assortments are selected
simultaneously in order to maximize the total profit. We show that this problem is
strongly NP-hard and present a heuristic that is able to tackle large-scale instances
that cannot be solved within a reasonable amount of time by applying a state-of-the-
art commercial solver to a standard integer programming formulation. Moreover, we
evaluate the quality of our algorithm in several computational experiments.
We present several approximation results for the APP including a 23 -approximation
algorithm. Moreover, we show that, for a fixed number of stores, the problem can be
123
solved in polynomial time. Additionally, we analyze how substitution effects can be

included in our model.
From a managerial perspective, our results show that using an optimized mixed
assortment planning strategy as suggested here can lead to substantial gains in profit
for a retailer with multiple chain stores.
Our computational results indicate that the relative gain in profit of such a mixed
strategy compared to one of the basic strategies can be up to 20%. The magnitude of
this benefit depends on demand differences between stores and scaling effects that are
included in our model as “bonus values,” which measure the gain in profit obtained
by putting a product into the common assortment.
2 Formulation of the APP as an integer program
In this section, we formulate a stylized model for the APP that is obtained by simpli-
fying the original problem using some reasonable assumptions (see also Corsten and
Kasper 2016; Corsten et al. 2017).
Assumption 1 We develop our model using the following assumptions:
– The assortment consists of standardized products, offered at standardized shelf
space (i.e., all products have unit size).
– Every store has the same capacity.
– Demand can be estimated for every product and every store.
– Products that are assigned to the common assortment are offered in every store,
while products in a local assortment are offered only in this particular store.
The most restrictive assumption seems to be that all products have unit size. How-
ever, we can view the unit size as a standardized area or facing (e.g., 1m 2 ) that is
occupied by each item. Then, an item corresponds to the number of units of a particu-
lar product that fit into this area (e.g., 1000 pencils, or 5 toasters) and the profit obtained
from the item is scaled accordingly (see, for example, Borin et al. 1994 for a similar
approach). Moreover, in some cases, only a single unit of each product is displayed in
a store within a standardized area. For example, a retailer for luxury goods confirmed
in a recent discussion that it has only one unit of every product for show in each store,
arranged on uniform displays. Considering the APP without the assumption of unit
size products is part of our future work (see also Sect. 7).
Concerning the assumption that every store has the same capacity, we note that the
size of the common assortment is always bounded by the smallest store capacity. Any
additional capacity that might exist in some of the stores always has to be used by the
respective local assortment. Thus, this assumption is not too restrictive. Nevertheless,
we show in Sect. 4.5 how our algorithms can be generalized to the case of different
store capacities.
The third assumption regards demand estimation, which is itself an important field
of research for which a magnitude of methods have been developed (cf. Sect. 1.1.1).
Primary demand data for every product as used here can, for example, be obtained
from readily available data such as scanner data by using the models presented in
Chong et al. (2001), Kök (2003), Kök and Fisher (2007), Vulcano et al. (2012).
123
2.1 Integer programming formulations
Now, we describe the APP as a packing problem. We are given m bins (stores) with
size K (capacity of the store) in which we want to pack unit size items (products). In
total, there are n ≥ K different items. With each item j, we associate m + 1 different
profits w j ≥ 0 and v jk ≥ 0 for k = 1, . . . , m. We obtain profit w j if item j is packed
into all bins (i.e., packed into the common assortment) and profit v jk if item j is
packed into bin k, but there is at least one bin into which it is not packed. Using this
notation, we can model the problem as the following integer program:

n
n
m
maximize wjxj + v jk y jk
j=1 j=1 k=1
n
subject to (x j + y jk ) ≤ K ∀k ∈ {1, . . . , m}
j=1
x j + y jk ≤ 1 ∀ j ∈ {1, . . . , n}, k ∈ {1, . . . , m}
x j , y jk ∈ {0, 1} ∀ j ∈ {1, . . . , n}, k ∈ {1, . . . , m}
where

1, if item j is packed into the common assortment
xj =
0, else
and

1, if item j is packed into bin k (but not into the common assortment)
y jk =
0, else.
The profits w j and v jk are composed of estimated revenue and cost as follows:
Let r jk be the estimated revenue of product j in store k, cd jk the cost of type d (e.g.,
procurement, transportation, storage cost) of product j in store k, c x the assignment
y
cost for assigning a product to the common assortment, and ck the assignment cost for
assigning a product to the local assortment of store k. Then, we can write the objective
function of the APP as
y
r jk (x j + y jk ) − cd jk (x j + y jk ) − c x x j + ck y jk .
j k j k d j k
where

y
wj = r jk − cd jk − c x
and v jk = r jk − cd jk − ck .
k d d
Thus, the estimation of the parameters w j and v jk is essentially the estimation of the
revenue and cost of the products.
123
The main goal of this model for the APP is to decide which products to place in the
common assortment and which in the local assortments, given estimated parameters w j
and v jk . As mentioned earlier, the APP also models the case when a retailer wants to
decide which products to select for a single category. Then, all following results hold
with K being the size of the category instead of the size of the store. For convenience,
we mostly just refer to the case where K is the size of the store.
It is worthwhile to mention another integer programming formulation for the APP
that provides an alternative view on the profits obtained from the items in the common
assortment. In this alternative formulation, we use two vectors x and y of binary
decision variables, where x j is again set to one if and only if item j is packed into the
common assortment, but y jk is now set to one whenever item j is packed into bin k,
no matter if j is part of the common assortment or only part of the local assortment
of bin k. The objective is now to maximize nj=1 b j x j + nj=1 m
k=1 v jk y jk , where
m
b j := w j − k=1 v jk is the bonus value of item j.
Moreover, given a packing with common assortment C, we refer to B := j∈C b j
as the total bonus value of this packing. The new formulation can be obtained from
the original one by the transformation y jk := x j + y jk and x j := x j for all j, k.
The bonus value b j of a product j is a measure of the magnitude of scaling effects
such as economies of scale, recognition value, customer loyalty, or coordinated adver-
tisement. Economies of scale Stigler (1958) refer to cost reductions that result from
improved contracts with suppliers as discussed, for example, in Cadeaux (1999), Hin-
gley (2005). The recognition value of a product describes that a customer chooses to
enter a particular chain store because she knows that a certain product will be avail-
able in all of its stores. As argued, for example, in Porter and Claycomb (1997), the
availability of anchor brand products greatly improves retail store image and, thus,
customer loyalty.
3 Computational complexity of the APP
In this section, we analyze the computational complexity of the APP. Theorem 2 states
that the problem is strongly NP-hard. Thus, unless P = NP, there is no algorithm that
solves the APP exactly in polynomial time.
Theorem 2 The APP is strongly NP-hard.
Proof We use a reduction from the satisfiability problem SAT. An instance of SAT
consists of n Boolean variables x1 , . . . , xn and m disjunctive clauses C1 , . . . , Cm of
arbitrary length.
Given an instance of SAT, we construct an instance of the APP with 2n + 1 items
and n + m + 1 bins of size n + 1 as follows: There is an item for each positive literal
denoted by xi , an item for each negative literal denoted by x i , and one additional
item z. The first n bins A1 , . . . , An correspond to the variables and the next m bins
B1 , . . . , Bm correspond to the clauses. Additionally, there is one extra bin Z . We define
the following profits for the items:
123

1, if k = i 1, if x i ∈ Ck
vxi ,Ak := vxi ,Bk := vxi ,Z := 0
0, else 0, else

1, if k = i 1, if xi ∈ Ck
vx i ,Ak := vx i ,Bk := vx i ,Z := 0
0, else 0, else
vz,Ak := 0 ∀k vz,Bk := 0 ∀k vz,Z := N
For the items corresponding to the literals, we set wxi := M and wx i := M, where
M > n + m + 1 is a large integer. For the additional item z, we set wz := N , where
N > M is an even larger integer.
We now show that the given instance of SAT admits a satisfying assignment if and
only if there exists a packing with objective value at least n M + n + m + N for the
constructed instance of the APP.
First, suppose that the SAT instance admits a satisfying assignment. Then, we pack
item xi (x i ) into the common assortment if variable xi from the SAT instance is set
to TRUE (FALSE) in this assignment. This results in a total profit of n M from the
common assortment. Now, exactly one additional item fits into each bin. We put item z
into bin Z and, thus, obtain an additional profit of N . For each i, exactly one of the
two items xi , x i is packed into the common assortment and, thus, we can put the
other one into bin Ai , which yields an additional profit of n when considering all
bins A1 , . . . , An . Moreover, we know that each clause C j is satisfied, so, for each j,
there exists a literal xi ∈ C j (x i ∈ C j ) for which the corresponding variable xi is set to
TRUE (FALSE). By the construction above, this means that item xi (x i ) is packed into
the common assortment and x i (xi ) is not. Thus, we can pack item x i (xi ) into bin B j
and obtain an additional profit of 1. Hence, since all clauses are satisfied, we obtain
an additional profit of m in this way, which yields a total profit of n M + n + m + N .
Now suppose that there exists a packing with objective value at least n M +n+m+N
for the constructed instance of the APP. In this packing, item z cannot be in the
common assortment since, otherwise, the most profitable way to pack items is to put
n items into the common assortment (as M > n + m + 1), which results in an overall
profit of just N + n M. Moreover, we can assume without loss of generality that item z
is packed into the last bin Z since this yields a profit of N and the alternatives of
packing another item into bin Z or packing another item into the common assortment
yield profits 0 and M < N , respectively. Also, since M > m + n + 1, we have to put
n items into the common assortment in order to obtain profit n M. In order to obtain
the remaining profit of n + m, each item packed into one of the remaining n + m empty
slots of the bins A j and B j has to yield a profit of one. Thus, for each i, one of the
items xi and x i must not be packed into the common assortment since, otherwise, no
item can generate further profit in bin Ai . Hence, we can define a truth assignment of
the variables in the SAT instance by setting variable xi to TRUE (FALSE) if item xi
(x i ) is packed into the common assortment. This truth assignment satisfies all clauses
since an item xi (x i ) that generates a profit of 1 is packed into each bin B j , which
means that the literal x i ∈ C j (xi ∈ C j ) satisfies the clause.

123
Note that the reduction presented in the proof of Theorem 2 does not lead to an
inapproximability result for the APP. In particular, we do not know whether the APP
is APX-hard or if there exists a PTAS. In the next section, however, we provide a
constant factor approximation.
In the following, we show that the APP can be solved in polynomial time for a fixed
number of bins (i.e., if the number of bins is considered as a constant that is not part
of the input). First, we show that, for the case m = 2, i.e., the case of only two bins,
the problem can be solved efficiently. Then, we present a combinatorial algorithm that
solves the problem in polynomial time for any fixed number of bins.
Lemma 1 The APP can be solved in polynomial time for m = 2.
Proof For m = 2, the APP can be written as a minimum cost flow problem as shown
below. However, we first show the polynomial-time solvability by using the second
integer programming formulation mentioned in Sect. 2. Given this formulation, for
m = 2, we obtain the constraint matrix
⎛ ⎞
1 1 ... 1
⎜ 1 1 ... 1⎟
⎜ ⎟
⎜1 −1 ⎟
⎜ ⎟
⎜1 −1 ⎟
⎜ ⎟
A=⎜ 1 ⎜ −1 ⎟
⎟
⎜ 1 −1 ⎟
⎜ ⎟
⎜ ... ... ... ⎟
⎜ ⎟
⎝ 1 −1 ⎠
1 −1
which is totally unimodular. To see this, we partition the 2n + 2 rows of A into the
odd rows I1 and the even rows I2 . Then, it is immediate that for every subset J of the
rows,
the partition of J into J1 := I 1 ∩ J and J2 := I 2 ∩ J satisfies | j∈J1 A ji −
A
j∈J2 ji | ≤ 1 for each column i, which implies that the matrix is totally unimodular
(cf. Nemhauser and Wolsey 1988). Since all right-hand sides of the constraints are
integral, this implies that the associated polyhedron is integral and the problem can be
solved in polynomial time by solving its linear programming relaxation.

Moreover, as mentioned above, we can write the APP with m = 2 as a minimum
cost flow problem. In order to construct the corresponding graph, we introduce a source
node s, a sink node t, and two nodes v j1 and v j2 for each item j. There is an arc from
s to every node v j1 (for every j), and an arc to t from every node v j2 (for every j).
Additionally, from every node v j1 (for every j), there is an arc to every node vl2 (for
every l), which yields n 2 arcs between these nodes in total. The capacity of every arc
is set to one. The arcs adjacent to s and t have cost zero. The arc between nodes v j1
and vl2 has cost −(v j1 + vl2 ) if j
= l and −(v j1 + vl2 + b j ) if j = l. Then, it is easy
to see that an optimal solution of the APP can be obtained by computing a minimum
cost flow with flow value K in this graph.
For m ≥ 3, the constraint matrix of the IP formulation above is not totally uni-
modular anymore. Indeed, the following example shows that the integrality gap of the
formulation from the proof of Lemma 1 is greater than zero for m ≥ 3 bins.
123
Example 1 Consider the case that m = 3, n = 3, and K = 2. Let a > 0 be some

arbitrary number and let the values v jk and w j be given by
⎛ ⎞ ⎛ ⎞
2a 0 0 3a
(v jk ) j,k =⎝ 0 a 0 ⎠ and (w j ) j = ⎝ 3a ⎠ .
0 0 a 3a
Then, the optimal objective value of this instance of the APP is 6a, whereas the
optimal objective value of the linear programming relaxation is 13
2 a. This leads to an
13
integrality gap of 12 ≈ 1.08. Extending this idea to more bins does not amplify the
gap.
Although the integrality gap of the canonical IP formulation from the proof of
Lemma 1 is greater than zero for m ≥ 3, the following theorem shows that, for
general fixed m, the APP is still solvable in polynomial time. The proof is provided
in “Appendix.”
m +m−1
Theorem 3 The APP can be solved in time O(n 2 ).
Theorem 3 directly yields the following corollary:
Corollary 1 If the number of bins m is fixed, the APP can be solved in polynomial
time.

Note that the algorithm provided in the proof of Theorem 3 is not suitable for
solving instances of realistic size because its running time is doubly exponential in
the number of bins. Algorithms more suitable for practice will be discussed in the
following section.
4 Algorithmic approach
As it turns out (see Sect. 6), commercial solvers are unable to obtain optimal solutions
of the APP within a reasonable amount of time even for instances with 300 stores and
15.000 products. This motivates the development of algorithms that run fast and pro-
duce close-to-optimal solutions since, for example, large supermarket chains consist
of more than 10.000 stores and more than 50.000 products.
In this section, we first present several algorithms for the APP including a 2-
approximation algorithm, an algorithm that works well if the common assortment of
an optimal solution is small, and a 23 -approximation algorithm. Afterward, we provide
a greedy heuristic that iteratively uses local improvements and yields close-to-optimal
solutions for most instances considered in our experimental results in Sect. 6.
4.1 A 2-approximation algorithm
The reduction presented in the proof of Theorem 2 does not lead to an inapproximabil-
ity result for the APP. Indeed, a polynomial-time 2-approximation algorithm is easy
123
value of a given instance of the APP

to obtain. Let W denote the optimal objective
when the objective function is changed to j w j x j (i.e., only the profit obtained from
the common assortment is maximized), and let V denote the optimal objective value of
the instance when the objective function is changed to jk v jk y jk (i.e., only the profit
obtained from the local assortments is maximized). These values can be computed
easily in polynomial time (optimal solutions are given by packing the K items with
the largest values w j into the common assortment for the first objective function and
by packing the K items with the largest values v jk into every bin k for the second
objective function). From now on, we refer to the corresponding optimal solutions as
the trivial solutions. Letting OPT denote the optimal objective value of the original
instance of the APP, we then have
OPT ≤ W + V ≤ 2 · max{W, V }.
Therefore, choosing the better one of the two trivial solutions yields a 2-approximation.
With a more careful analysis, one can show the following stronger result:
V W
Lemma 2 OPT ≤ 1 + min W , V · max{W, V }.
Proof It holds that

V V V
OPT ≤ V + W = W · +W = 1+ ·W ≤ 1+ · max{W, V } and
W W W

W W W
OPT ≤ V + W = V + V · = 1+ ·V ≤ 1+ · max{W, V }.
V V V

Lemma 2 shows that the worst-case approximation ratio of 2 can only occur if
W = V . Indeed, the approximation factor of the algorithm that chooses the better
of the two trivial solutions is heavily dependent on the distribution of the values w j
and v jk . In particular, if all the values w j and v jk are contained in some small interval,
the approximation ratio is small:
Lemma 3 Suppose that there exist ω and 0 ≤ < 1 such that
(1 − )ω ≤ w j ≤ ω ∀j and
(1 − )ω ≤ m · v jk ≤ ω ∀ j, k.
Then, OPT ≤ 1
1− · max{W, V }.
Proof Assume that there are L items in the common assortment of a fixed optimal
solution. We define W L as the sum of the L largest values w j , and, for each k, we
define VkK −L as the sum of the (K − L) largest values v jk .
Then, we have

m
OPT ≤ W L + VkK −L .
k=1
123
However, since w j ≥ (1 − )ω for all j and m · v jk ≥ (1 − )ω for all j, k, we also
have
W ≥ W L + (K − L)(1 − )ω and

m
V ≥ VkK −L + L(1 − )ω.
k=1
Hence, if W ≥ V , we obtain
K −L
OPT OPT WL + m k=1 Vk
= ≤ L
max{W, V } W W + (K − L)(1 − )ω
W + (K − L)ω
L 1
≤ L ≤ ,
W + (K − L)(1 − )ω 1−
where the second inequality follows since m · v jk ≤ ω for all j, k.

If V ≥ W , we obtain the same inequality by using that w j ≤ ω for all j.

Observe that the main difficulty of the APP is to choose the items that belong to the
common assortment: If the set of items in the common assortment of some optimal
solution is known, filling the remaining space in each bin k with the items with the
largest values v jk among the items not already contained in the common assortment
yields an optimal solution. This argumentation also yields the following result:
Corollary 2 The APP can be solved in polynomial time if the store capacity K is
fixed independently of the input size.
K n
i ∈ O(n ) possible sets of items in the common
Proof We enumerate all the i=0 K
assortment, which are polynomially many as long as K is fixed. For each given set
of items in the common assortment, we can then fill the remaining space in each bin
greedily as above, and choosing the best one among the solutions obtained in this way
yields an optimal solution.

4.2 An approximation in terms of the size of the common assortment
We now present an algorithm that roughly guarantees an approximation factor of

∗
(1 + LK ), where L ∗ denotes the size of the common assortment of a fixed optimal
solution. In particular, for instances where the size of the common assortment of an
optimal solution is small, this algorithm improves upon the approximation factor of 2
obtained by choosing the better one of the two trivial solutions.
Given a parameter > 0, the algorithm enumerates all possible choices for the
common assortment containing at most a := 2 + 2 items while filling the remaining
space in each bin k with the items with the largest values v jk among the items not
already contained in the common assortment (hence, the algorithm will return an
optimal solution as long as the size of the common assortment of some optimal solution
is at most a). Afterward the algorithm generates one solution for each possible size i >
123
common assortment by packing the i items j with the largest bonus values b j =
a of the
wj − m k=1 v jk into the common assortment and filling the remaining capacity as
before. Note that, since only a polynomial number of solutions is enumerated for a
fixed parameter value > 0, the running time of the algorithm is polynomial in the
instance size for every fixed .
Algorithm 1 Algorithm GUESS AND PACK

1: Given > 0, set a := 2 + 2 .
2: for i = 0 to K do
3: if i ≤ a then
4: Enumerate all solutions where the size of the common assortment is equal to i and the remaining
K − i places in each bin k are filled by the K − i items j with the largest local profits v jk among
the items not already packed into the common assortment. Let sol(i) denote the solution with the
highest total profit among these solutions.
5: else
6: Pack the i items j with the largest bonus values b j = w j − m k=1 v jk into the common
assortment.
7: for k = 1 to m do
8: Among the items not already packed into the common assortment, pack the K − i items j with
the largest local profits v jk into bin k.
9: end for
10: Let sol(i) denote the resulting solution.
11: end if
12: end for
13: return Solution with the highest total profit among sol(0), . . . , sol(K ).
Lemma 4 Let > 0 and let L ∗ denote the number of items in the common
∗
assortment of an arbitrary optimal solution. Then, Algorithm 1 is a (1 + )· 1 + LK -
approximation algorithm for the APP.
Proof If L ∗ ≤ a, Algorithm 1 returns an optimal solution. Hence, we may assume for

the rest of the proof that L ∗ > a.
Consider an optimal solution with L ∗ items in the common assortment and let
B , V ∗ denote the total bonus value and the total individual profit, respectively,
∗
∗K ∗
obtained in this optimal solution. Since 0 ≤ LL∗ +K ≤ LL ∗K = K , there exists
∗
an iteration of the algorithm in which i = LL∗ +K
K
. Let B denote the total bonus value
and V the total individual profit in the solution sol(i) produced in this iteration. Then,
L∗ K
B∗ ≤ · B and V ∗ ≤ ·V
i K −i
since the i items j with the largest bonus values b j are contained in the common
assortment of sol(i) and, for each bin, the K −i items j with the largest local profits v jk
are contained in bin k of sol(i). Moreover, we have
123
K K K · (L ∗ + K ) L∗ + K L∗
≤ ∗K = = = 1 + and
K −i K − LL∗ +K K · (L ∗ + K ) − L ∗ K L∗ + K − L∗ K

L∗ L∗ L∗ 1
≤ L∗ K = 1+ · ∗
i L ∗ +K − 1
K 1 − LL ∗+KK

L∗ 1 L∗
≤ 1+ · = 1+ · (1 + )
K 1 − a2 K
where we used that a ≤ L ∗ ≤ K and a = 2 + 2 . Since OPT = B ∗ + V ∗ , this shows

the claim.

Algorithm 1 can be used to obtain a bicriteria approximation algorithm for the APP.
Here, for α, β ≥ 1, a (bicriteria) (α, β)-approximation algorithm is a polynomial-time
algorithm ALG such that OPT ≤ α · ALG holds for every instance of the APP and
ALG packs at most β · K items into each bin (i.e., it violates the bin capacities by a
factor of at most β).
Lemma 5 For , γ > 0, applying Algorithm 1 to the instance

with modified bin

capacity K := (1 + γ ) · K yields a bicriteria (1 + ) · 1 + 1+γ
1
, (1 + γ ) -
approximation algorithm for the APP.
Proof For a given instance of the APP, let L ∗ denote the number of items in the
common assortment in a fixed optimal solution of the instance (with the original bin
capacity K ). Then
L∗ K 1
1+ ≤1+ =1+ .
(1 + γ ) · K (1 + γ ) · K 1+γ

√
In particular, letting Φ := 5+1
2 denote the golden ratio and choosing γ := Φ − 1
in Lemma 5 immediately yields the following corollary:
√
Corollary 3 Let > 0 and let Φ := 5+1 2 denote the golden ratio. Then, applying
Algorithm 1 to the instance with modified bin capacity K := Φ · K yields a bicriteria
((1 + )Φ, Φ)-approximation algorithm for the APP.

4.3 A 23 -approximation algorithm
In the following, we develop a polynomial-time algorithm that achieves an approxi-

mation factor of 23 for the APP. To this end, we first present three different algorithms
for the APP that do not achieve constant approximation guarantees. However, we then
show that applying all three algorithms to an instance of the APP and choosing the
best among the three solutions yields an approximation factor of 23 .
123
For a given instance of the APP, let OPT denote the optimal objective value of the
instance. We fix an arbitrary optimal solution and let C ∗ denote the common assortment
of this optimal solution and, for each bin k, let Ik∗ denote the local assortment of this
∗
optimal solution. Moreover, we let B := j∈C ∗ b j denote the total bonus value of

the optimal solution, where, as before,
bj = w j − m k=1 v jk denotes the bonus value of
item j. Similarly, we let V1∗ := m k=1 ∗ v
j∈Ik jk denote the total profit obtained from

items in the local assortments of the optimal solution and V2∗ := m k=1 j∈C ∗ v jk
denote the total local profit obtained from the items in the common assortment of the
optimal solution. The value V ∗ := V1∗ + V2∗ then corresponds to the total local profit
of the optimal solution and we have
OPT = V ∗ + B ∗ = V1∗ + V2∗ + B ∗ .
When considering the solution returned by one of our algorithms, denote the analogous
profit values of this solution by B, V1 , V2 , and V = V1 + V2 .
We now present the three algorithms that are used in order to obtain our 23 -
approximation algorithm for the APP. The first two of these algorithms simply produce
the best packings with C = ∅ and |C| = K that were already used at the beginning of
this section in order to obtain an approximation factor of 2. We now show that using
an additional third algorithm improves the approximation factor to 23 .
Algorithm 2 Algorithm LOCAL

2: Pack the K items j with the largest local profits v jk into bin k.
3: end for
4: return Solution obtained.
Lemma 6 Algorithm 2 achieves an objective value of at least V ∗ and can be imple-

mented to run in O(mn log n) time.
Proof Since Algorithm 2 packs the K items with the largest local profits into each bin
(which corresponds to the trivial packing with respect to the local profits considered
before) and there can be at most K items in each bin of the optimal solution, the
objective value obtained by the algorithm solution must be at least V ∗ . The claimed
running time bound follows since, for each bins k, we can simply sort the n items by
nonincreasing local profits v jk in O(n log n) time in order to determine the K items
to pack into bin k.

Note that Algorithm 2 does, in general, not achieve a constant factor of the optimal
profit since the local profits can be very small compared to the bonus values of the
items (e.g., when all local profits are zero, the algorithm obtains no positive profit at
all).
123
Algorithm 3 Algorithm COMMON

1: Pack the K items j with the largest common assortment profits w j into the common assortment.
2: return Solution obtained.
Lemma 7 Algorithm 3 achieves an objective value of at least V2∗ + B ∗ and can be

implemented to run in O(n log n) time.
Proof Since Algorithm 3 packs the K items with the largest common assortment profits
into the common assortment (which corresponds to the trivial packing with respect to
the common assortment profits considered before) and there can be at most K items
in the common assortment of the optimal solution, the objective value obtained by the
algorithm must be at least the objective value obtained from the common assortment of
the optimal solution, which is V2∗ + B ∗ . The claimed running time bound follows since
we can simply sort the n items by nonincreasing common assortment profits w j in
O(n log n) time in order to determine the K items to pack into the common assortment.

Note that Algorithm 3 does, in general, not achieve a constant factor of the optimal
profit either since the common assortment profits can be very small compared to the
local profits of the items.
Algorithm 4 Algorithm BONUS

1: for L = 1 to K do
2: Pack the L items j with the largest bonus values b j = w j − m k=1 v jk into the common assortment.
4: Among the items not already packed into the common assortment, pack the K − L items j with
the largest local profits v jk into bin k.
5: end for
6: Let sol(L) denote the solution obtained.
7: end for
8: return Solution with the highest total profit among sol(1), . . . , sol(n).
Lemma 8 Algorithm 4 achieves an objective value of at least V1∗ + B ∗ and can be

implemented to run in O(mn 2 ) time.
Proof If C ∗ = ∅, then V2∗ = B ∗ = 0 and we are in the case of Algorithm 2. Otherwise,

we have 0 < |C ∗ | ≤ K , so there exists an iteration of the algorithm in which L = |C ∗ |
is considered. In this iteration, the algorithm packs the |C ∗ | items with the largest bonus
values b j into the common assortment, so the total bonus value B of the produced
solution sol(|C ∗ |) is at least B ∗ . Moreover, since the algorithm fills the remaining
K − |C ∗ | places in each bin k with the K − |C ∗ | items with the largest local profits v jk
among those it has not already packed into the common assortment, the total local
profit V of the solution sol(|C ∗ |) is at least as large as the local profit V1∗ that the
optimal solution obtains only from the items in its local assortments (which contain
at most K − |C ∗ | items for each bin). Hence, the profit of the solution sol(|C ∗ |) is at
least V1∗ + B ∗ , which proves the claimed bound on the objective value.
123
The claimed running time bound follows since we can sort the n items by nonin-
creasing local profits v jk once for each bin k and sort the items once by nonincreasing
bonus values b j in a total of O(mn log n) time. Afterward, each of the n iterations of
the algorithm can be performed in O(mn) time, which yields a total running time of
O(mn log n + mn 2 ) = O(mn 2 ).

We now show that applying all three algorithms to an instance of the APP and
choosing the best among the three solutions yields an approximation factor of 23 :
Theorem 4 If LOCAL, COMMON, and BONUS denote the objective values

obtained by Algorithms 2, 3, and 4, respectively, on an instance of the APP, then
3
· max{LOCAL, COMMON, BONUS} ≥ OPT.
2
Hence, the algorithm that chooses the best among the three solutions obtains an
approximation factor of 23 -approximation for the APP. It can be implemented to run
in O(mn 2 ) time.
Proof If ALG := max{LOCAL, COMMON, BONUS} denotes the objective value

obtained by the algorithm that chooses the best among the three solutions, then Lem-
mas 6, 7, and 8 yield that

OPT V ∗ + B∗ V ∗ + B∗ V ∗ + B∗
≤ min , ∗ , .
ALG V∗ V2 + B ∗ V1∗ + B ∗
Assume for the sake of a contradiction that the minimum on the right-hand side is
strictly larger than 23 . Then, we obtain
1 ∗
B∗ > V
2
1 ∗
V1∗ > V2 + B ∗
2
1 ∗
V2∗ > V + B∗ .
2 1
However, adding the last two inequalities and using that V1∗ + V2∗ = V ∗ yields
V ∗ > 2B ∗ , which contradicts the first inequality. Thus, we must have ALGOPT
≤ 23 ,
which proves the claimed approximation factor. The claimed running time bound
follows directly from the bounds on the running times of Algorithms 2, 3, and 4 given
in Lemmas 6, 7, and 8, respectively.

4.4 A greedy heuristic
We now present a greedy heuristic that is used in Sect. 6 to solve large-scale instances
of the APP. The idea of the algorithm is to first neglect the advantages of using the
123
common assortment and start with the best local assortment for each store (independent
of the others). Then, in each step, the algorithm chooses an item that currently grants
the largest gain in total profit when being added to the common assortment and adds
this item to the common assortment (while removing either the item itself or the least
profitable item from the local assortment of each store).
Denoting the set of items that may still be added to the common assortment by S, the
current common assortment by C, and the current set of items in the local assortment
of bin k by Ik , the algorithm can be formulated as follows:
Algorithm 5 Algorithm GREEDY

1: Let S = {1, . . . , n}, C = ∅ and, for each bin k, let Ik be the set containing the K items j with the highest
values v jk .
2: for j ∈ S do
3: Let u j = w j − / k minl∈Ik vlk − k: j∈Ik v jk .
k: j ∈I
4: if u j ≤ 0 then
5: Remove item j from S.
6: end if
7: end for
8: if S = ∅ or |C| = K then
9: return Common assortment C and local assortments Ik .
10: else
11: Add an item j ∈ S with maximum value u j to the common assortment C, remove j from S, and
update the sets Ik by removing j from Ik if it is contained in Ik and removing an item l with minimum
value vlk from Ik otherwise, and go to line 2.
12: end if
For each item i ∈ S, the value u j calculated in line 3 represents the change of the
total profit of the current solution when item j is added to the common assortment,
removed from all local assortments, and, in each bin k in which item j was not
contained in the current local assortment Ik , an item l with lowest local profit vlk is
removed from Ik .
Observe that, when u j ≤ 0 for some item j in line 4 in some iteration of the
algorithm, item j is removed from the set S in line 5 and, thus, will not be added to the
common assortment in any future iteration. This is motivated by the following result:
Lemma 9 The value u j of any item j considered in Algorithm 5 can never increase.
Proof Fix an item j and consider the set Bin( j) := {k : j ∈ Ik } of all bins k that
contain item j in their local assortment Ik . As long as the set Bin( j) remains unchanged
during the algorithm, the value u j cannot increase since the sets Ik can never become
larger during the execution of the algorithm, so the term minl∈Ik vlk can only increase.
If Bin( j) changes, this must be because item j is removed from some of the sets Ik .
Removing item j from a set Ik decreases the second sum in the definition of u j by v jk
and increases the first sum by minl∈Ik :l
= j vlk . However, since item j will only be
removed from Ik in line 11 if minl∈Ik vlk = v jk holds before removing j from Ik , we
must have minl∈Ik ,l
= j vlk ≥ v jk , which shows that u j does not increase.

By Lemma 9, an item j for which u j ≤ 0 holds in some iteration of the algorithm
will never increase the total profit of the solution when being added to the common
123
assortment in any future iteration since u j will never become positive again. Hence,
the algorithm removes such an item j from the set S of items to be considered in later
iterations.
We note that Algorithm 5 starts with the best solution with C = ∅ and the total profit
of its current solution can never decrease in any iteration. Hence, since choosing the
better one among the best solutions with |C| = K and C = ∅ yields an approximation
ratio of 2 as shown at the beginning of this section, Algorithm 5 can easily be turned
into a 2-approximation algorithm for the APP by comparing the objective value of the
returned solution to objective value of the best solution with |C| = K and choosing
the better solution. The running time of Algorithm 5 can easily be seen to be in
O(nm(log(n) + K )).
4.5 Generalization to different store capacities
All algorithms presented in this section can be generalized to stores with different
capacities. In this case, the size of the common assortment is obviously bounded by
the size of the smallest store. Therefore, for larger stores, there is always some part of
the capacity that is reserved for the local assortment.
Algorithm 5 can easily be generalized to the case where each store k has a different
capacity K k . Then, in the first step, for each bin (store) k, we pack the best K k items.
Moreover, the second stopping criterion in line 8 changes to |C| = mink K k .
It is also straightforward to see how Algorithms 2, 3, and 4 can be generalized to
different store capacities. Again, the size of the common assortment that is consid-
ered will be bounded by the capacity of the smallest store. The same holds true for
Algorithm 1. The same analyses as for the single capacity case show that the same
approximation ratios are still obtained in the case of different store capacities.
5 Modeling substitution effects
If a requested product is unavailable for a customer, she might buy another one instead,
i.e., she substitutes the requested product with another one. Here, we consider out-of-
assortment substitution, which describes the case of substitution if a requested product
is not part of the assortment (in contrast to out-of-stock substitution, which describes
the case where a product is temporarily unavailable).
We model the substitution effects by using the exogenous demand model. Utility
models like the MNL suffer from various limitations, most notably the independence
of irrelevant alternatives property, which states that the choice probability ratio for
two products is independent of the other products on offer (i.e., in a category, substi-
tution within a subgroup is not more likely)—a restriction that does not hold in many
situations. The exogenous demand model does not suffer from these drawbacks. Addi-
tionally, the exogenous demand model is a more universal approach where demand
and substitution behavior for every product are directly specified. By setting substitu-
tion parameters accordingly, it even incorporates, among others, choice models like
the MNL model. Thus, it is a flexible approach for modeling substitution effects in
123
various settings. For a closer examination, we refer to Hübner et al. (2016), Kök and
Fisher (2007), Kök et al. (2009).
We now formulate a model for the regionalized assortment planning problem with
substitution effects under the exogenous demand model. Note that our model also
covers the general case of defining any substitution effect between any pair of products.
The parameter K can both indicate the size of the stores as well as the predefined size of
a category. In more detail, we regard the following setting. If a product is unavailable,
the customer decides with some probability to buy another product: if product j is not
in the assortment but product l is, we obtain some extra profit s jl with probability α jl ,
i.e., an expected extra profit of s jl := α jl s jl . Note that it would also be possible
to define these substitution profits store-dependent, which can be easily incorporated
into our model by replacing s jl with s jlk . However, it is unclear whether this has
practical relevance, so we stick to store-independent substitution profits here. We call
this variant of the APP the assortment planning problem under substitution (S-APP).
The S-APP can be formulated as the following integer program:

n
n
m
m
n
n
maximize wjxj + v jk y jk + s jl z jlk
j=1 j=1 k=1 k=1 j=1 l=1

n
subject to (x j + y jk ) ≤ K ∀k ∈ {1, . . . , m}
j=1
x j + y jk ≤ 1 ∀ j ∈ {1, . . . , n}, k ∈ {1, . . . , m}
z jlk ≤ xl + ylk ∀ j, l ∈ {1, . . . , n}, k ∈ {1, . . . , m}
z jlk ≤ 1 − x j − y jk ∀ j, l ∈ {1, . . . , n}, k ∈ {1, . . . , m}
x j , y jk , z jlk ∈ {0, 1} ∀ j, l ∈ {1, . . . , n}, k ∈ {1, . . . , m}
where

xj =
0, else
and

1, if item j is packed into bin k (but not into the common assortment)
y jk =
0, else
and

1, if item j is not packed into bin k, but item l is packed into bin k
z jlk =
0, else.
The parameter s jl describes the additional profit obtained from customers substi-
tuting product j with product l. Therefore, we naturally have s j j = 0. The third and
fourth sets of constraints ensure that the additional profit s jl can only be obtained for
a store k if product j is not offered in store k, but product l is part of the assortment
of this store.
123
Note that the S-APP is an extension of the APP, which is obtained from the S-APP
by setting s jl := 0 for all j, l. Concerning the complexity of the S-APP, we show the
following result:
Proposition 1 The S-APP is APX-hard.
Proof We prove the claim by a reduction from the maximum cut problem (MAX CUT).
Given an instance of MAX CUT specified by an undirected graph G = (V, E), we
construct an instance of the S-APP as follows: There is a product for every node of G
and we set m := 1, K := n, and w j := v jk := 0 for all j, k. The substitution profits
are given by setting s jl := sl j := 1 if ( j, l) ∈ E and s jl := sl j := 0, otherwise. Then,
for every cut (S, V \S) in G, packing exactly the items corresponding to the nodes
in S (into the common or into the local assortment) yields a solution of the constructed
instance of the S-APP whose profit equals the number of edges crossing the cut.
Conversely, for any solution of the constructed instance of the S-APP, defining S as
the set of nodes whose corresponding items are part of the assortment yields a cut
(S, V \S) for which the number of edges crossing the cut equals the total profit of this
solution. Since MAX CUT is APX-hard (Papdimitriou and Yannakakis 1991), this
shows the claim.

5.1 Some approximation results
In order to obtain a polynomial-time constant-factor approximation algorithm for

the most general case of the S-APP, we generalize the idea of our 2-approximation
algorithm for the APP presented at the beginning of Sect. 4. In addition to the vari-
ants of the problem where only the first part of the objective function j w j x j (the
profitobtained from the common assortment) or the second part of the objective func-
tion jk v jk y jk (the profit obtained from the local assortments) is maximized, we
now consider
m n a third variant in which only the third part of the objective function
n
k=1 j=1 l=1 s jl z jlk (the profit obtained from substitution) is maximized. For a
given instance of the S-APP, we let S denote the optimal objective value of this third
variant. Denoting the optimal objective values of the first two variants by V and W as
before, and letting OPT denote the optimal objective value of the original instance of
the S-APP, we then obtain that
OPT ≤ V + W + S ≤ 3 · max{V, W, S}.
Unfortunately, as seen in the proof of Proposition 1 (where all local profits v jk and
all common assortment profits w j were equal to zero in the constructed instance
of the S-APP), the problem ofcomputing
n an optimal solution for the third variant
with objective function nj=1 m k=1 l=1 s jl z jlk is already APX-hard, so it cannot
be solved optimally in polynomial time unless P = NP. However, since OPT ≤
V + W + α · S ≤ (2 + α) · max{V, W, S} for α ≥ 1, any α-approximation algorithm
for the third variant of the problem directly yields a (2 + α)-approximation algorithm
for the S-APP.
123
By a construction similar to the one used in theproof of Proposition 1, we can view
the third variant with objective function nj=1 m k=1
n
l=1 s jl z jlk as the following
(weighted) graph bisection problem: Given a directed graph with a nonnegative weight
on each arc, partition the nodes into two disjoint sets A and B such that |A| ≤ K and the
total weight of the arcs with start node in B and end node in A is maximized (here, the
weight of the arc from the node corresponding to item j to the node corresponding to
item l is given by s jl ). Even though we are not aware of any approximation algorithms
for the general version of this problem, several approximation algorithms exist for
the analogous problem in an undirected graph,n whichn results from the third variant
of the APP with objective function m k=1 j=1 l=1 s jl z jlk in the case where the
substitution profits are symmetric, i.e., s jl = sl j holds for all items j, l (in this case, the
weight of the (undirected) edge between the nodes corresponding to items j and l is
given by s jl = sl j ). This symmetric variant is a special case of the budgeted maximum
cut problem and admits a 1.42-approximation algorithm (Halperin and Zwick 2002).
Thus, by using this algorithm to approximate the third variant of the APP, we obtain
a 3.42-approximation algorithm for the S-APP in the case that the substitution profits
are symmetric.
5.2 Bounded substitution profits
Selling a substitution product can generate higher, identical, or lower revenue depend-
ing on how the substitute is priced compared to the first choice. Studies show that
customers prefer to buy a substitute in the same price range as originally intended or
not to buy at all (with the chance of buying higher- and lower-priced substitutes each
less than 3%, cf. Walter and Grabner 1975; similar results can be found in Emmel-
hainz et al. 1991). Thus, it is reasonable to assume that retailers cannot generate more
revenue (or profit) with selling a substitute than when offering the customers’ first
choice, as a portion will not buy any product at all. Empirical studies give evidence
that this assumption holds, i.e., revenue decreases when products are unavailable and
substitution occurs (see, for example, Anderson et al. 2006; Corsten and Gruen 2003;
Verbeke et al. 1998).1
Hence, in this subsection, we study the S-APP under the additional assumption that
s jl ≤ mink vlk holds for all products j, l. In this setting, the algorithm that chooses the
better of the two trivial solutions with profits V and W presented at the beginning of
Sect. 4 now yields a linear approximation in n: Using the fact that s jl ≤ mink vlk ≤ vlk
for all k ∈ {1, . . . , m}, setting Δ := maxl=1,...,n |{ j : s jl > 0}|, and denoting the set
of items in bin k (including the items in the common assortment) in a fixed optimal
solution by OPT(k), we can upper bound the profit obtained from substitution in this
1 Other studies show that an increase in sales can be generated by reducing (Boatwright and Nunes 2001;
Broniarczyk et al. 1998) or expanding (Baumol and Ide 1956; Oppewal and Koelemeijer 2005; Ton and
Raman 2010) assortment size. Yet, a decrease in assortment size can also lead to no changes in sales (Sloot
et al. 2006) or reduced sales (Borle et al. 2005), depending on how many items are currently offered in the
corresponding category. Therefore, as effects of assortment variety depend on specific problem parameters,
we do not consider these consequences in our stylized setting.
123
optimal solution by

m
m
s jl ≤ |{ j : s jl > 0}| · vlk
k=1 l∈OPT(k) j ∈OPT(k)
/ k=1 l∈OPT(k)
m
≤Δ· vlk ≤ Δ · V.
k=1 l∈OPT(k)
Hence, we can upper bound the total profit of the optimal solution as

m
OPT ≤ V + W + s jl
k=1 l∈OPT(k) j ∈OPT(k)
/
≤ (1 + Δ) · V + W ≤ (2 + Δ) · max{W, V },
which shows that the algorithm that chooses the better of the two trivial solutions with
profits V and W obtains an approximation ratio of (2 + Δ) in this case. As before, we
can strengthen the analysis for the case V
= W as shown in the following lemma:

Lemma 10 OPT ≤ 1 + min (1 + Δ) W V
,Δ + WV · max{W, V }.
Proof We have

V V
OPT ≤ (1 + Δ) + 1 · W ≤ (1 + Δ) + 1 · max{W, V } and
W W

W W
OPT ≤ (1 + Δ) + ·V ≤ 1+Δ+ · max{W, V }.
V V

Lemma 10 shows that the worst-case approximation ratio of (2 + Δ) can only occur if
W = V . Moreover, the proof of Lemma 10 shows that if W
V
→ 0, then max{V,W
OPT
} → 1.
However, if W V → 0, we only obtain that max{V,W } → Δ + 1. Thus, we can observe
OPT
some asymmetry between V and W under our assumption on the s jl . Observe that, in
case there are no substitution effects, i.e., Δ = 0, we obtain the worst-case factor of 2
from Sect. 4.
Note that, in theory, the value of Δ can be as large as n − 1. However, Δ = n − 1
would mean that there is at least one product l that can (at least partially) substitute
every other product, which is very unrealistic. Thus, it is natural to assume that Δ
n −1, i.e., a product can only substitute a relatively small fraction of the other products.
5.3 A greedy heuristic
Considering the results of the previous subsections, it seems much harder to develop
approximation algorithms for the S-APP than for the APP. Nevertheless, we now
123
provide a heuristic that can solve practical instances of the S-APP quickly with a
sufficient quality of approximation. To obtain this heuristic, we modify our greedy
heuristic for the APP (Algorithm 5) by also considering substitution profits. The
resulting algorithm is shown below (Algorithm 6).
Algorithm 6 Algorithm S-GREEDY

1: Let C = ∅ and, for each bin k, let Ik = ∅.
3: for j = 1 to n do
4: Let u j = v jk + l
= j sl j .
5: end for
6: while |Ik | < K and max j ∈I / k u j > 0 do
7: Add an item j ∈ / Ik with maximum value u j to Ik .
8: for j ∈ Sk do
9: Let u j = v jk − l∈Ik s jl + l∈Sk \{ j} sl j .
10: end for
11: end while
12: end for
14: Let Bk = Ik ∪ C.
15: end for
16: for j ∈
/ C do ⎛

17: u j = w j − v jk + sq j − s jp + max ⎝sl j + slp
l∈Ik
k: j∈Ik k: j ∈I
/ k q ∈B
/ k k: j ∈I
/ k p∈Bk k: j ∈I
/ k p∈Bk
⎞

− sql − vlk ⎠
q ∈B
/ k ∪{ j}
18: end for
19: if u j ≤ 0 for all j ∈
/ C or |C| = K then
20: return Common assortment C and local assortments Ik .
21: else
22: Add an item j ∈ / C with maximum value u j to the common assortment C, update the sets Ik
by removing
j fromIk if it is contained in Ik and removing an item l ∈ Ik with maximum value
sl j + p∈Bk slp − q ∈B / k ∪{ j} sql − vlk from Ik otherwise, update the sets Bk accordingly, and go
to line 16.
23: end if
The idea of the algorithm is again to first neglect the advantages of using the
common assortment and generate a local assortment for each store interdependently
of the others. These local assortments are computed in lines 2 to 12 of the algorithm,
where items are added greedily to each local assortment taking into account the local
profits as well as the substitution profits. Afterward, in each step, the algorithm chooses
an item that currently grants the largest gain in total profit when being added to the
common assortment and adds this item to the common assortment (while removing
either the item itself or the least profitable item from the local assortment of each
store).
At the end of the next section, we compare the solution quality and running time of
Algorithm 6 to the solution obtained by applying a state-of-the-art commercial solver
to the integer programming formulation of the S-APP provided at the beginning of
this section.
123
6 Experimental results
In this section, we present computational experiments on randomly generated instances

in order to analyze the benefits of the simultaneous selection of a common assortment
and local assortments compared to the two established strategies for specifying an
assortment where either every store has a customized assortment or the assortments
in all stores are the same. In particular, we are interested in how large the common
assortment profits w j must be in relation to the local profits v jk in order to see substan-
tial benefits from mixing common and local assortments. Moreover, we analyze the
performance of our greedy heuristic GREEDY from Sect. 4. In the experiments that
include substitution effects, we analyze the performance of the algorithm S-GREEDY
from Sect. 5.
For all generated instances, we compute the solutions for the two basic strategies
with the algorithms LOCAL and COMMON presented in Sect. 4. An optimal mixed
strategy using a common assortment that is supplemented by a specific local assortment
for each store is determined by using the first integer programming formulation of the
APP presented in Sect. 2 or, if substitution effects are considered, by using the integer
programming formulation of the S-APP presented in Sect. 5.
6.1 Generation of random instances
We randomly generate the values v jk and w j . Even though the cost of providing item j
might vary for different stores (e.g., due to different transportation costs), it seems to be
a reasonable assumption that the local profits v jk of a fixed item j in different stores k
are not completely independent. Therefore, we consider three scenarios where we
draw a value v j uniformly at random from [0, 1] independently for each item j and
then
– set v jk := v j for all k (total dependence), or
– independently for each store k, draw rk uniformly at random from [−0.5 p, 0.5 p]
and set v jk := max(0, v j + rk ), where p is a model parameter (intermediate
dependence), or
– draw all values v jk uniformly and independently from [0, 1] (total independence).
Observe that, for p = 0, the first and the second scenario are identical. In order
to generate the values w j , we draw a valueq j uniformly at random from [0.95, 1.05]
for each item j and set w j := q j · b · k v jk , where the factor b represents the
relative financial gains obtained when putting a product into the common assortment
(e.g., from economies of scale or recognition value). We consider 100 equidistant
values of b in [1, 2] and generate 100 instances for each of these values and four
different settings concerning the dependence of the values v jk (total independence,
intermediate dependence with p = 0.75 and p = 0.95, and total independence). The
optimal profit of our mixed strategy for each instance is determined by solving the
integer programming formulation of the APP presented in Sect. 2 using Gurobi 6.5.
We observe that, when b is too small or too large, algorithms COMMON and/or
LOCAL already yield close-to-optimal solutions (cf. Fig. 1). If a company estimates
123
OPT
Fig. 1 On the x-axis, the value b is shown. On the y-axis, the ratios COMMON OPT (crosses)
(stars) and LOCAL
are shown. We observe that, in the case of total independence shown here, min(COMMON, LOCAL) is
maximal when b ≈ 1.35
the value b in such a way, the assortment planning that we propose might not be neces-
sary. Therefore, for each of the four settings concerning dependence of the values v jk ,
we concentrate on three values of b that are neither too small nor too large. In the
setting of total independence, we consider b ∈ {1.2, 1.35, 1.5}, for p = 0.75 and
total dependence, we consider b ∈ {1.01, 1.05, 1.09}, and for p = 0.95, we consider
b ∈ {1.04, 1.09, 1.14}. Moreover, we consider two different instance sizes (small and
large), where (n, m, K ) = (1500, 50, 750) and (n, m, K ) = (50.000, 150, 25.000),
respectively. The small instances approximately represent the decision process for one
product category, while the large instances model the selection of a profit-maximizing
assortment for the whole store (i.e., products for all categories are selected simultane-
ously, neglecting minimum or maximum category sizes).
6.2 Setup of the experiments
All experiments were conducted on a standard desktop computer equipped with an Intel
Core i7-3770 processor and 16 GB of RAM. In order to compute optimal solutions,
we solved the integer programming formulations by using Gurobi 6.5.
We conducted experiments with and without substitution effects. In the first case, we
implemented the algorithms GREEDY, COMMON, and LOCAL, and compared their
solution quality with each other and with the optimal solution, which was obtained from
the first integer programming formulation presented in Sect. 2. However, only the small
instances could be solved by Gurobi within a time frame of 1 h. Indeed, Gurobi already
needs about 30 h on average to solve a single instance of size much smaller than the
large size (cf. Fig. 3). For the large instances with (n, m, K ) = (50.000, 150, 25.000),
Gurobi failed to produce solutions due to lack of memory. On the other hand, GREEDY
was able to solve such instances within a time frame of 1 h.
123
For the experiments with substitution, we additionally implemented the algo-

rithm S-GREEDY. In order to obtain optimal solutions, we solved the integer
programming formulation presented in Sect. 5. It turned out to be much harder for
Gurobi to solve instances of the S-APP than similarly sized instances of the APP.
Indeed, solving an instance of the S-APP of size (n, m, K ) = (80, 10, 40) with Gurobi
already takes more than 50 h on average. On the other hand, GREEDY was able to
solve instances of size roughly (n, m, K ) = (900, 50, 450) within a time frame of 1 h.
6.3 Results
In this section, we discuss the results of the computational experiments for the APP
and the S-APP.
6.3.1 Experimental results without substitution effects
In this subsection, we consider the basic APP without substitution effects. As shown
in the following, we identify several settings where our mixed strategy provides a
considerable increase in profit compared to the basic strategies given by COMMON
and LOCAL. Furthermore, GREEDY manages to find good solutions fast, even for
large instances that cannot be solved by Gurobi.
Comparison of COMMON and LOCAL
The two basic strategies in retail for specifying an assortment for multiple chain
stores are either one unified assortment for all stores (as computed by COMMON)
or customized assortments in all stores (as computed by LOCAL). One expects that,
depending on the extent of scaling effects such as recognition value or economies of
scale, the solution quality of COMMON and LOCAL will differ greatly. Indeed, this
effect can be observed in our experiments. The scaling effects are modeled by the
value b. The larger the value of b (which corresponds to the bonus value), the better
the solution produced by COMMON becomes in comparison with the one produced
by LOCAL. Figure 2 displays these results. On the x-axis, the value b is shown, and
on the y-axis, the ratio COMMON
LOCAL is shown. We observe that the ratio is monotonically
increasing with b.
Comparison of OPT and GREEDY
Since the APP is strongly NP-hard, we cannot expect GREEDY to always obtain
optimal solutions. However, for all considered instances, GREEDY obtains nearly
OPT
optimal solutions, i.e., the average ratio GREEDY of the profit of an optimal solution
and the profit of the solution returned by the algorithm is below 1.01.
The great benefit of GREEDY is its favorable running time compared to Gurobi.
For example, for instances of size (n, m, K ) = (15.000, 300, 7.500), Gurobi already
needs 30 h on average to solve a single instance, whereas GREEDY finishes in less
than ten minutes. For even larger instances, Gurobi often fails to produce solutions
123
Fig. 2 Comparison of the solution quality of COMMON and LOCAL for different values b. Increasing
the bonus value does not affect the value of the solution computed by LOCAL, but leads to a nearly linear
increase in the objective value of COMMON
Fig. 3 Comparison of the running times of Gurobi and GREEDY. We consider instances with m = 300
and K = n2 , where n is given on the x-axis. The running time is given in hours on the y-axis
due to lack of memory. GREEDY, on the other hand, finishes in less than 1 h even
when applied to instances of size (n, m, K ) = (50.000, 300, 25.000). Even on very
large instances of size (n, m, K ) = (150.000, 300, 75.000), GREEDY still finishes
in about 14 h.
We illustrate this effect in Fig. 3, where we used instances of size (n, m, K ) =
(n, 300, n2 ). On the x-axis, the number n of items is shown. On the y-axis, we plot the
running times of Gurobi and GREEDY in hours. While the running time of GREEDY
only increases very moderately with increasing number of items, the running time of
Gurobi reaches 30 h already for about 15.000 items.
123
Table 1 Average ratios COMMON OPT OPT (for small instances) as well as GREEDY and GREEDY
and LOCAL COMMON LOCAL
(for large instances) for different values b and different degrees of dependence of the values vjk
Value b : Small Medium Large Small Medium Large

Instance size: Small Small Small Large Large Large
v jk :
Total dependence 1.01|1.02 1.00|1.05 1.00|1.09 1.01|1.02 1.00|1.05 1.00|1.09

p = 0.75 1.06|1.01 1.04|1.03 1.03|1.06 1.06|1.01 1.04|1.03 1.03|1.06
p = 0.95 1.07|1.02 1.05|1.05 1.04|1.08 1.08|1.02 1.05|1.05 1.04|1.08
Total independence 1.18|1.01 1.09|1.05 1.04|1.11 1.20|1.01 1.10|1.03 1.05|1.09
Comparison of OPT, COMMON, and LOCAL
In order to illustrate the benefits of the simultaneous selection of a common assortment

and local assortments, we now compare the solutions produced by COMMON and
LOCAL to an optimal mixed strategy that supplements a common assortment by a
specific local assortment for each store.
Note that, since Gurobi is unable to compute solutions for the large instances, we
can directly perform this comparison only for the small instances. However, since the
OPT
ratio GREEDY is less than 1.01 for all instances that could be solved to optimality by
Gurobi, it is reasonable to instead compare the solutions produced by COMMON and
LOCAL to the solution of GREEDY for the large instances.
OPT OPT
Table 1 shows the average ratios COMMON and LOCAL for the small instances and
GREEDY GREEDY
the average ratios COMMON and LOCAL for the large instances. Here, we observe that
the relative gain in profit obtained from using a common assortment supplemented by
a specific local assortment for each store compared to one of the solutions produced
by the algorithms COMMON and LOCAL can be up to 20%. As one would expect,
algorithm COMMON provides good results when either the demand variation between
OPT GREEDY
stores is low or the bonus value is high. Thus, the ratios for COMMON and COMMON are
near one. In contrast, for high demand variation or low bonus values, LOCAL performs
well. Especially for intermediate dependence and a medium bonus, our approach
provides clearly better results than algorithms COMMON as well as LOCAL. Note
that GREEDY is always at least as good as both COMMON and LOCAL since it
incorporates these two basic strategies—and oftentimes even better.
Table 2 provides some objective values obtained by the different algorithms indi-
vidually in the case of total independence and medium-sized bonus value. Naturally,
the profit increases with increasing store capacity and increasing number of items.
Next, we regard mixed scenarios that are represented by a (4 × 3)-matrix, where
each row represents one of the four scenarios used to generate the values v jk (total
dependence, p = 0.75, p = 0.95, total independence), and each column corresponds
to a choice of the value b (small, medium, large). This provides us with twelve cat-
egories, and each entry of the matrix indicates the fraction of items that are drawn
from this category. Table 3 shows the results. For all scenarios, combining a common
assortment with local assortments leads to better results. In fact, for some scenarios,
123
Table 2 Average objective values obtained by the algorithms GREEDY, COMMON, and LOCAL for
medium-sized value b (b = 1.35) and total independence of the v jk . The instance size is (n, 150, n2 ) for
varying values of n. The displayed values are the average values over 100 instances
Number of items n 10.000 20.000 30.000 40.000 50.000
GREEDY 58.166 116.340 174.520 232.690 290.860

COMMON 52.862 105.730 158.600 211.470 264.330
LOCAL 56.245 112.500 168.750 224.990 281.240
our approach leads to a profit gain of up to 7% compared to the better of the two basic
strategies.
Managerial insights
Our results show that close-to-optimal mixed-strategy assortments can be computed

in reasonable time for instances of a realistic size. This enables the use of optimized
combinations of a common assortment and specific local assortments in practice.
Moreover, combining a common assortment with local assortments can lead to substan-
tial increases in profit compared to the basic strategies that use either one assortment
for all stores or a completely customized assortment for every store. As we include
both basic strategies, our approach is always at least as good as the current planning.
If management assumes intermediate or mixed demand dependencies between stores
or a bonus value in the range provided, it can generate even more profitable assort-
ments. The extend of the profit gains achievable seems to be relatively independent of
the store size and mainly depends on the size of the relative financial gains obtained
when putting a product into the common assortment (e.g., from economies of scale,
customer loyalty, recognition value, or coordinated advertisement) and on the degree
to which the profits of a single product differ from store to store.
6.3.2 Experimental results with substitution effects
We now present computational results of the assortment planning problem under sub-
stitution (S-APP) and analyze the solution quality and running time of our greedy
heuristic S-GREEDY. Here, we consider the case where the assortment of a prede-
fined category has to be determined, so substitution effects might be present between
any pair of items.
The common assortment profits and the local profits are generated as before. For
the substitution profits, we consider the following two settings:
j j
– Independently for each item j, we draw a vector r j = (r1 , . . . , rn ) uniformly at
j j
random from [0, 1]n such that i ri = 1. Then, we set s jl := rl · m1 · m k=1 v jk
for l
= j and s j j := 0.
– Independently for each item j, all values s jl for l
= j are drawn uniformly at
random from [0, 1] and s j j := 0.
123
OPT , OPT , GREEDY , and GREEDY for different mixed scenarios. Again, we consider small and large instances
Table 3 Ratios COMMON LOCAL COMMON LOCAL

OPT OPT
GREEDY GREEDY
Scenario Description
COMMON LOCAL COMMON LOCAL
⎛ ⎞
0 0 0
⎜0 0 0 ⎟
⎜ ⎟ Item profits are rather independent between 1.021.24 1.021.23
⎝0 0.25 0.25 ⎠
the stores and economies of scale apply
0 0.25 0.25
well
Assortment planning for multiple chain stores
⎛ ⎞
0 0 0
⎜ 0 0 0 ⎟
⎜ ⎟ Item profits are independent between the 1.041.11 1.051.09
⎝ 0 0 0 ⎠
stores and, for some items, economies of
0.33 0.33 0.33
scale apply
⎛ ⎞
0 0 0
⎜ 0.33 0 0⎟
⎜ ⎟ Item profits are rather dependent between the 1.071.09 1.071.09
⎝ 0.33 0 0⎠
stores and economies of scale do not apply
0.33 0 0
well
⎛ ⎞
0 0 0
⎜ 0.11 0.11 0.11 ⎟
⎜ ⎟ There is a mixture of dependent and 1.011.30 1.021.29
⎝ 0.11 0.11 0.11 ⎠
independent item profits and, for some
0.11 0.11 0.11
items, economies of scale apply
123
905
The motivation for the first setting was given in detail in Sect. 5 and mainly originates
from studies showing that customers prefer to buy a substitute in the same price range
j
as originally intended. Note that the value r j corresponds to the portion of customers
who do not buy a substitute for product j.
After generating the substitution profits as above, we scale all substitution profits
by a factor μ ≥ 0 that expresses the extent of substitution. A large parameter value μ
represents a setting in which substitution is important, whereas a value close to 0
indicates that substitution effects are negligible. We found that values of μ around 0.3
generate the most interesting settings since, otherwise, the substitution effects are
either too overwhelming or nonexistent.
Comparison of OPT and S-GREEDY
Our first observation is that, when using the integer programming formulations pre-
sented earlier, Gurobi needs much more time for solving instances of the S-APP than
it needs for solving similarly sized instances of the APP. For example, solving an
instance of the S-APP of size (n, m, K ) = (80, 10, 40) to optimality with Gurobi
already takes more than 50 h on average, which means that applying Gurobi to the
integer programming formulation is impractical for instances of a realistic size. In
contrast, our greedy heuristic S-GREEDY only needs seconds to finish with reason-
able solution quality on instances of size (n, m, K ) = (80, 10, 40). In summary, for
the small instances for which we can compare Gurobi and our greedy heuristic, we
found that the solution produced by the heuristic was on average within 3% of the
OPT
optimum. In Table 4, the column “Small instances” shows the ratio S-GREEDY for
different settings. The instance size is (n, m, K ) = (15, 5, 8), and we regard the case
of medium-sized bonus values, different degrees of independence, and values 0.2 and
0.4 for μ.
For the APP, GREEDY is able to solve instances of size roughly (n, m, K ) =
(50.000, 150, 25.000) in about an hour. In this time frame, S-GREEDY can only
solve instances of the S-APP of size about (n, m, K ) = (900, 50, 450). However,
since we consider the optimization of the assortment for a predefined category here,
considering instances with about 900 items might be sufficient for solving real-world
instances in this case.
Comparison of S-GREEDY, COMMON, and LOCAL
To illustrate the benefits of combining a common assortment with customized local

assortments that vary from store to store, we again compare the solutions produced by
COMMON and LOCAL to an optimized mixed strategy. However, in the presence of
substitution effects, Gurobi is unable to produce solutions for instances of any relevant
size. Hence, we now compare the solutions produced by COMMON and LOCAL to
the solution of S-GREEDY for instances of size (n, m, K ) = (200, 30, 100), which
are referred to as large instances in Table 4. As expected, the basic strategies perform
much worse in the presence of substitution effects than for the regular APP.
123
Table 4 We examine the solution quality of Algorithm S-GREEDY. The size of the small instances is
(n, m, K ) = (15, 5, 8), and the size of the large instances is (n, m, K ) = (200, 30, 100). The substitution
j
profits are drawn as described in the first setting (s jl = μ · rl · m1 · m k=1 v jk ). We display the results
for medium-sized bonus values as defined in Sect. 6.1. The results for the other settings (as in Table 1)
are similar. The ratios given are the average over 100 runs. The maximum over these 100 runs is given in
brackets
Dependence of v jk Value b μ·m Small instances Large instances

OPT S-GREEDY S-GREEDY
S-GREEDY COMMON LOCAL
Total dependence 1.05 1 1.01(1.05) 1.81(1.90) 1.90(1.99)
p = 0.75 1.05 1 1.00(1.03) 2.02(2.09) 2.00(2.05)
p = 0.95 1.09 1 1.00(1.01) 2.00(2.07) 1.99(2.04)
Total independence 1.35 1 1.00(1.01) 2.20(2.23) 2.16(2.17)
Total dependence 1.05 2 1.03(1.17) 3.36(3.49) 3.53(3.57)
p = 0.75 1.05 2 1.01(1.11) 3.55(3.64) 3.52(3.61)
p = 0.95 1.09 2 1.01(1.03) 3.47(3.54) 3.46(3.52)
Total independence 1.35 2 1.00(1.03) 3.41(3.47) 3.34(3.38)
Managerial insights
In the presence of substitution effects, the size of the instances for which close-
to-optimal solutions can be quickly obtained decreases drastically compared to the
setting without substitution effects. However, since substitution usually takes place
only within one category, the presented algorithms are still sufficient for solving
instances of a realistic size. Moreover, the financial gains obtained from using an
optimized mixed-strategy assortment are even higher in the presence of substitution,
which provides an even stronger motivation for using an assortment planning as pro-
posed here. These financial gains now depend not only on the bonus obtained when
putting a product into the common assortment and the degree to which the profits of
a single product differ from store to store, but also on the substitution probabilities. If
a retailer finds itself in one of the settings analyzed above, our approach is especially
appealing.
7 Conclusion and future research
We have formulated a stylized model for the regionalized assortment planning problem
(APP) and have shown that solving the APP can lead to significant profit gains for
a retailer with multiple chain stores. However, it seems computationally impossible
to solve its natural integer programming formulation for instances of a realistic size.
Therefore, we proposed a greedy heuristic that computes close-to-optimal solutions
efficiently within a reasonable amount of time. We believe that our findings are also
interesting from a practical point of view, i.e., from a managerial perspective since our
methods can be used to optimize the assortment planning and, thus, generate more
total profit.
123
Additionally, we provided several approximation algorithms for the APP including

a 23 -approximation algorithm. Concerning the computational complexity of the prob-
lem, we showed that the APP is strongly NP-hard but can be solved in polynomial
time if the number of stores is fixed. Moreover, we showed how substitution effects
can be included in our model.
In future research, we plan to consider extensions of the model such as sub-regions
(where additional profit is obtained when a product is placed in all stores within a
predefined set) or individual item sizes (i.e., loosening the same size assumption). To
do this, it is necessary to include shelf space planning into our model and decide on
how many individual-sized facings to offer of every product in the assortment. It could
also be worthwhile to consider the simpler variant of a fixed number of different item
sizes representing partly standardized facings. Moreover, extensions such as pricing
or inventory decisions under limited shelf space could be included in our regionalized
assortment planning model.
Appendix
We present the proof of Theorem 3, which states that the APP can be solved in
m
time O(n 2 + m − 1).
Proof Consider an instance of the APP with m stores of size K and n products. In the
following, we assume that, in every optimal packing, the common assortment has size
strictly less than K . This can be done since, if it has size exactly K , we can simply
pack the K items with the largest values w j into the common assortment to obtain an
optimal solution.
We start by sorting, for each bin k individually, the profits v jk for all j in non-
increasing order to obtain m vectors ν 1 , . . . , ν m of length n. For each bin, our first
packing consists of the first K items with respect to theordering
K obtained for this bin.
The objective value of this solution is given by V := m k=1 j=1 (ν ) j .
k
Next, we “guess” the least valuable item lk in each bin k of a fixed optimal solution
that is not part of the common assortment. To do so, we enumerate over all possible
combinations of these items in all bins, which are n m possible combinations. Thus, we
assume in the following that we know the least valuable item lk that is not part of the
common assortment in each bin k of a fixed optimal solution. Let ek be the index of
this item with respect to the ordering of the items by nonincreasing profits v jk in bin k.
Without loss of generality, we can assume that ek ≤ K since, otherwise, lk could be
substituted in the optimal solution by an item with larger local profit in bin k, at least
one of which is not part of the common assortment of the optimal solution. Moreover,
we can assume that no item with index larger than ek with respect to the ordering of
bin k is part of the individual packing of bin k in the optimal solution since lk is the
least valuable such item in bin k and items with index larger than ek provide no more
individual profit than lk . A consequence is that, in the optimal packing, all items with
index smaller or equal to ek are packed (either into the common assortment or into
the individual assortment of bin k). Moreover, if there is an item j such that, for each
bin k, its index with respect to the ordering of bin k is smaller than ek , then this item
123
must be part of the common assortment of the optimal solution. We denote the set of
these items with index smaller than ek in the ordering of each bin k by J.
Now, we still have to decide which items to pack into the common assortment.
Here, we can now restrict our search to the items that have an index larger than ek in
at least one bin k since the other items (which are contained in J ) are automatically
part of the common assortment. Let Sk be the set of the n − ek items with index larger
than ek in the ordering of bin k. Thus, our remaining task is to decide which items
from S := ∪k Sk should be packed into the common assortment such that, for each k,
exactly K − ek of these items belong to Sk and the total profit obtained by this process
is maximized (no item with index smaller than ek has to be removed from the packing
during this process). We can formulate this problem as the following integer program:

maximize ωjxj
j∈S

subject to x j = K − ek ∀k ∈ {1, . . . , m}
j∈Sk
x j ∈ {0, 1} ∀j ∈ S
where

xj =
0, else
and ω j is the additional profit obtained by packing item j into the common assortment.
More formally, for an item j that is not one of the least valuable items l1 , . . . , lm in
the bins of the optimal solution, we let T j denote the set of bins wherethe index of j
in the corresponding ordering is smaller than ek . Then, ω j = w j − k∈T j v jk . If j
is among the least valuable items l1 , . . . , lm in the bins of the optimal solution that
we guessed in the first step, we set ω j := 0 since we know that j is not part of the
common assortment in the optimal solution.
We solve the integer program given above by clever enumeration of a suitable
subset of the feasible solutions. For each nonempty subset L of the bins, let S(L)
denote the set of all items that are contained exactly in those sets Sk corresponding
to the bins k ∈ L. Then, by definition, every item in S belongs to exactly one of the
sets S(L).
The important observation now is that, if we decide to choose a certain number h of
items from some set S(L), the objective value is always maximized by choosing the h
items with the largest values ω j (breaking ties arbitrarily). Thus, for each nonempty
subset L of the bins (of which there are 2m − 1 many), we only have to choose a
number of items from S(L) to pack into the common assortment, which results in
n 2 −1 potential solutions of the integer program. For each of these potential solutions,
m
we can check in constant time whether it is feasible (i.e., whether it satisfies the m
constraints of the integer program) and then select the best solution among the feasible
ones in order to obtain an optimal solution.
123
The running time of the algorithm is dominated by two steps: The “guessing” of the
least valuable items takes time O(n m ) and solving the integer program by the method
described above takes time O(n 2 −1 ), which yields the claimed running time.
m

Remark 1 Unless P = NP, there does not exist an algorithm that solves the integer
program in the proof of Theorem 3 in polynomial time in n and m. This can be shown
by using a reduction from the independent set problem: Given a graph G = (V, E),
we construct a set Se = {xu , xv , q} for each edge e = (u, v) in G, where q is a unique
dummy item for each set. Further, we set ωu = ωv = 1, ωq = 0, and K − ek = 1.
Then, it is easy to see that an independent set of size p corresponds to a solution of
value p of the integer program and vice versa.
References
Aastrup J, Kotzab H (2010) Forty years of out-of-stock research—and shelves are still empty. Int Rev Retail
Distrib Consum Res 20:147–164
Anderson ET, Fitzsimons GJ, Simester D (2006) Measuring and mitigating the costs of stockouts. Manag
Sci 52:1751–1763
Baumol WJ, Ide EA (1956) Variety in retailing. Manag Sci 3:93–101
Boatwright P, Nunes JC (2001) Reducing assortment: an attribute-based approach. J Mark 65:50–63
Borin N, Farris PW (1995) A sensitivity analysis of retailer shelf management models. J Retail 71:153–171
Borin N, Farris PW, Freeland R (1994) A model for determining retail product category assortment and
shelf space allocation. Decis Sci 25:359–384
Borle S et al (2005) The effect of product assortment changes on customer retention. Mark Sci 24:616–622
Broniarczyk SM, Hoyer WD, McAlister L (1998) Consumers’ perception of the assortment offered in a
grocery category: the impact of item reduction. J Mark Res 35:166–176
Cadeaux JM (1999) Category size and assortment in US macro supermarkets. Int Rev Retail Distrib Consum
Res 9:367–377
Campo K et al (2004) Dynamics in consumer response to product unavailability: do stock-out reactions
signal response to permanent assortment reductions? J Bus Res 57:834–843
Chen M-C, Lin C-P (2007) A data mining approach to product assortment and shelf space allocation. Expert
Syst Appl 32:979–986
Chong J-K, Ho T-H, Tang CS (2001) A modeling framework for category assortment planning. Manuf Serv
Oper Manag 3:191–210
Corsten D, Gruen T (2003) Desperately seeking shelf availability: an examination of the extent, the causes,
and the efforts to address retail out-of-stocks. Int J Retail Distrib Manag 31:605–617
Corsten H, Kasper B (2016) Entwurf mehrstufiger Sortimentsplanungsmodelle in Handelsunternehmungen.
Schr. zum Produktionsmanage. 96, Kaiserslautern
Corsten H, Hopf M, Kasper B, Thielen C (2017) Regionalized assortment planning for multiple chain stores.
Operations Research Proceedings 2016 (forthcoming, 2017)
Emmelhainz MA, Stock JR, Emmelhainz LW (1991) Consumer response to retail stock-outs. J Retail
67:138–148
Farias VF, Jagabathula S, Shah D (2013) A nonparametric approach to modeling choice with limited data.
Manag Sci 59:305–322
Fisher M, Vaidyanathan R (2014) A demand estimation procedure for retail assortment optimization with
results from implementations. Manag Sci 60:2401–2415
Gaur V, Honhon D (2006) Assortment planning and inventory decisions under a locational choice model.
Manag Sci 52:1528–1543
Grewal D, Levy M, Mehrotra A, Sharma A (1999) Planning merchandising decisions to account for regional
and product assortment differences. J Retail 75:405–424
Halperin E, Zwick U (2002) A unified framework for obtaining improved approximation algorithms for
maximum graph bisection problems. Random Struct Algorithms 20:382–402
123
Hingley MK (2005) Power to all our friends? Living with imbalance in supplier–retailer relationships. Ind
Mark Manag 34:848–858
Honhon D, Gaur V, Seshadri S (2010) Assortment planning and inventory decisions under stockout-based
substitution. Oper Res 58:1364–1379
Hübner AH, Kuhn H (2012) Retail category management: state-of-the-art review of quantitative research and
software applications in assortment and shelf space management. Omega Int J Manag Sci 40:199–209
Hübner AH, Kuhn H, Kühn S (2016) An efficient algorithm for capacitated assortment planning with
stochastic demand and substitution. Eur J Oper Res 250:505–520
Huh YE, Vosgerau J, Morewedge CK (2016) More Similar but less satisfying. Comparing preferences for
and the efficacy of within-and cross-category substitutes for food. Psychol Sci 27(6):894–903
Hwang M, Bronnenberg BJ, Thomadsen R (2010) An empirical analysis of assortment similarities across
US supermarkets. Mark Sci 29:858–879
Kök AG (2003) Management of product variety in retail operations. Pennsylvania
Kök AG, Fisher ML (2007) Demand estimation and assortment optimization under substitution: method-
ology and application. Oper Res 55:1001–1021
Kök AG, Fisher ML, Vaidyanathan R (2009) Assortment planning: review of literature and industry practice.
In: Agrawal N, Smith SA (eds) Retail Supply Chain Manag, vol 223. Springer, USA, pp 99–153
Li G, Rusmevichientong P (2014) A greedy algorithm for the two-level nested logit model. Oper Res Lett
42:319–324
Mantrala MK, Levy M, Kahn BE, Fox EJ, Gaidarev P, Dankworth B, Shah D (2009) Why is assortment
planning so difficult for retailers? A framework and research agenda. J Retail 85:71–83
Nemhauser GL, Wolsey LA (1988) Integer and combinatorial optimization. Wiley-Interscience, New York
Oppewal H, Koelemeijer K (2005) More choice is better: effects of assortment size and composition on
assortment evaluation. Int J Res Mark 22:45–60
Papdimitriou C, Yannakakis M (1991) Optimization, approximation, and complexity classes. J Comput
Syst Sci 43:425–440
Porter SS, Claycomb C (1997) The influence of brand recognition on retail store image. J Prod Brand Manag
6:373–387
Rajaram K (2001) Assortment planning in fashion retailing: methodology, application an analysis. Eur J
Oper Res 129:186–208
Rooderkerk RP, Heerde HJv (2016) Robust optimization of the 0–1 knapsack problem: balancing risk and
return in assortment optimization. Eur J Oper Res 250:842–854
Rooderkerk RP, van Heerde HJ, Bijmolt THA (2013) Optimizing retail assortments. Mark Sci 32:699–715
Rusmevichientong A, Shmoys D, Tong C, Topaloglu H (2014) Assortment optimization under the multi-
nomial logit model with random choice parameters. Prod Oper Manag 23:2023–2039
Shin H, Park S, Lee E, Benton WC (2015) A classification of the literature on the planning of substitutable
products. Eur J Oper Res 246:686–699
Sloot LM, Verhoef PC, Franses PH (2005) The impact of brand equity and the hedonic level of products on
consumer stock-out reactions. J Retail 81:15–34
Sloot LM, Fok D, Verhoef PC (2006) The short- and long-term impact of an assortment reduction on
category sales. J Mark Res 43:536–548
Smith SA, Agrawal N (2000) Management of multi-item retail inventory systems with demand substitution.
Oper Res 48:50–64
Stigler GJ (1958) The economies of scale. J Law Econ 1:54–71
Syring D (2004) Bestimmung effizienter Sortimente in der operativen Sortimentsplanung. PhD thesis, Freie
Universität Berlin, Berlin
Talluri K, Ryzin Gv (2004) Revenue management under a general discrete choice model of consumer
behavior. Manag Sci 50:15–33
Ton Z, Raman A (2010) The effect of product variety and inventory levels on retail store sales: a longitudinal
study. Prod Oper Manag 19:546–560
Urban TL (1998) An inventory-theoretic approach to product assortment and shelf-space allocation. J Retail
74:15–35
Verbeke W, Farris P, Thurik R (1998) Consumer response to the preferred brand out-of-stock situation. Eur
J Mark 32:1008–1028
Vulcano G, van Ryzin G, Ratliff R (2012) Estimating primary demand for substitutable products from sales
transaction data. Oper Res 60:313–334
Walter C, Grabner JR (1975) Stockout cost models: empirical tests in a retail situation. J. Mark 39:56–60
123
Wang R (2012) Capacitated assortment and price optimization under the multinomial logit model. Oper
Res Lett 40:492–497
Zinn W, Liu PC (2001) Consumer response to retail stockouts. J Bus Log 22:49–71
123

Corsten2018 Article AssortmentPlanningForMultipleC

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Corsten2018 Article AssortmentPlanningForMultipleC

Uploaded by

Copyright:

Available Formats

OR Spectrum (2018) 40:875–912

Assortment planning for multiple chain stores

Hans Corsten2 · Michael Hopf1 · Benedikt Kasper2 ·

Abstract In retail, assortment planning refers to selecting a subset of products to offer

1 Department of Mathematics, University of Kaiserslautern, Paul-Ehrlich-Str. 14, 67663

Keywords Retail management · Computational complexity · Approximation

In retail, assortment planning refers to selecting a subset of products to offer in each

1.1 Previous work

The assortment planning problem (APP) is considered in both operations research

1.1.1 Demand estimation

1.1.2 Single store assortment planning

1.1.3 Regionalization and store-level assortment optimization

Regionalization and store-level assortment optimization lead to competitive advan-

An important issue in assortment planning is substitution. If a product is not available,

1.2 Our contribution

solved in polynomial time. Additionally, we analyze how substitution effects can be

2 Formulation of the APP as an integer program

2.1 Integer programming formulations

3 Computational complexity of the APP

Theorem 2 The APP is strongly NP-hard.

vz,Ak := 0 ∀k vz,Bk := 0 ∀k vz,Z := N

Example 1 Consider the case that m = 3, n = 3, and K = 2. Let a > 0 be some

Theorem 3 directly yields the following corollary:

4.1 A 2-approximation algorithm

 value of a given instance of the APP

W ≥ W L + (K − L)(1 − )ω and

where the second inequality follows since m · v jk ≤ ω for all j, k.

4.2 An approximation in terms of the size of the common assortment

We now present an algorithm that roughly guarantees an approximation factor of

Algorithm 1 Algorithm GUESS AND PACK

Proof If L ∗ ≤ a, Algorithm 1 returns an optimal solution. Hence, we may assume for

where we used that a ≤ L ∗ ≤ K and a = 2 + 2 . Since OPT = B ∗ + V ∗ , this shows

Lemma 5 For , γ > 0, applying Algorithm 1 to the instance

4.3 A 23 -approximation algorithm

In the following, we develop a polynomial-time algorithm that achieves an approxi-

OPT = V ∗ + B ∗ = V1∗ + V2∗ + B ∗ .

Algorithm 2 Algorithm LOCAL

Lemma 6 Algorithm 2 achieves an objective value of at least V ∗ and can be imple-

Algorithm 3 Algorithm COMMON

Lemma 7 Algorithm 3 achieves an objective value of at least V2∗ + B ∗ and can be

Algorithm 4 Algorithm BONUS

Lemma 8 Algorithm 4 achieves an objective value of at least V1∗ + B ∗ and can be

Proof If C ∗ = ∅, then V2∗ = B ∗ = 0 and we are in the case of Algorithm 2. Otherwise,

Theorem 4 If LOCAL, COMMON, and BONUS denote the objective values

Proof If ALG := max{LOCAL, COMMON, BONUS} denotes the objective value

4.4 A greedy heuristic

Algorithm 5 Algorithm GREEDY

4.5 Generalization to different store capacities

5 Modeling substitution effects

Proposition 1 The S-APP is APX-hard.

5.1 Some approximation results

In order to obtain a polynomial-time constant-factor approximation algorithm for

OPT ≤ V + W + S ≤ 3 · max{V, W, S}.

5.2 Bounded substitution profits

5.3 A greedy heuristic

Algorithm 6 Algorithm S-GREEDY

In this section, we present computational experiments on randomly generated instances

6.1 Generation of random instances

6.2 Setup of the experiments

For the experiments with substitution, we additionally implemented the algo-

6.3.1 Experimental results without substitution effects

Comparison of COMMON and LOCAL

Comparison of OPT and GREEDY

Value b : Small Medium Large Small Medium Large

value of a given instance of the APP

Lemma 5 For , γ > 0, applying Algorithm 1 to the instance