

Strategic Retail Location
LocationAllocation Modelling
Briefly, locationallocation has two
components. The location
model systematically examines
various combinations of retail locations, while the allocation
model
predicts human behaviour and allocates demand to the most likely supply
point, i.e. it assigns a shopper to one or more stores.
Integrating
the location and allocation models, we generate a scenario, evaluate
its
effectiveness, modify and reevaluate it hundreds or thousands of
times,
until we find a suitable configuration of stores. In effect it is
a trialanderror simulation of a large number of store location
scenarios.
The location of
competing stores can be fed into the system. The corporate
objective
may be to avoid competition, or to go headtohead against it—these
and
other strategies can be examined. Assuming that the models are
properly
constructed and calibrated, this is a far less risky method of
selecting
a store location than to
openinthefirstavailableshoppingcentreandpray.
Discrete and Continuous Space
There are two modes of locationallocation
modelling: continuous space and discrete space. In the continuous
space implementation, a store is free to roam anywhere in the study
area; the algorithm pushes the store incrementally over the map until
it
finds the optimal position. Depending on the data, this just might be
an
unlikely location: on a railway line, in a lake
or
atop the CN Tower. GIS procedures can constrain
the
domain of wandering to acceptable locations. In the discrete
space
implementation we begin with a set of candidate locations, e.g.
greensites
available for purchase, and select the location best suited to our
purpose.
Discrete space mode is usually the more practical
option for
urbanscale applications,
while continuous space mode is better suited to macrolevel
analysis, or when going into a market with no clear idea where to look.
Technical Detail
Suppose we already have
10 stores, and there are 5 greensite locations from which we wish to
select
2, for a total of 12 stores. Let's call the 15 sites A thru
O.
A thru J are the 10 existing stores. Initially we select K and
L.
This can be represented as follows, with black letter representing
existing outlets, red letters representing the
selected sites, and green letters the unselected candidates:
A
B C D E F G H
I J 
K
L
M N O 
< Existing
stores >

< Picked the first two
>

This is the location step:
we have located stores at K and L. Now the allocation
step is
to assign demand to these stores. Often we assume that shoppers
go
to the nearest store, but in reality shopper behaviour is more complex
and we have to use some type of gravity model. Assume that on
the basis of some rule we assign all
demand
to at least one store. Then the above configuration can be
evaluated
and associated with some measure of effectiveness (e.g. the total
distance
travelled by all shoppers, or our market share or profitability), which
we call the objective function.
Similarly we evaluate the objective function for other configurations:
A B C D E
F G H I J 
K
L M N O 
A B C D E
F G H I J 
K
L M N O 
A B C D E
F G H I J 
K
L M N O 
etc
The configuration with the
best objective function is the winner.
This seems easy enough conceptually. When applied
to a practical retail problem there are several technical aspects
that require training and experience:
 Determining the objective function and
optimization
criterion: is it to minimize distance travelled, to minimize the number
of people beyond a certain reach (say 50 km), to maximize market share,
to maximize sales, to maximize profit, to batter the competition in the
short term? Each criterion probably produces a different optimal
configuration.
 Modelling demand and allocation. The outcome
is only as good as these steps.
 The method of evaluating all
those combinations—choosing 2 from 5 is not a big problem, but when
you
consider larger networks of stores, the competition, or the option of
closing
some existing stores, the number of combinations can be in the
millions, and
can challenge even today's fast computers.
 The quality of data that feed
the problem, e.g. the positional accuracy of geographic data (street
networks)
and the sensitivity of the outcome to data quality.
