DecisionPower Inc. MarketSim
by David
M. Raab
DM News
October, 2006
Agent-based modeling uses computer programs to
simulate quasi-independent entities as they interact with their environment and
each other. It is frequently used to
model things like ecosystems and network traffic, although it is probably most
familiar from computer games such as SimCity.
In an agent-based model, the actions of each entity (agent) are
determined by rules that process inputs including the agent’s own state
(“hungry”), external factors (“food on that rock”), and other agents (“crowd
around the food”). Results over time can
produce complex patterns of behavior that are not explicitly defined in the
rules themselves. In the example I just
gave, the creature might balance how hungry it is, the distance to the food,
and the size of the crowd to decide whether to move towards or away from the
rock.
MarketSim (DecisionPower Inc.,
408-379-9200, www.decisionpower.com)
applies agent-based modeling to consumer behavior. In particular, it is used to predict results
in a product category such as snack foods or headache remedies. In a model of this sort, the primary agents
are consumers, while the environment includes brands, products and
channels. Rules describe how consumers
make purchase decisions, taking into account factors such as product
attributes, consumer preferences, marketing activities, distribution, and
external influences like the weather. When
the model is run with specific inputs such as particular set of products and
marketing activities, the system simulates what consumers will buy over time
and generates outputs for market share, revenue, profits, unit sales, and other
business measures.
This may sound quite simple, but in practice it is
not. MarketSim’s
must match the results of actual consumer markets with enough accuracy for
businesses to determine the likely results of a particular marketing plan or
product launch. This means models all
the competitors in a product category, as well as the different distribution
channels and types of customers. Each
has its own characteristics.
The most complex entity is the consumer. MarketSim provides
over 100 prebuilt rules, of which 25 or so might be
used in any particular model. These
rules describe how customers make their purchase decisions, taking into account
their preferences for different product attributes; information gathered from
advertising and personal contacts; consumption volume and frequency; shopping
behavior such as channels used and responsiveness to in-store displays and
price differences; responsiveness to coupons; and previous experiences. Each model includes multiple customer
segments with their own settings for rules such as the weight assigned to different
product attributes.
The number of agents assigned to a segment would be
proportionate to the size of that the segment in the actual marketplace,
although the total number of agents need not equal the number of actual
customers. A market of many millions could
be accurately modeled with 100,000 to 150,000 agents.
A model would also include dozens or even hundreds of
specific products, depending on the complexity of the marketplace and the
degree of detail required. Products are
linked to brands. Each brand will have
its own attributes, which use the same categories as consumer preferences. This provides the connection between brands
and consumers that is needed to model consumer choice. Each brand also has a marketing plan with
price, display, distribution, media, and coupon details for each time period.
All these rules and attributes must be set so they result
in an accurate prediction. DecisionPower does this by gathering historical
data—typically three year’s worth—for actual sales, marketing activities,
product attributes, distribution, and external factors. This data covers all competitors, not just
the model sponsor. It is fed into the
model and the rules are tuned until the system gives acceptably realistic
results. Modelers usually feed in two
and a half years of actual data and then compare simulated results for the
final six months with the known actual results for the same period.
This calibration process occupies most of the three
to four months it takes to build a major MarketSim
model. Most of the work involves adding
new data sources or events to reduce anomalies in results. A calibrated model can give reliable
predictions about one year into the future, and will remain valid—assuming the
inputs such advertising spend are updated—for one or two years before it must
be rebuilt.
Constructing a MarketSim
model is largely a task for experts, either employed by DecisionPower
or in the research department of a client.
The system offers a graphic user interface that is more than adequate
for such purposes. A separate module
called BrandManager lets non-technical users specify
model inputs, such as alternative marketing plans, on an Excel
spreadsheet. BrandManager
then imports the spreadsheet, loads the data into MarketSim
proper, and runs a model with the new assumptions. Users can save their inputs as scenarios,
allowing them to easily compare different options. They can view results on the screen, selecting
the scenarios, products and measures to compare, and displaying them as a table
or graph. Results can also be exported
back to Excel for further analysis.
MarketSim was introduced in
1996 and has been the primary focus of DecisionPower’s
business since 2001. The end-user
portion of the system runs on a Windows PC, while data—which can be quite
voluminous because of the all the historical detail—is typically stored on a
central server. Small models can run on
the end-user system but larger ones are often sent to a separate modeling
server. DecisionPower
offers hosted options for clients that prefer not to install the system
in-house. Running a single scenario may
takes about ten minutes for a simple model and 45 minutes to an hour for a very
complicated one. Companies often run
multiple scenarios—sometimes hundreds—as they explore alternatives to identify
optimal marketing strategies.
DecisionPower charges from
$100,000 to $300,000 to develop a MarketSim model,
depending on the project. Clients can
license the completed model to run in-house for $100,000 per year. DecisionPower has
sold 60 to 80 MarketSim models to date, mostly to
consumer package goods manufacturers. A lower cost option, MarketSimExpress,
is available for quicker, simpler projects.
* * *
David Raab is president of
Client X Client, a marketing technology and consulting firm specializing in
helping companies understand and improve how they treat their customers. He based in