Professional Documents
Culture Documents
OWL RESEARCH v
Prepared
by:
Geoffrey
Knight
Prepared
for:
Lucky
Duck
Entertainment
Date:
February
20,
2013
Executive
Summary
Lucky
Duck
Entertainment
(LDE)
is
faced
with
a
problem
of
maximizing
the
profitability
of
slot
machines
at
each
of
their
eight
Nevada
locations.
The
number
of
possible
machine
mixes
is
too
great
to
rely
on
ad-hoc
optimization.
For
that
reason,
Owl
Research
(OR)
proposes
an
analytics
approach
to
this
problem.
First,
all
machines
will
be
grouped
by
profit
level
and
demand.
This
will
make
the
problem
more
tractable,
by
reducing
the
number
of
possible
combinations.
Second,
an
Excel
based
forecasting
tool
will
be
developed,
to
predict
the
performance
of
a
given
machine
mix.
This
is
essential
for
comparing
any
proposed
changes
to
the
performance
of
the
current
mix.
Finally,
a
mathematical
optimization
problem
will
be
formulated
and
solved
to
find
the
optimal
machine
mix,
given
capacity
and
management
constraints.
After
the
optimal
mix
is
verified
with
the
forecasting
tool,
it
will
provide
a
clear
implementation
strategy
for
each
casino
manager.
The
rest
of
this
document
provides
high-level
description
of
the
projects
methodology,
data
and
other
requirements.
For
this
project,
a
number
of
analytical
tools
will
be
used.
In
particular,
to
categorize
the
machines,
a
large
amount
of
data
will
need
to
be
processed.
SAS
will
be
used
for
this
purpose,
as
it
provides
a
platform
to
format
the
data
and
to
perform
the
necessary
analysis.
Owl
Research
has
a
fully
licensed
version
of
SAS
installed
on
an
in-house
server.
No
investment
by
LDE
is
required
at
this
stage.
Monte
Carlo
Simulation
will
be
employed
in
forecasting.
A
combination
of
VBA
and
Microsoft
Excel
will
be
used
for
this
purpose.
Since
VBA
is
built
into
Excel,
LDE
simply
requires
a
recent,
licensed
copy
of
Microsoft
Office.
Finally,
depending
on
the
size
of
the
final
problem,
either
Open
Solver
or
AMPL
will
be
used.
OR
has
both
of
these
installed
locally.
AMPL
and
its
packages
are
available
for
commercial
use
for
a
fee.
The
necessity
of
AMPL
will
become
evident
after
machine
categorization.
This
section
contains
information
regarding
the
resources
required
to
complete
the
project,
the
final
deliverables,
and
the
overview
of
the
analytical
approach.
2.1
Information
Flow
To
provide
final
recommendations
to
LDE,
Owl
Research
will
require
accurate
data.
First,
available
historical
transaction
records,
going
back
at
least
one
year,
will
be
necessary
to
complete
categorization
and
forecasting.
Separate
data
sets
for
overall
and
cardholder-
only
transactions
will
benefit
the
analysis.
At
the
optimization
stage,
approximate
data
regarding
cost
of
installation
and
decommissioning
of
each
machine
will
be
necessary.
Ms.
de
la
Luz
shall
be
the
primary
point
of
contact
to
provide
technical
data
to
OR.
Lastly,
since
management
preferences
play
an
important
role
in
casinos
performance,
optimization
cannot
be
completed
without
it.
A
survey
will
be
sufficient
to
gather
these
preferences.
This
survey
will
be
compiled
by
OR
and
administered
by
LDE.
OR
will
collect
back
the
survey,
organize
the
data,
and
use
it
for
modeling.
2.2
Analytical
Component
This
project
is
broken
down
into
three
parts:
machine
categorization,
performance
forecasting,
and
machine
mix
optimization.
All
three
parts
of
the
project
depend
on
one
another,
and
will
be
completed
in
the
following
order.
2.2.1
Machine
Categorization
Currently
there
are
hundreds
of
machine
types,
each
to
one
of
six
denominations,
among
four
distinct
locations,
within
each
of
eight
casinos.
If
no
further
categorization
were
to
be
performed,
this
would
result
in
a
problem
with
hundreds
of
thousands
of
variables.
While
it
is
not
infeasible
to
solve
such
large
problems,
it
is
problematic
to
collect
valid
data
for
them.
Therefore,
it
will
be
necessary
to
break
down
the
machine
types
into
a
manageable
number
of
categories.
Most
of
the
effort
will
be
directed
to
understanding
which
games
and
models
can
be
grouped
together,
as
there
are
hundreds
of
these
combinations.
Preliminary
goal
is
to
have
no
more
than
20
machine
groups.
This
will
allow
managers
to
have
a
great
deal
of
flexibility
in
terms
of
what
games
and
themes
they
employ,
as
long
as
these
machine
adjustments
do
not
cross
group
boundaries.
The
process
of
grouping
will
be
profit
and
demand
driven.
First,
each
machine
group
will
have
certain
profit
range
associated
with
it.
This
profit
range
will
be
derived
from
the
historical
performance
for
all
machines
of
the
given
group.
Furthermore,
the
profit
range
will
vary
according
to
casino,
location
within
that
casino,
denomination
setting,
and
time
of
year.
If
any
of
these
parameters
are
found
to
be
persistent
in
a
group,
the
parameter
may
be
excluded
after
further
analysis.
This
is
why
historical
data
is
more
precise,
in
this
case,
than
vendors
estimates.
Second,
each
machine
group
will
be
examined
based
on
its
player
demographic.
This
will
require
analyzing
cardholder
transaction
data
to
determine
how
it
differs
(if
at
all)
from
the
overall
player
behavior.
If
significant
difference
is
found,
a
number
of
machine
groups
can
be
established
for
cardholders
benefit.
OR
will
provide
a
document
that
arranges
all
machines
into
a
small
number
of
categories,
along
with
the
explanation
as
to
why
these
categories
are
practical.
2.2.2
Performance
Forecasting
Forecasting
performance
of
a
machine
mix
will
be
an
important
component
of
the
project.
It
will
be
a
management
tool,
and
could
be
used
to
test
the
final
recommendations
against
the
current
mix
performance.
Therefore,
the
model
will
have
to
be
rigorously
verified
and
validated.
After
the
deliverable
of
machine
categorization
is
completed,
the
work
on
the
forecasting
will
begin.
OR
will
build
a
Monte
Carlo
simulation
model,
which
evaluates
the
performance
of
a
particular
machine
mix.
The
model
will
have
a
variable
time-horizon
cycle:
from
one
week
to
a
year
(shorter
or
longer
forecasts
will
be
increasingly
inaccurate).
The
mechanics
of
a
Monte
Carlo
model
are
simple.
First,
the
profit
range
for
each
machine
group,
location
and
denomination
will
be
entered
into
the
model.
Next,
the
model
will
draw
normally
distributed
random
numbers
from
each
profit
range
that
corresponds
to
a
given
machine
mix.
If
there
is
more
than
one
machine
in
the
group,
a
random
number
will
be
drawn
for
each
machine.
The
sum
of
all
these
numbers
will
estimate
the
profit
for
one
cycle.
Repeating
this
procedure
a
sufficient
number
of
times
(determined
analytically)
will
allow
the
construction
of
a
confidence
interval
(of
desired
width)
that
can
be
used
as
a
forecast
of
the
given
machine
mix
performance.
Before
discussing
solving
the
problem,
one
must
be
formulated.
LDE
will
pursue
two
approaches
here.
The
first
will
rely
on
deterministic
assumption.
With
this
approach,
it
is
assumed
that
all
data
regarding
revenue
is
known
with
certainty
(average
values
will
be
used).
The
solution
will
be
crude,
as
it
does
not
take
into
account
variability.
However,
it
will
at
least
provide
a
basic
benchmark
for
any
additional
optimization.
Furthermore,
the
optimal
mix
solution
under
the
deterministic
assumption
can
be
used
to
test
the
forecasting
model
described
in
section
2.2.2.
Second
approach
will
use
stochastic
optimization.
This
procedure
assumes
that
the
revenue
is
variable
(a
random
variable).
Depending
on
the
data
analysis
and
desired
time
horizon,
stochastic
model
may
include
seasonal
effects
and
cost
variability.
Both
approaches
will
have
similar
problem
statements,
but
the
stochastic
method
will
have
a
more
complex
objective
function.
There
is
a
number
of
ways
to
compute
the
total
profit
mentioned
above.
One
way
is
to
consider
how
much
revenue
a
machine
group
generates
per
unit
time.
This
is
practical,
since
machines
are
often
relocated
within
each
casino,
and
different
locations
lead
to
different
revenues.
It
will
be
straightforward
to
compute
these
parameters
based
on
the
transaction
data.
From
these
profits,
costs
will
be
subtracted.
Costs
are
associated
with
moving
and
setting
up
of
machines.
If
it
is
established
that
set-up
costs
are
low,
compared
to
revenues,
they
may
be
excluded
from
the
model.
To
quantify
the
management
constraints,
a
survey,
that
is
to
be
distributed
among
LDEs
casino
managers,
will
be
prepared
by
OR.
This
survey
will
contain
the
information
on
the
categorization
effort
and
instructions
on
how
to
perform
inventory
according
to
this
categorization.
Additionally,
each
manager
will
be
asked
to
provide
an
upper
and
a
lower
bound
on
the
number
of
machines
in
each
group
that
the
manager
is
willing
to
have
at
that
location.
It
is
important
to
note
that
the
amount
of
change
at
each
location
is
completely
up
to
the
manager
and
is
very
flexible.
Knowing
the
current
machine
mix,
a
manager
can
set
such
upper
and
lower
bounds
that
only
a
minimal
change
to
the
mix
is
allowed.
Moreover,
since
the
machine
groups
will
be
broad,
managers
will
be
free
to
mix
games
and
models
within
each
group.
The
completion
of
the
optimization
part
of
the
project
culminates
in
a
report
that
recommends
an
optimal
machine
mix
along
with
its
analysis
and
comparison
to
the
current
mix.
This
comparison
will
be
performed
using
the
forecasting
tool,
which
will
be
completed
before
this
deliverable.
Input
Data
Structure
For
machine
categorization,
the
overall
transaction
data
and
cardholder
transaction
data
is
necessary.
The
overall
aggregate
transaction
data
should
contain
the
following
fields:
machine
name,
casino,
section,
denomination,
manufacturer,
model,
period
of
record,
realized
gross
revenue
for
the
period,
number
of
plays
for
the
period.
For
simplicity,
the
above
fields
can
be
separate
for
each
type
of
machine.
Transaction
data
for
cardholders
can
be
provided
in
the
following
format:
cardholder
id,
age,
locality,
section,
casino,
machine
name,
denomination,
period
of
record,
gross
revenue
and
number
of
plays.
Complete
data
sets
should
include
as
many
records
as
there
are
available
(going
at
least
one
year
back).
Additional
data
may
be
required,
but
LDE
will
be
notified
in
advance
of
any
such
possibility.
Forecasting
model
will
be
based
on
categorization
data
set
(see
next
section).
No
input
is
required
from
LDE.
At
optimization
stage,
cost
data
is
necessary.
These
data
can
be
in
the
following
format:
machine
name,
manufacturer,
model,
casino,
section,
decommissioning
cost,
procurement
cost.
Note
that
if
a
particular
model
is
already
available
at
a
given
casino
section,
its
procurement
cost
is
zero.
Otherwise,
if
a
machine
is
not
available,
the
cheapest
procurement
cost
is
to
be
used.
Costs
may
be
estimates
and
same
numbers
can
be
provided
for
machine
of
similar
types.
Output
Data
Structure
A
deliverable
of
categorization
component
of
the
project
will
be
a
data
set
where
each
machine
is
listed
under
some
group.
Every
machine
group
will
have
associated
with
it
a
profit
range
(2
fields),
depending
on
casino
location,
position
within
casino,
denomination,
and
time
of
year.
For
example,
one
group
of
machines
may
have
a
profit
range
between
20,000
and
30,000
in
3rd
quarter,
given
they
are
at
Aries
casino,
in
the
interior,
and
with
denomination
setting
of
$0.01.
Majority
of
groups
will
not
have
overlapping
machine
types.
However,
based
on
demographics
analysis,
a
group
may
be
created
to
recognize
the
preferences
of
cardholders.
This
group
will
contain
overlapping
entries
with
other
groups.
This
data
set
will
lay
the
groundwork
for
both
forecasting
and
optimization
components.
Given
a
machine
mix,
grouped
based
on
the
above
categorization,
the
output
of
running
the
forecasting
model
is
an
average
and
a
95%-confidence
interval
for
profits
due
to
the
mix.
The
profits
will
be
broken
down
by
casino
and
section,
as
that
is
more
meaningful
from
management
perspective.
Solution
of
the
optimization
problem
will
give
an
optimal
machine
mix.
That
is,
how
many
machines
from
each
group
are
to
be
placed
at
each
location,
and
at
what
denomination.
Solution
to
stochastic
problem
will
also
include
a
schedule
that
reflects
how
the
mix
should
be
adjusted
in
different
periods.
To
make
the
result
more
readable,
the
final
report
will
contain
a
breakdown
of
machine
groups
by
casinos
and
sections.
Assumptions
As
the
project
is
in
early
stages,
a
number
of
assumptions
have
to
be
made.
Some
of
these
will
be
verified
or
rejected
as
the
project
progresses,
while
others
will
remain.
First
and
foremost,
OR
assumes
that
the
transaction
data
provided
by
LDE
is
accurate.
A
tolerable
amount
of
data
may
be
missing,
in
which
case
one
of
two
tactics
will
be
used.
If
the
missing
data
is
not
essential,
it
can
be
ignored.
Otherwise,
missing
data
will
be
collected
through
expert
opinion
and
estimation.
OR
expects
that
machines
generate
different
revenues
depending
on
the
section
of
the
casino
they
are
in.
Therefore,
transaction
data
should
reflect
this
level
of
detail.
Second,
it
is
assumed
that
operating
cost
is
approximately
the
same
for
all
machine
types.
This
simplification
allows
the
optimization
model
to
be
simplified
by
excluding
the
operating
cost
from
the
objective
function.
The
goal
is
to
maximize
machine
performance,
which
is
independent
of
operating
cost.
Third,
linear
optimization
assumptions
of
proportionality
and
additivity
hold.
That
is,
if
a
machine
generates
certain
amount
of
revenue
per
day,
then
using
two
such
machines
will
generate
double
the
amount
on
the
same
day.
Similarly,
if
one
machine
has
revenue
10,000
and
other
has
revenue
30,000,
then
using
them
together
generates
revenue
40,000.
While
these
may
not
reflect
reality
perfectly,
as
demand
may
be
dependent
on
machine
mix,
high
traffic
of
customers
and
stochastic
optimization
will
smooth
the
covariance
effects.