Uncovering Hidden Profits:
Th Next
The
N t Step
St
in
i Optimizing
O ti i i
Direct Marketing
Dmitry Krass
Rotman School of Management
University of Toronto
and
C
Custometrics
i Inc.
I
Outline
Direct marketing and Analytical Models
• Introduction
• Ranking vs. Forecasting Models
MO: Multi-Product Optimization
Multi-Product Multi-Period Optimization
Summary
2
Direct Marketing and Analytics:
A brief Introduction
1:1 Direct Marketing: marketing action customized to an
individual level
Typical Channels: Telemarketing (TM), Addressed/Unaddressed
Direct Mail (DM), E-mail, Salesperson contact, etc
“Analytics”
l i
(mainly
( i l statistical
i i l / data
d mining
i i techniques)
h i
)
have long been accepted in DM community
Acceptance
p
byy other marketing
g areas has been slower
However, analytics is usually compartmentalized and (in our
opinion) underused
3
Typical 1:1 DM Campaign
Planning / Design: Decide on
offer,, creative,, budgets;
g ; comm.
channels
•Traditional
“home”
home of
analytics
• Targeting
Models
Targeting: select targets from
(typically) a large universe of
potential targets
Execution: mailing schedules, etc.
Where analytics
belongs Requires
belongs.
• New data flows
•New / better
statistical
stat
st ca models
ode s
•Optimization
Models
•DSS Tools
•Automation
Evaluation: Did the message work?
Did the
h targeting
i work?
k? Overall
O
ll ROI?
4
Targeting Models: A brief Introduction
Goal: select “targets” (customers or
geographical regions) to include in the
p g from a ((much)) larger
g “universe”
campaign
Assume: contacting the whole universe is
infeasible due to cost and/or capacity
constraints
How should
H
h ld these
th
“best
“b t prospects”
t ” be
b
selected?
Decision Rules
Targeting Models: much more effective
Typical targeting model is a statistical
model relating expected response to
Gains Chart: Sample Program
5.00%
Exp. A
Activation Rate
e
4.00%
3.00%
2.00%
1 00%
1.00%
0.00%
Past behavioural history (RFM)
Target characteristics (demographics, etc.)
Other factors (seasonality, campaign specifics)
Goal: separate “best”
best from the “rest”
rest
Know: 20% of customers
often account for 80% of
response
1
2
Contact
top two
deciles
3
4
5
6
7
deciles
5
8
9
10
Targeting Models:
Ranking vs. Forecasting
Ranking: separating “high” from “low” responders
Often sufficient for Campaign Targeting
Forecasting: predicting actual response rates
A typical Targeting Model does a good job of ranking, but
is a poor forecasting tool
Forecasting is much harder!
Requires much richer data; frequent model re-builds
Poor Forecasting
Good Ranking
Gains Chart: Sample Program
2.0%
Ac
ctual Resp. R
Rate
Actual Resp. Ra
A
ate
2.5%
2.5%
1.5%
1.0%
0.5%
0.0%
1
2
3
4
5
6
deciles
7
8
9
10
Actual vs. Expected
2.0%
1 5%
1.5%
1.0%
Model
0.5%
Ideal
0.0%
0.00%
2.00%
4.00%
6.00%
Expected Response Rate
6
Why Ranking is not Enough
Simple planning question: how many targets to
include in the campaign?
Simple answer: communicate to margin
Margin = ($ Benefit of Response)*(Prob. Of Response) – ($
Communication Cost)
Include target if Margin > 0
But poor forecast of response undermines this!
Example
Benefit = $50, Comm. Cost = $1
Top 3 deciles should be included
(based on exp. response rates)
Actually should only include 2
deciles
Exp. ROI = 66%
Actual ROI = -1%
Margins: Sample Program
$1.50
Exp. Margin
$1.00
Act. Margin
$0.50
$0.00
$0.50
$1.00
1
2
3
4
5
6
7
deciles
$1.50
7
8
9
10
Building Forecasting Models
Relative target quality is often quite stable => ranking is
much easier then forecasting
T picall models mis-forecast
Typically
mis forecast beca
because
se
They fail to include communication history/ offer specifics
They fail to include recent economic/ competitive trends
Th data
The
d they
h were calibrated
lib d on is
i too old
ld
The coefficients/ set of significant factors is too old
Using predictive models for planning support /
optimization
i i i requires
i
Maintaining full communication history
Richer variable set
Rapid data refreshes
Rapid model rebuilding cycle
More data management
g
/ modelingg sophistication
p
8
Outline
Direct marketing and Analytical Models
MO: Multi-Product Optimization System
•
•
•
•
DM Planning and Analytics
System Description
Strategic Support
Beyond MO
Multi-Product Multi-Period Optimization
Summary
9
DM Planning Tasks
An Integrated DM campaign involves a number of
offers executed simultaneously (or nearly so)
H
Here, an ““offer”
ff ” may involve
i l different
diff
products
d
or
variations of the same product
Each “target” (customer or region) can receive at most
one offer
ff in
i a given
i
time
ti period
i d
Issues:
• Construct targeting lists for each offer
• Which customers to include? In which list?
• How large should different lists be?
• Given fixed overall budget, find optimal
targeting list design while meeting side
constraints
– These constraints often include minimal / maximal
list sizes, certain markets to include /exclude, etc.
10
Example
A telecommunications company runs monthly
unaddressed direct mail campaigns to acquire new
customers
Campaigns are targeted at a geographical region
level (postal walks or other distribution units)
• Each walk contains
i a number off households
• If a walk is targeted, every household must be mailed
The current campaign involves five different
product offers
• Products have different profitabilities in different
regions
Total budget is $1,000 (ok to exceed by a bit)
Average contact cost is $.175 per household
Constraint: the spend on each product should be
between $100 and $300
11
Data for Example
Assume first only ranking models are available
Product profit margins are Prod1>….>Prod5
12
Typical “Solution”: Priority Targeting
A typical company solves this problem by “priority
targeting”
Prioritize products (by profit margins or … less
objectively)
Prepare targeting lists one product at a time
Problems:
Targets that are desirable (highly ranked) for one
product are often desirable for many products
• Lists for higher-priority products “steal” all the best
targets; lower-priority products left with very poor lists
• High profit margin does not mean high profitability (more
expensive products often have lower response rates)
• Meeting all the constraints makes designing priorityassigned
i d li
lists
t non-trivial
t i i l
13
List Based on Priority Targeting
Product Priority in order of margin: Prod 1> …>Prod 5
Total budget is $1000
Each product spend should be between $100 and $330.
$330
14
From Ranked-Based Lists to MultiProduct Optimization
Issues with Ranked-based lists
There were a number of conflicts between products. Were they
resolved correctly (i.e., in profit
profit-maximizing
maximizing way)?
What is the expected number of orders for the created list?
Expected profit?
Need Forecasting, not Ranking models
Models should be able to
• Estimate effect of communication (i.e., sending offer X to
region Y)
• Account for cross-product
cross product effects (e.g.,
(e g Prod1 offer may
lead to Prod2 sale and vice versa)
To estimate Profitability need LTV (life-time value) estimates
p
for each product
Also need an Optimization Model to “enumerate” all possible lists
satisfying our requirements and pick the best one
These pieces together make up the “MO System”
15
Optimization Models:
Key components
Decision variables: represent target-to-list assignments
Objective Function (MOS): what are we trying to optimize?
Basis: response curve estimated by the forecasting model (for any
given assignment must be able to compute expected response for
each product)
Poor objective: maximize total response to the campaign
(
(responses
to
t products
d t A andd B may have
h
very different
diff
t financial
fi
i l
impact)
Good objective: maximize overall Profit
Need an estimate of LTV impact of response for each product;
costs (both contact and product delivery) may differ in different
locations
Constraints
B
Budget
d
Business Constraints (min / max support for each product , etc.)
Modeling Constraints – crucial. Ensure that optimization does not
leave the “trust
trust region
region” of the underlying forecasting model
16
Optimization Model - Comments
Software: “solvers” are becoming more prevalent, but
require highly-skilled user; interface with statistical
models a big issue
Dimensionality often very large (while number of offers
is usually small – less than 20, number of potential
targets can be in the millions)
May require customized optimization algorithms
Natural extensions: allow the model to decide which
offers to use
Makes the problem even harder to solve
Evaluation / control groups – much more intricate than
f the
for
th one-product
d t case
Key to the whole process: accurate and sophisticated
parameter estimates (i.e., good forecasting models built
on goodd ddata)
t )
17
Example Cont.
Optimized vs. Ranked List
18
Optimized
li t
list
performs
over 2x
better
Profitable in
all areas
V
Very
different
from the
ranked list
Strategic Decision Support:
Scenario Simulators
Cost of constraints: What is the impact of min/max spend
constraints?
In this example,
example removing constraints increases the opt.
opt profit by 6%
• Constraints are not very costly
The only products used in the unconstrained run are Prod3 and
Prod5
• These are the key products in spite of low unit profit margins
What is the optimal budget?
Plug in different budget levels into the model
Should consider increasing budget to $4000 range
19
MO and Planning Support
The true benefit of MO is that it allows
analytics to take an earlier seat at the
“analytics”
table:
Before budgetary decisions and offers are
finalized and while constraints are still flexible
By the “targeting”
targeting stage,
stage many of these
decisions are frozen, sharply reducing
potential benefits of analytical models
20
Issues with MO
List fatigue
Repeated communication with the same customers
induces “fatigue”
“Blackouts” – typical fatigue-fighting strategy
Can quickly lead to deteriorating list quality
Would like a more scientific approach to fatigue
management
• For some targets it may be optimal to suspend
communication for one or more periods; for others
repeated communication may be fine
Past communication affects future response
Need to ensure long-term,
g
, not one-period
p
profitability
21
Issues with MO – cont.
Spacing out offers may be better
Multi-wave
M lti
communication
i ti
Need to balance that versus potential fatigue
Multi-period planning
DM budgets are often quarterly or bi-annual,
while execution is monthly
Little flexibility in monthly budgets
Need multi-period, multi-channel approach
22
Outline
Direct marketing and Analytical Models
MO: Multi-Product Optimization System
M20: Multi-Product Multi-Period Optimization
System
•
•
•
•
Settingg
Thinking in sequences
Impact to date
Beyond M20
Summary
23
M2O: Setting
Assume k-period planning horizon
Total budget is specified
Have a set of offers (products) to select from and a
universe of targets
Often have additional constraints (new product
intros require certain minimal support, etc.)
Questions
Who?
When?
How?
H ?
Which offers should be used?
g be allocated?
How should the budget
24
The Essence of M2O
Right Customers
Right Products
Right Timing
25
M2O: Key Insight
Must think in terms of k-period communication
sequences
s = {Prod
{P d 11, P
Prod*,
d* P
Prod2,...Prod1}
d2 P d1}
Prod* - is an artificial “offer” signifying a decision not
to communicate
In our example, assuming 3-period planning horizon,
some of the sequences are
•
•
•
•
{Prod1, Prod1, Prod1}
{P d1 P d2 P d3}
{Prod1,Prod2,Prod3}
{Prod*, Prod4,Prod*}
Etc.
For each
F
h sequence s mustt be
b able
bl to
t estimate
ti t kk
period expected response pattern:
Expected response for each product in each period if
sequence s is
i usedd
26
M2O: Main Challenges
Estimation:
Statistical models must be sophisticated enough to capture full
p
cross-period
p
effects
cross-product
Data management
Must capture long communication history in the data
Must have enough
g variation in histories at the target
g level to enable
accurate estimation
Optimization
With m products and k planning periods there are (m+1)k potential
communication sequences for each target
The dimensionality of the Optimization model becomes huge
Even in our “toy” example with 3 planning periods, 5 product and
g
have 15*(6)3
( ) = 3240 decision variables
15 targets
For practical problems need effective heuristics (specialized
algorithms)
Planning cycles are often compressed in time: quick running times
a must
27
Payoff: M2O in Action
M2O Starts
Payoff: M2O in Action, an example
The system has now been implemented for 6
quarters
G
Gradually
d ll gained
i d trust, allowed
ll
d more participation
i i i in
i
planning / budget allocation
The system is very profitable for the client
About 2/3 of the total impact comes from better
targeting, about 1/3 comes from planning support
ROI of between 1,690% to 2,141% based on 12 month
net cash;
ROI of between 4,063% to 5,099% based on LTV
During each planning period, on average 50% of the
i iti l budget
initial
b d t gets
t re-allocated
ll t d based
b d on system’s
t ’
recommendations
The client is now much more attuned to costing out
constraints before imposing them,
them etc.
etc
29
Beyond M20: Multi-product, Multiperiod, multi-channel system (M30)
Natural next step is to integrate multiple
communication channels
Easy part: if “at most one communication per
target per period” rule holds, channels can be
regarded as new offers
Dimensionality grows, but the system still works
Hard ppart: communications can be overlaid
E.g., a direct mail and E-mail in the same period
Need to estimate cross-channel effects; may differ
by communication sequence
Dimensionality jumps
Work in progress...
30
Outline
Direct marketing and Analytical Models
MO: Multi-Product Optimization System
M20: Multi-Product Multi-Period Optimization
System
Summary
• Decision Support for DM
• Other challenges
g
31
Typicall
MO
M30
Decision Support for Direct Marketing
Basic
Execution: LTV-based cross-product cross-channel
multi-period optimization
Strategic
g support:
pp
Comprehensive
p
analysis
y of
budgetary scenarios, offer timing, channel
interactions
Prevalence: We (Custometrics) are getting there…
Execution: LTV-based analytical models with cross-product
single-channel optimization within each period
Strategic support: Limited (single-period, single-channel
scenarios can be analyzed)
Prevalence: state of the art; only a few leading-edge
organizations have access to this technology
Execution: Product-specific
p
rankingg models
Strategic support: data-base pulls only
Prevalence: typical operating mode for an organization with
strong analytics support team (major banks, etc.)
Execution:
E
i
B i
Business
ddecision
i i rules
l
Strategic support: data-base pulls only
Prevalence: very common 32
Other challenges
Integrating channels with different target
“concepts”
p
E.g., mass and 1:1 channels
Evaluation
Many challenging problems
As the sophistication of the Decision Support
S t
Systems
grows, so do
d the
th evaluation
l ti challenges:
h ll
how do you measure impact of a truly strategic
system?
Execution
Many tough and interesting analytical problems at
this stage as well
33
Optimizing
Direct Marketing Campaigns
Questions?
34
© Copyright 2026 Paperzz