Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
MKTG 521- Spring 2016 - 1
Individual-Level Targeting
Individual-level targeting.
Review choice modeling framework.
Using choice models for individual-level targeting (predictive modeling).
Demonstration of software for BBBC case.
Arvind Rangaswamy
www.arvind.info
MKTG 521- Spring 2016 - 2
Three Approaches to Targeting
Macro-level targeting (e.g., segment selection as in Pacific Brands)
GE-McKinsey Portfolio Matrix
Micro-level targeting using Classification models
Discriminant function
CART – Classification and Regression Trees
Micro-level targeting using RFM, Regression, Choice Models, Neural networks, Data mining, etc.
MKTG 521- Spring 2016 - 3
Models for Targeting Individual Customers
Create database of customer responses (choices) based either on test mailing to a sample of prospects/customers, or historical data of past customer purchases.
Step 1
Use models such as regression, RFM, and Logit to assess the impact of independent variables (drivers) of customer response. Step 2
Score each customer/prospect based on the drivers identified in Step 2 - the higher the score, the more likely is the predicted response.
Step 3
Classify customers into deciles (or smaller groupings) based on their scores. Step 4
Based on profitability analyses, determine the top deciles to which a marketing action (e.g., mailing of brochure) will be targeted. Step 5
MKTG 521- Spring 2016 - 4
Database for BookBinders Book Club Case
Predict response to a mailing for the book, Art History of Florence, based on the following variables accumulated in the database and the responses to a test mailing:
Gender
Amount purchased
Months since first purchase
Months since last purchase
Frequency of purchase
Past purchases of art books
Past purchases of children’s books
Past purchases of cook books
Past purchases of DIY books
Past purchases of youth books
Step 1
MKTG 521- Spring 2016 - 5
Attributes in ABB’s Choice-Segmentation Model
Invoice price
Energy losses
Overall product quality
Availability of spare parts
Clarity of bid document
Knowledgeable salespeople
Maintenance requirement
Ease of installation
Warranty
MKTG 521- Spring 2016 - 6
Drivers of the RFM Model
Recency
Frequency
Monetary
Value
Time/purchase occasions since the last purchase
Number of purchase occasions since first purchase
Amount spent since the first purchase
R
F
M
Total RFM Score: R Score + F score + M Score
Step 2
MKTG 521- Spring 2016 - 7
Example RFM Model Scoring Criteria
R
Months from last purchase
13-max
10-12 7-9 3-6 0-2
Max Score = 50
F
Frequency > 30 21-30 16-20 11-15 0-10
Max Score = 30
M
Amount purchased
> 400 301-400
201-300
101- 200
100
Max Score = 20
Implement RFM model using Nested If statements in Excel
Step 2
Present
State
Behaviors
Ignore
Postpone
Engage in
Purchase Process
Desired
State
Functional
and
Economic
Needs
Perceived
and
Psychological
Needs •Search for options
•Evaluate options
•Choose product
•Purchase product
•Use product
Customer
Value
Measurement
Approaches
Objective
Measures
of Value
Perceptual
Measures
of Value
Behavioral
Measures
of Value
Customer Needs and Buying Process
Motivation
Customer Needs and Customer Value Measurement
MKTG 521- Spring 2016 - 9
Choice Models vs Needs Surveys
With standard survey methods . . .
preference/ importance
choice weights perceptions predict observe/ask observe/ask
But with choice models . . .
importance
choice weights perceptions observe infer observe/ask
MKTG 521- Spring 2016 - 10
Using “Choice Models” to Determine the Value of a Customer
1. Observe choice:
Buy/not buy -- direct marketers Brand bought -- packaged goods
2. Capture related characteristics data:
demographics attitudes/perceptions market conditions (price, promotion, etc.) pattern of previous choices
3. Link
1 to 2 via “choice model” – the model predicts customers’ probabilities of purchase and also reveals importance weights of characteristics.
MKTG 521- Spring 2016 - 11
Logit Model of Response to Direct Mail
Probability of function of (past response behavior,
responding to = marketing effort, characteristics
direct mail of customers)
solicitation
Step 2
MKTG 521- Spring 2016 - 12
The Multinomial Logit Choice Model (Marketing Engineering Software)
The probabilities lie between 0 and 1, and sum to 1.
The model is consistent with the proposition that customers pick the choice alternative that offers them the highest utility (value) on a purchase occasion, but the utility has a random component that varies from one purchase occasion to the next.
The model has the proportional draw property -- each choice alternative draws from other choice alternatives in proportion to their utility.
The primary objective of the model is to predict
the probabilities that an individual will choose
each of several choice alternatives. The model
has the following properties:
Step 2
MKTG 521- Spring 2016 - 13
Conceptual Background for Choice Model
Uij = Vij + eij
where:
Uij = Utility (or hidden value) that customer i has for choice alternative j. In Bookbinder’s case, choice alternatives are “Buy” and “Don’t Buy.” We can set j = 1 for Buy and j = 0 for “Don’t Buy.” We do not observe Uij.
For “identification,” we will also set 𝑼𝒊𝒋 = 𝟎 for “Don’t Buy”
Vij = Deterministic (or observable) component of utility that is a function of several attributes associated with choice alternatives.
eij = An error term that reflects the non-deterministic component of utility.
Step 2
Value for
Gender Value of Buying
Art History of Florence = +
Value for previous
purchase of art books + + …..
Value of
unknown factors
MKTG 521- Spring 2016 - 14
Specification of the Deterministic Component of Utility
where:
i = an index to represent customers.
k = an index to represent an independent variable.
𝑿𝒊𝒋𝒌 = The values of the K independent variables for customer i
for choice alternative j.
wk = estimated coefficient to represent the impact of 𝑿𝒊𝒋𝒌 on
the value (utility) realized by customer i.
𝑽𝒊𝒋 = 𝒘𝒌𝑋𝑖𝑗𝑘
𝑲
𝒌=𝟏
Step 2
MKTG 521- Spring 2016 - 15
The Special Case of Binary (Two Alternatives) Logit Model
If customers maximize the value associated with their choices and the non-deterministic component error (𝒆𝒊) has a specific form called double exponential, then:
𝒑 𝒊𝟏 =𝒆𝑽 𝒊𝟏
𝒆𝟎 + 𝒆𝑽 𝒊𝟏=
𝒆𝑽 𝒊𝟏
𝟏 + 𝒆𝑽 𝒊𝟏
where:
𝒑 𝒊𝟏 = predicted probability that customer i will respond (j = 1) to the direct mail.
𝑽 𝒊𝟏 = estimated value of the offering to customer i (i.e., based on estimates of wk) obtained from maximum likelihood estimation. It varies from - (minus infinity) to .
The value to the customer of not having the offering (j = 0) is set to 0 (base value).
Step 2
MKTG 521- Spring 2016 - 16
More General Case
(Multinomial Logit Model)
𝒑 𝒊𝒋 =𝒆𝑽 𝒊𝒋
𝒆𝑽 𝒊𝒋𝑱
𝒋=𝟏
If there are three choice alternatives, we will get three
probability predictions:
𝒑 𝒊𝟏 =𝒆𝑽 𝒊𝟏
𝒆𝑽 𝒊𝒋𝟑
𝒋=𝟏
; 𝒑 𝒊𝟐 =𝒆𝑽 𝒊𝟐
𝒆𝑽 𝒊𝒋𝟑
𝒋=𝟏
; 𝒑 𝒊𝟑 =𝒆𝑽 𝒊𝟑
𝒆𝑽 𝒊𝒋𝟑
𝒋=𝟏
MKTG 521- Spring 2016 - 17
An Important Property of the Logit Model
Probability of Individual i Choosing the product 𝒑𝒊
0.0 0.5 1.0
Low
High
The marginal impact is highest when the customer is “sitting on
the fence,” i.e., when 𝒑𝒊 = 0.5.
Question: Is this a good property to have?
Marginal impact of
variable j (e.g., price)
on 𝒑𝒊
Step 2
MKTG 521- Spring 2016 - 18
Computing Scores Based on Regression
Run regression model to predict probability of purchase:
Choice (0 or 1) = Function of variables (e.g., Gender) + error
i.e., the estimated regression equation is given by:
𝑷 𝒊𝒋 = 𝒘 𝟎 + 𝒘 𝒌𝒌 𝑿𝒊𝒋𝒌
w's are estimated regression coefficients. Pij is the probability
that individual i will choose alternative j. Note that predicted
choice probabilities from the regression model need not
necessarily lie between 0 and 1, although most of the
probabilities will fall in that range.
Step 3
MKTG 521- Spring 2016 - 19
Compute Prediction Scores from Different Models
RFM Score: Use computed score as an index of the probability of purchase (i.e., higher the RFM score, the greater the probability of purchase).
Regression:
Logit:
𝒘 's are weights estimated by the Regression or Logit models.
RFM and Regression models can be implemented in Excel.
Also, all three scoring procedures for “probability of purchase” can be implemented in Excel.
k
ijkk XwwicustomerforScore ˆˆ)( 0
ijkk
ijkk
Xww
Xww
e
ety)(probabiliscore si' Customer
ˆˆ
ˆˆ
0
0
1
Step 3
MKTG 521- Spring 2016 - 20
Score Customers for their Potential Profitability (Example)
A B C D Score Average Customer (Purchase Purchase Expected $ Customer Probability) Volume Margin = A B C
1 30% $31.00 0.70 6.51
2 2% $143.00 0.60 1.72
3 10% $54.00 0.67 3.62
4 5% $88.00 0.62 2.73
5 60% $20.00 0.58 6.96
6 22% $60.00 0.47 6.20
7 11% $77.00 0.38 3.22
8 13% $39.00 0.66 3.35
9 1% $184.00 0.56 1.03
10 4% $72.00 0.65 1.87
Average expected purchase per customer = $3.72
Step 3
MKTG 521- Spring 2016 - 21
Decile Classification Example
Decile Customer(s) $
1 5 6.96
2 1 6.51
3 6 6.20
4 3 3.62
5 8 3.35
6 7 3.22
7 4 2.73
8 10 1.87
9 2 1.72
10 9 1.03
If the marketing cost to reach a customer is $3, at what decile will you stop
your targeting effort? How is this targeting plan different from one based
on average purchases of customers ($3.72)?
Step 4
MKTG 521- Spring 2016 - 22
Decile Classification
Standard Assessment Method
Apply the results of approach and calculate the “score” of each individual (calibration versus test sample).
Order the customers based on “score” from the highest to the lowest.
Divide into deciles.
Calculate/graph hit rate and profit.
Customer 1 Score 1.00
Customer 2 Score 0.99
….
Customer 230 Score 0.92
Customer 2300 Score 0.00
Decile1
Decile10
…..
…..
Step 4
MKTG 521- Spring 2016 - 23
Determine Targeting Plan (Example shows potential profitability of top 6 deciles)
Model
Number of hits (favorable responses at 60th percentile of
ordered scores)
Expected response rate by mailing the
top 60% of customers in the ordered list
% of likely responders recovered at
60th percentile
RFM
Regression
MNL
Compute profit/ROI for the models based on the number of mailings recommended
by each model and compare that to mailing to the entire list (equivalently to a
randomly selected list of the same size).
Step 5
MKTG 521- Spring 2016 - 24
Develop Lift Charts and Choose Model for Implementation
0%
20%
40%
60%
80%
100%
1 2 3 4 5 6 7 8 9 10
Decile
Cu
mu
lati
ve
Hit
Ra
te (
%)
0
5,000
10,000
15,000
20,000
25,000
30,000
35,000
40,000
45,000
50,000
55,000
Cu
mu
lati
ve
Pro
fit
($)
Hit Rate Random Hit
Profit Random Prof
Step 5
MKTG 521- Spring 2016 - 25
Database Approach to Implementing Individual Targeting
Sales, customer support,
shipments, payments, etc.
Appended
Company Data
Third-Party
Data (e.g.,
Credit Scores)
Model-based
Scores
Offers and
Responses
Customer
Database
Operational
Database
Transactions General
Ledger