0

1

This simple definition regards predictive analytics’ output and its value

proposition, rather than the technology behind it, i.e., how the predictive scores

are generated. But that technology is really the defining characteristic of

predictive analytics – so, look ahead in this slide deck for our complete working

definition.

3

4

6

7

8

10

i.e., “Should we contact this customer?”

Synonyms of uplift modeling:

Uplift modeling

Differential response modeling

Impact modeling

Incremental impact modeling

Incremental lift modeling

Net lift modeling

Net response modeling

True-lift modeling

True response modeling

11

Does contacting the customer make them more likely to respond?

MEDICAL:

Will the patient survive if treated?

"My headache went away!“ Proof of causality by example.

Driving medical decisions with personalized medicine: selecting treatment, e.g.,

treating heart failure with betablockers

Personalized medicine. Naturally, healthcare is where the term treatment

originates. While one medical treatment may deliver better results on average

than another, personalized medicine aims to decide which treatment is best suited

for each patient, since a treatment that helps one patient could hurt another. For

example, to drive beta-blocker treatment decisions for heart failure, researchers

"use two independent data sets to construct a systematic, subject-specific

treatment selection procedure." (Claggett et al 2011) Certain HIV treatment is

shown more effective for younger children. (McKinney et al 1998) Cancer

treatments are targeted by the patient's genes. (Winslow 2011)

Brian Claggett, Lihui Zhao, Lu Tian, Davide Castagno, and L. J. Wei.

"Estimating Subject-Specific Treatment Differences for Risk-Benefit Assessment

with Competing Risk Event-Time Data." Harvard University Biostatistics

12

13

15

The results of A and B trials, shown on the left, are simply the results of AB

testing, along with information about each customer they were tested on. This

two data sets serve as two distinct training sets for two distinct predictive

modeling tasks.

16

Source of second example: Kim Larsen, Uplift Workshop at Predictive

Analytics World

_______________________________________________

US BANK EXAMPLE

… to existing customers

Resulting improvements over prior conventional analytical approach:

Campaign ROI increased over 5 times previous campaigns (75% to 400%)

Cut campaign costs by 40%

Increase incremental cross-sell revenue by over 300%

Decrease mailings to customers who would purchase whether

contacted or not, and customers who would purchase only if not contacted.

Sources: Radcliffe & Surry (2011), Tsai (2010), Patrick Surry (Pitney Bowes

Business Insight), Michael Grundhoefer (US Bank)

18

Graph image reproduced with permission, courtesy of Kane et al (2011), as

depicted in their Predictive Analytics World presentation

19

20

Uplift modeling

Courtesy Portrait Software

21

22

The effectiveness of retention outreach targeted with a churn model is often itself

not modeled. So, it may be ineffective or lead to an adverse response, “waking

sleeping dogs,” i.e., triggering to leave customers who would otherwise stay.

A chain reaction best left untriggered.

24

25

Improvements are relative to their existing best-practice retention models.

Case study presented at Predictive Analytics World, February 2009, San

Francisco.

Case study and graph courtesy of Pitney Bowes Business Insight.

30

31

32

33

A more general, encompassing definition of uplift modeling.

34

35

Could be the results of a simple A/B test.

36

An example interstitial promotion. If the user accepts the offer, he/she is

allowing the host to pass profile information directly to the sponsor (in addition

to the fields shown).

37

A few additional percentage points can be tough to get, in the face of fairly adept

existing systems, but can make a big difference. Consider the insurance business,

where predictive analytics aims to reduce the loss ratio by 2 to 5 points beyond

that attained by standard actuarial methodology, or the engineering of jet engines,

where a 1% increase in efficiency would be a huge bite out of annual fuel

consumption.

The revenue results above are for interstitial ads only; many more ads are

embedded within functional product web pages, and could also be targeted with

only a slight alteration to the analytical system and deployment integration

developed for this project.

The large 25% increase in acceptance rates means formerly less "popular" ads are

now being given a better chance, leading to success; these sponsors are likely to

appreciate the increase in customer leads now coming from advertising with the

client.

Likewise, user satisfaction is likely higher, since users are seeing more ads in

which they are provably more interested.

38

39

Persuasion; influence.

40

41

The results of A and B trials, shown on the left, are simply the results of AB

testing, along with information about each customer they were tested on. This

two data sets serve as two distinct training sets for two distinct predictive

modeling tasks.

42

Net weight of evidence (a measure of uplift) varies by a customer's number of open revolving accounts. Graph courtesy of Larsen (2011).

Example variables that may generate uplift:

Engagement: Upside-down U such as the graph

above is common. Those customers towards the

right are "tapped out".

Other variables with similar upside-down U

phenomena:

Recency: Purchased their last car between 4 and

6 years ago as a target for marketing a new car.

Age: Direct mail for certain financial products

may have a stronger uplift for customers higher

43

Thanks to Patrick Surry at PBBI for this example segment.

Contacting entire list produces a slight downlift, but the segment above produces

an uplift.

This example is simplified for this illustration.

Both training sets A and B have the same variables.

Instead of identifying a “hot” segment with more purchasers/respondents than

average (i.e., predicting behavior), identify segments like this one within which

customers are more likely to be positively influenced by marketing contact, i.e.,

for which there is a higher purchase rate in training set A (the active treatment –

contact) than in training set B (the passive treatment – no marketing contact) for

the same segment.

44

45

(This paper in turn references all the core technical papers on this topic.)

Free white paper: www.predictiveanalyticsworld.com/signup-uplift-

whitepaper.php

46

47

48