12
Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing costs. www.portraitsoftware.com A white paper by Portrait Software™ December 2006

Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

  • Upload
    others

  • View
    7

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

Optimal Targeting

through Uplift Modeling:

Generating higher demand andincreasing customer retention whilereducing marketing costs.

www.portraitsoftware.com

A white paper by Portrait Software™

December 2006

Page 2: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

2

Most direct marketing targets the wrong people.

It wastes money by focusing effort on many who will not reactpositively, and some who may even react negatively, while neglectingothers who would respond favorably if targeted. This paper shows

how a revolutionary new modeling technique – uplift modeling – canbe used to optimize targeting to maximize the returns from direct marketing. It achieves this by predicting, at an individual level, the change in behavior likely to result from a particularmarketing intervention.

Abstract

We need to target customers whose behavior we canchange, rather than taking credit for outcomes thatwould have occurred even without our intervention.

This paper discusses the use of a new kind of modeling – upliftmodeling – to improve the profitability and effectiveness of directmarketing spend. We start from five observations about targetedmarketing aimed at demand generation (cross-selling, up-selling anddeep-selling). These are:

1. Direct marketing works better for some customers than others.

2. Direct marketing has negative effects on some customers,reducing their level of spend.

3. The customers who spend most when targeted are notnecessarily the best targets for marketing: some of those whosespend is high when targeted would spend as much, or more,without treatment.

4. Not all purchases by targeted customers are caused by our directmarketing. Some would have bought anyway. To measure theincremental impact or uplift in sales resulting from a specific,targeted action we need to employ control groups. However, thisis merely a post-campaign assessment of impact: it does not tellus how to target better in the future.

5. The standard approach to targeting involves modeling only thetreated group from a previous campaign. This allows us toestimate each customer’s conditional probability of purchase iftargeted, but not the change in that probability resulting from ourtargeted action. Mathematically, such “response” models predict

Prob (purchase treatment) (1)

(the probability of purchase given treatment) while our returns arebased on uplift, defined as

Prob (purchase treatment) – Prob (purchase no treatment). (2)

In light of the points above, it is clear that traditional methods of targetingare suboptimal, based as they are on so-called “response” models thatactually predict conditional purchase probability1 (equation 1).

Ideally, we should model the difference in behavior between thetreated and control groups. This would allow us to predict the changein each customer’s purchase probability when treated (equation 2).We call such models uplift models.

In this paper, we show real-world examples of how uplift modelsgenerate higher returns on marketing spend, often from lowertargeting volumes. This constitutes a multiple win, because salesincrease while targeting costs are reduced. Additionally, whereincentives are used, the cost of these is usually significantly reducedbecause they are not “wasted” on customers who would purchaseanyway. Customer satisfaction also tends to improve as customerswho dislike being contacted are less likely to be targeted.

The same shift in approach can also bring even larger benefits in thearea of customer retention, where similar considerations apply, andwhere it is distressingly common for retention activity to drive upattrition rates. Adoption of uplift modeling involves a move fromtargeting on the basis of estimated attrition probability to an approachbased on savability. We illustrate that such a change can have a verysignificant impact on the bottom line, which is natural given thecounterproductive effect of spending money to drive customers away.

1 whether weighted by spend or not

Uplift Modeling: An Overview

Page 3: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

Uplift modeling requires controlgroups and is of greatest benefit incompetitive markets wherecustomers are subject to manyinfluences. It can be used toreduce costs, to increase revenuesand to enhance customersatisfaction and retention.

Uplift modeling allows businesses to optimizetheir targeting of customers in such a way asto maximize the return on investment ofmarketing spend or to optimize somealternative goal. The benefits from this in theareas of demand generation and customerretention can include:

• Reducing the number of customers requiredto achieve a given level of businessstimulation and thus reducing costs.

• Increasing the level of business generationachieved for any given level of spend.

• Lowering customer dissatisfaction bydecreasing the level of negativelyreceived material.

• Enhancing understanding within thebusiness of the effectiveness of variouskinds of marketing spend.

• Eliminating many or all of the negativeeffects associated with mis-targetedcampaigns.

• Increasing customer retention.

Additionally, there is the potential to usethese methods in other areas, such asoptimizing the management of actions basedon behavioral credit scores, optimizingcollections targeting and optimizing responseto customer complaints.

There are four qualification criteria importantfor evaluating the potential impact of upliftmodeling:

• Control groups: Uplift modeling absolutelyrequires the systematic use of randomizedcontrol groups, so is only of benefit tobusinesses that already use, or areprepared to adopt this approach.

• Size of customer base: Businesses withlarge customer bases will tend to benefitdisproportionately more than those withfewer customers, because reasonablevolumes are needed to model thissecond-order effect. Though dependenton response rates, uplifts and the ratio oftreated to control group sizes, we wouldnot normally expect the technique to workwell on campaigns targeting fewer thanabout 100,000 people.

• Influences: Uplift modeling is of particularbenefit in situations in which many factorsinfluence a customer’s behavior.Companies with a largely single-channel,single-communication route to customers,in less competitive markets, will benefitless from uplift modeling than those withmany channels and communication pathsand greater competition.

• Overall level of uplift: Portrait believesthat the approach of uplift modeling iscorrect even when the overall level of upliftis high and there are no obvious negativeeffects involved in a campaign. In thesecircumstances there are alternatives, albeitsuboptimal ones. In circumstances wherecampaigns show little overall uplift, ornegative overall uplift, other than acomplete change of approach, there is noknown alternative to uplift modeling.

33Who Can Benefit from Uplift Modeling?

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

Page 4: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

4

An uplift model for a catalog mailingallowed the incremental spend tobe raised by 15% for a fixed-sizetarget group.

A retailer used a catalog mailing to drivegreater spend activity amongst activecustomers – an example of deep-selling.2

The customers were selected on the basis ofa conventional “response” model – the so-called “champion” model – built on data fromcustomers mailed in a similar previouscampaign. Approximately 100,000 of thoseranked as likely high “responders” by themodel were targeted, and approximately50,000 were held back as a control group.On average, spend increased by $8 perhead among those mailed. An uplift modelwas then built, using data from the samecampaign as that used to construct thechampion model. In contrast to thechampion model, however, the uplift modelused information about the control group aswell as the mailed group.

Graph 1 shows a gains chart for “uplift inspend”.3 As with a conventional gains chart, the horizontal axis shows theproportion of the populated target.Customers are sorted by score, with thosehaving the best expected outcomes on theleft. However, whereas the vertical axis of aconventional gains chart shows the proportion of total spend achieved whentargeting a given proportion of the population,this uplift gains chart shows the proportion of total incremental spend achieved. (This is assessed by comparison with the control group.)

As is the case with a conventional gainschart, the diagonal line shows the result ofrandom targeting. Under this scenario,targeting any x% of the customers shouldyield the same x% of the total incrementalrevenue. This acts as a baseline, and any

useful score will yield a curve that is “bowed”above this diagonal.

The blue line (dashed) shows the result oftargeting with the “champion” model and thered line (solid) shows the effect of targetingwith the uplift model.4 Up to about 20%, thetwo models perform similarly. Thereafter, themodels diverge. For example, if 50% aretargeted, the uplift model manages to identifycustomers delivering about 16% more

revenue than the champion model –approximately $11.84 against $10.24 perhead. And if 80% are targeted, the upliftmodel manages to retain 97% of theincremental spend ($9.70 per head) againstonly 85% ($8.50 per head) from thechampion model.

Clearly, the uplift model is significantly betterat identifying customers for whom marketingspend generates a positive return.

Uplift Results Summary

Proportion of population targeted 50% 80%

Incremental spend per head – champion model $10.24 $8.50

Incremental spend per head – uplift model $11.84 $9.70

Extra spend per head gained from using uplift model $ 1.60 $1.20

Percentage increase by using uplift model 16% 14%

Graph 1: Demand Generation

Pro

port

ion

of t

otal

ext

ra s

pend

ach

ieve

d

1.0

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0.0

random q = 0.0%

Uplift Model (validation) q0 = 35.2%

Champion Model (validation) q0 = 20.1%

Proportion of population targeted

0.0 0.2 0.4 0.6 0.8 1.0

Example 1: Deep-Selling

2 While cross-selling tries to sell new products to customers, and up-selling aims to drive customers to upgrade, deep-selling simply tries toincrease the frequency or size of their transactions.3 Because the results shown here are from real companies, some results have been systematically rescaled to protect client anonymity andconfidentiality. 4 Both results are shown for validation data, which is also known as “holdout” or “test” data.

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

Page 5: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

5

A mobile phone operator ran acounterproductive retention mailingthat increased churn5 from 9% to10%. An uplift model finds a 30%sub-segment which, if targetedalone, will allow an overall reductionin churn from 9% to 7.8%.

A mobile phone company was experiencingan annual churn rate of approximately 9% ina segment. It targeted the entire segmentwith a retention offer, holding back only acontrol group. The net result was an increasein churn to 10% among the treated group,while the churn rate in the untreated groupremained at 9%. Obviously, this is the exactopposite of the desired effect, but we havewitnessed this phenomenon repeatedly. Itseems that retention interventions oftenbackfire because they variously remindcustomers of their ability to terminate,provide a catalyst to help overcome inertiaand annoy customers through intrusiveness.

Clearly, one way to improve the situation is tostop this retention offer entirely. However,there was a strong belief within the businessthat the offer was valid and did work forsome groups of customers. Also, no moresuccessful approach had been identified. Anuplift model was therefore built to try toidentify a sub-segment within which thetreatment was effective.

As in the previous example, Graph 2 is agains chart for uplift. The horizontal axisshows the proportion of the populationtargeted, while the vertical axis shows theresulting increase in total churn. (The goal, ofcourse, is to find a negative increase, i.e. areduction in churn.) The mailing wasuntargeted, so the diagonal line (blue) showsthe effect of random targeting of customers.Of course, with such random targeting, extracustomers are lost in proportion to the

number treated. However, the red line(concave) shows the effect of targeting onthe basis of the uplift model. The resultsare striking.

The model shows that retention activity waseffective for about 30% of the customers: ifonly the 30% identified as “most savable” bythe model are treated, churn across theentire segment falls by 1.2 percentage

points, from 9% to 7.8%, i.e. over 13%fewer customers churn. Compared totargeting everyone, churn actually reducesfrom 10% to 7.8%—a proportionate fall of22%. The segment contained approximately1 million customers, so using industry-standard ARPU of $400/year, the financialimpact of moving to uplift modeling, for onesegment alone, is as shown in the resultssummary above.

Example 2: Retention

Graph 2: Churn Reduction

Cum

ulat

ive

Upl

ift (%

)

2.00

1.60

1.20

0.80

0.40

0.00

-0.40

-0.80

-1.20

-1.60

-2.00

Random (existing) q0 = 0

Uplift Model (validation) q0 = 164.3%

Proportion of population targeted (%)0 20 40 60 80 100

Uplift Results Summary

Churn AnnualStrategy rate saving

Target everyone 10.0% $ 0

Target no one 9.0% $4,000,000

Target right 30% 7.8% $8,800,000

5 “Churn” is the term mobile phone companies tend to use to refer to customer loss or attrition

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

Page 6: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

6

The value of a new banking productis often high enough to justifycross-sell mailings, even whenthese generate additional sales inonly one in a thousand cases.However, comparatively smallsegments are often responsible formost or all of the extra salesgenerated. We have used upliftmodels to identify such segments,resulting in increases in campaignprofitability by up to five times.

Portrait is not alone in having concluded thatdirect marketing is best targeted bypredicting uplift. We have been approachedby a number of large, successful banks whohave themselves realized this, and in somecases have tried numerous techniques formodeling uplift, without much success. Inevery case, Portrait’s approach to upliftmodeling has far exceeded the performanceof the best alternative.

A typical scenario involves cross-sellingactivity aimed at increasing product holding.The value of many banking products is high,so that even an increase in product take-upas low as a tenth of a percentage point canprovide a positive return on investment formailings. However, we have shown that withappropriate targeting, we can usually achievebetween 80% and 110% of the sameincremental sales while reducing mailingvolumes by factors ranging from 30% to80%. Because the banks in question havethemselves attempted to model uplift, theytypically have historical data allowing fulllongitudinal validation6 of results.

One of the complicating factors in thesescenarios is that the uplift from the campaignis usually significantly smaller than the natural

purchase rate of the product being promoted– typically by a factor of five to twenty. It istherefore not unusual to see a purchase ratein the control group of around 1%, and in thetreated group of 1.1%. However, drivers ofthe base purchase rate are often quitedifferent from those of the incrementalpurchases resulting from the campaign.Because of this, non-uplift approaches totargeting are often doomed to failure, andsometimes actually perform less well thanrandom targeting.

Graph 3 shows an example of one suchcampaign. Here, the net effect of thecampaign was to increase the uptake of theproduct by a quarter of a percentage point.However, the uplift model shows that over60% of the increase in sales comes from just

10% of the targeted population, 90% comesfrom 40% of the population and 99%comes from 70% of the population. Noticealso that the “champion” model producedsubstantially worse results than randomtargeting – in fact, in this case, reversingthe ranking would have been very muchmore effective than using its actual output.This suggests that this campaign is beingeffective in stimulating demand from thevery people who tend not to purchasewithout intervention.

Example 3: Cross-selling High-value Products

6 Models are often validated by dividing a single historical dataset into two random subsets, building on one and measuring the quality of thepredictions on the other. With a longitudinal validation, data from two different time periods is used. Models are built using data from the oldertime period and validated using data from the more recent time period. This constitutes a stronger validation because it verifies that the modelsidentify patterns that are stable over time and really can be used to improve targeting in real-world situations.

Graph 3: Banking Cross-Sell

Cum

ulat

ive

Upl

ift (p

c pt

)

random Q = 0 q0 = 0

Uplift Model (validation) Q = 19.03%

Champion Model (validation) Q = 15.29%

0 20 40 60 80 100

0.25

0.20

0.15

0.10

0.05

0.00

-0.05Proportion of population targeted (%)

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

Page 7: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

7

Conventional models wastefullytarget some customers whoseoutcomes would be almost asgood or better if not treated. Upliftmodels identify an extra group ofcustomers whose outcomes canbe profitably improved but who areoverlooked by conventionalmodels because the absolute levelof their outcome falls below somearbitrary threshold.

A good way to get a feel for the differencebetween the approach encouraged by upliftmodeling and conventional targeting is toconsider the “swap sets” – customers whoare targeted under one approach but notthe other. Revisiting retention, thefundamental contrast is between targetingcustomers at high risk of attrition andtargeting those whose likelihood of attritioncan be materially reduced.

Graph 4 shows the probability of attrition7 onthe horizontal axis, and savability – decreasein churn likelihood when treated vertically. Wecan identify three swap sets.

1. Uplift models target a set of peoplewhose probability of attrition is nothuge, but who can be made much lesslikely to leave8 (segment 1, top left,burgundy). Retention activity can bevery effective here. In the context ofdemand-generation, the correspondingpeople are those whose probability ofpurchase when targeted is not high, butwho are unlikely to purchase at allwithout inducement.

2. The second swap set consists of peoplewho are at considerable risk of attrition,but are unlikely to be saved by our action(segment 2, thin pink strip above thehorizontal axis). Trying to retain themcosts more than the likely return. In ademand-generation setting, thecorresponding people are goodcustomers, who spend money, butwhose behavior is largely unaffected byour actions. Uplift models discouragetargeting these groups while conventionalmodels recommend them.

3. The third swap-set is segment 3 (blue,below the axis), which is targeted usingthe conventional approach, but not anuplift approach. This consists of peoplewhose outcomes are worse if treated, sowe actually spend money to drivecustomers and revenue away in thissegment. There seems no argument atall for such expensive andcounterproductive activity.

Comparison of Targeting Strategies

+100%

0%

-100%

Upl

ift T

arge

ts

Sav

abilit

y

CommonTargets

Conventional Targets

Conventional Targets

100%

Attrition Probability

1People at moderaterisk of attrition, but

highly savable

2People likely to

leave who can’t besaved economically

3People who will bemore likely to leave

if targeted

Graph 4: Uplift vs. Conventional Targeting

7 When not treated.8 Of course, everyone’s probability of being saved is limited to their probability of leaving, so the disparity can never be huge. In reality, therefore,a portion of the top left part of this schematic is empty.

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

Page 8: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

8

Response codes do not prove thata customer purchase is caused bythe associated marketing action.This is especially true whenincentives are offered, as customershave a strong motive for usingdiscount codes even when theywould have purchased anyway.Only control groups allow reliableproof of causality.

Rational marketers want to understandwhich of their actions work best, and inwhat circumstances. One of the methodsused to increase traceability is that ofresponse codes, and their adoption hasundoubtedly been of benefit. Unfortunately,however, response codes are not withouttheir problems.

When a customer quotes a response codeor uses a coupon, this clearly shows that heor she has been exposed to, and has takensome note of, the associated marketingintervention. Indeed, it is likely that asubstantial proportion of transactions orother outcomes tied to a response codewere caused by the intervention itself.However, in general it would be wrong toassume that all outcomes connected with agiven response code are incremental.

Consider two situations. In the first, a littleknown company launches a new product in anew category. Its only marketing is a directmail piece and there is no media coverage. Inthis situation it seems clear that all purchasesapparently in response to the mailing areactually caused by the mailing, because thereis no other way for a prospective customer toknow about the product.

Contrast this with the situation in which awell-known credit card company, with a largespend on mass advertising, sends out adirect mail piece. The fact that a customerapplies for a card using a mailed applicationform may mean that the mailing was crucialto that customer’s recruitment; but it may not.There are many influences on that customer –including competitor behavior – and themailed material may simply have been theeasiest way to carry out a decision made onthe basis of one or more of those otherinfluences. As with the banking examplesdiscussed earlier, we commonly see apparent“response” rates of 1% in this sort of

situation, but uplifts of more like 0.1percentage point (e.g. an account openingrate of 1.1% in a treated group against 1.0%in a control group).

The potential for being misled about theeffectiveness of marketing is largest when acampaign involves a discount or otherincentive. Most people will, if they are awareof a discount, use it even if they would havepurchased the same product anyway.Attributing a response to a campaign on thebasis of a discount coupon or responsecode associated with a discount, istherefore particularly dangerous, as there isa strong motivation for the customer totransact using that code even when there isabsolutely no causality.

The conclusion, therefore, is that rigorous useof control groups is the only safe way to assessthe true impact of a marketing campaign.9

Response Codes, Incentives, Causation and the Critical Importance ofControl Groups

9 Ideally, two control groups should be used. The first, which we call the “mailing” or “action” control group, is the normal one – a randomlychosen group of people who meet the selection criteria but who are omitted from the campaign to allow the uplift to be assessed. The second,which we call the “targeting” control group, is a randomly-selected group of people who do not meet the targeting criteria but are treated aspart of the campaign. This second form of control group is critical to allowing the effectiveness of the targeting – as opposed to the treatmentaction – to be assessed and improved.

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

Page 9: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

9

Negative effects are real and costly.They are most clearly seen in someretention activity, but can also behidden in marketing campaigns thatare successful overall. They bothreduce revenue and increase costs,so merit special attention.

There is no question that certain marketingactions have negative effects on somecustomers. This is most clearly illustrated bythe surprisingly common experience ofmobile telecommunications operatorsrunning retention campaigns thatdemonstrably increase overall churn. Theseeffects are actually quite easy to understand,as we shall explain below. It is less clear howoften campaigns with a positive overallimpact contain significant segments withinwhich there is a negative effect. Where thisdoes occur, it not only drives up campaigncosts, but actually leads to a lower grossoutcome than could be achieved bytargeting a smaller volume.

It’s not hard to understand why negativeeffects often occur in retention activity. Theconventional approach first identifiescustomers at high risk of attrition. By andlarge, these are people who are alreadydissatisfied in some way. An intervention,particularly by phone, runs a high risk of thecustomer’s responding by asking forimmediate cancellation – bringing forwardsome attrition and crystallizing othercancellations that may otherwise not havehappened at all because of customer apathy.

It is also the case that certain kinds ofintervention run intrinsically high risks ofnegative responses for some customers.Intrusive communication mechanisms, likephone calls, are often received badly. Nichecommunications, designed to appeal to aspecialist group, may well backfire on peoplewho do not belong to that group, or do notconform to that group’s stereotype. Onecreative’s “ingenious idea” is anothercustomer’s “inappropriate communication”.

One user of Portrait’s Uplift Modeling reportsthat even in demand-generation applications(cross-selling, deep-selling and up-selling)there are almost always material negativeeffects in the last one or two deciles by upliftscore, and on this basis would neveradvocate a 100% mailing, no matter howunarguably attractive the offer seems.

The financial impact of negative effects canbe large, because they not only cause acampaign to generate fewer incrementalsales or saved customers than it might: theyalso increase target volumes and costs.

Negative Effects

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

Page 10: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

10

There are a number of ways to useconventional models to improvetargeting for uplift. However, theseapproaches rarely match theperformance of models whose goalduring fitting is to predict uplift.

This paper argues that direct marketingactivity should be targeted so as to maximizethe desired change in behavior of customersfor every unit of marketing spend. Once thishas been accepted, there are a number ofpossible approaches. We discuss these inincreasing order of sophistication.

1. Target with a conventional model butkeep a control group. Needless to say,this is now standard practice, and doesnot constitute uplift modeling, but at leastallows assessment of the true impact ofmarketing after the intervention. If post-campaign analysis measures uplift on aper-segment basis, variations may evenallow improved targeting of future actions.

2. Build a conventional model but assessfor uplift. A subtle variation involvesbuilding a conventional model, but thenassessing uplift as a function of the score.If the uplift increases monotonically10 withscore, as is the case for Score 1 in Graph5 (upper, blue), we would feel fairlycomfortable targeting on the basis of thepurchase score. If, however, it were morelike Score 2 (lower, red), we might bemore nervous, and would likely choosenot to target customers with purchasescores between 7% and 10.5%.

3. Build two models and subtract. Perhapsthe most obvious approach is to buildtwo models, one for the treatedpopulation and one for the controlpopulation. These can then be subtractedto give an estimate of uplift. This is a fairlycommon technique among practitionerswho have realized the need for upliftmodeling, and it sometimes works well.However, the fundamental weakness ofthis approach is that the two models areindependent, each fitting a differentpopulation. Nothing specificallyencourages the models to fit the variationin uplift. This is true at all stages of themodeling process. Different variables maydrive uplift from purchase. Differentinteractions between variables may occur.Errors may occur in different places in the

models, introducing variation in predicteduplift where there is none in reality. Quitesimply, when we perform twoindependent regressions, we are not evenasking the models to fit the difference inoutcomes, so there is little reason tosuppose that they will.

4. Directly fit an uplift model. Naturally, themost powerful uplift models are thosethat are optimized (fitted) on the basis ofuplift. What is really needed is anapproach in which the fitting processdirectly uses information from both thetreated and the control populations andoptimizes the fit of the model to thedifference in behavior – the uplift. This isprecisely the approach that Portrait’sUplift Modeling solution takes.

10 i.e. if higher predicted outcomes always correspond to higher levels of uplift.

Building Uplift Models

Graph 5: Uplift as a function of Purchase Score

Upl

ift2.50%

2.00%

1.50%

1.00%

0.50%

0.00%

-0.50%

-1.00%

-1.50%

Score 1

Score 2

Purchase Score

3% 6% 9% 12% 15%

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

Page 11: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

11

Portrait Software markets theworld’s only uplift modelingsoftware, and has uniqueexperience in this field.

Quadstone, now part of Portrait Software,has pioneered uplift modeling, firstreleasing software containing suchcapabilities in 1998, and first publishing apaper on the approach in 1999. As wehave worked with customers and gainedmore experience with the techniques, wehave moved through three generations ofapproach, each progressively moresophisticated and powerful.

Portrait’s approach begins by modifyingstandard, greedy tree-building algorithms sothat they look at both a treated and a controlpopulation and directly model the difference.This involves modifying both splitting andpruning criteria, as well as the performancemeasure for the model itself.

Portrait is also acutely aware thattheoretically desirable data volumes are oftennot available in one or other of the treatedand control segments. (For roll-outcampaigns, control groups are often small;for trial campaigns, treated volumes tend tobe limited.) Moreover, even when overalloutcome rates are high, overall uplift is oftenlow, and sometimes zero or negative. We have therefore added a large number offeatures to the software to allow modeling forsmall populations, small effects and highnoise levels. We are not aware of any othercompany or group with comparableexperience in this area, or any other softwarewith packaged uplift modeling capability.

Portrait’s Uplift Modeling solution ispackaged software that can be used as astand-alone capability, as an integratedcomponent of the award-winning QuadstoneSystem or as an integrated component of aSAS environment.

Portrait Software’s Uplift Modeling Solution

December 2006A white paper by Portrait Software™ Optimal Targeting through Uplift Modelling

Page 12: Optimal Targeting through Uplift Modeling: Generating ... · Optimal Targeting through Uplift Modeling: Generating higher demand and increasing customer retention while reducing marketing

Contact Details

About Portrait Software™

Portrait Software™ is a global provider ofCustomer Interaction Optimization softwarethat helps B2C organizations provide a greatcustomer service, gain deep customerinsight and sell more, profitably.

Portrait Software™ solutions combineanalytical capabilities with a process-centric customer interaction managementplatform, providing organizations with theability to deliver the most highly optimizedcustomer processes across all sales andservice channels.

Portrait Foundation™ is the only specializedMicrosoft based Application Platform Suitethat enables organizations to quickly buildprocess centric, multi-channel interactionmanagement applications.

Portrait Interaction Optimizer™ enablesorganizations to exploit customer insight,evaluate in real time the “best next action”and automate the follow-up processes tomaximize the value of each and everycustomer interaction.

Portrait Customer Analytics™ deliversanalytic solutions that allows large consumerorganizations to explore and exploit customerinsight. It combines multiple sources of datato graphically discover, model, and predicttrends and relationships in customer behaviorin a collaborative environment.

Our 200+ customers are organizations thatlead the world’s most demanding customer-intensive sectors, including FinancialServices, Public Services, Telecoms and theIndependent Software Vendors that supportthese markets. Our customers includeKeyCorp, Wells Fargo Bank, Fidelity,Washington Mutual, Lloyds TSB, MerrillLynch, Nationwide Building Society, RainierPacific Bank, Chelsea Building Society,British Telecom, China AutomobileAssociation, T-Mobile, and Fiserv CBS.

Europe

Portrait SoftwareThe Smith CentreThe FairmileHenley-on-ThamesOxfordshire, RG9 6AB, UK

Email: [email protected]: +44 1491 416600Fax: +44 1491 416601

Americas

Portrait Software125 Summer Street16th FloorBoston, MA 02110USA

Email: [email protected]: +1 617 457 5200Fax: +1 617 457 5299

Asia Pacific

Portrait Software10 Faber Park129101Singapore

Email: [email protected]: +65 9752 1759Fax: +65 6872 6764

www.portraitsoftware.com

©Copyright Portrait Software International LimitedAll rights reserved.This document may contain confidential and proprietary information belonging to Portrait Software plc and/or its subsidiaries and associated companies.Portrait Software,the Portrait Software logo,Portrait and Portrait Software ’s Portrait brand are the trademarks of Portrait Software International Limited and may not be used orexploited in any way without prior express written authorisation of Portrait Software International Limited UK

*In December 2005, Portrait Software acquired Quadstone, a company recognized as best-of-breed forcustomer analytics. The acquisition provides Portrait with the ability to offer its customers real-timeintelligence and decision management tools for marketing, sales and service applications. The Quadstoneproduct suite has been integrated into the Portrait Collaborative Customer Analytics offering as well as otherproducts in the Portrait Customer Optimization suite.