Customer Profiling, Segmentation and Marketing Strategies ...homes.stat.unipd.it/sites/homes.stat.unipd.it... · Mobile multimedia operator of Hutchison Whampoa Limited. H3G. H3G

July 18, 2002 |

Mobile multimedia operator of Hutchison Whampoa Limited

Customer Profiling, Segmentation and Marketing Strategies in Telecommunications

International Conferenceon Current Advances

and Trends in Nonparametric Statistics

July 15July 15--19, 2002 19, 2002 -- Crete, GreeceCrete, Greece

Bruno ScarpaCustomer Intelligence Manager

July 18, 2002 |


outline

! aim of this talk:

a. investigate the role of a statistician in this kind of companies

b. to present the statistical problems I met in my experiences in

telecom companies

c. to takle one (some) specific problem and present how we tried to

solve it

July 18, 2002 |


H3G

H3G Italy s.p.a. is born to be the first mobile operator of 3rd generation in Italy,completey focused on the develop of innovative and mutilmedia services in UMTStechnology

H3G is part of an international project leaded by Hutchison Whampoa, one of themost important player in the global economy

Cirtel International1,83%

NHS Investment5,64%

BMI2,26%

Gemina0,56%

Tiscali Finance0,36%

Hutchison 3G Italy Investment88,22%

HDP1,13%

H3G will be a supplier of content/services/solutions

Customers will have a connection always on, anyware they will be

H3G is planning to launch services (with the brand “3”) in the 4th quarter 2002

ITALYITALY

SWEDENSWEDEN

AUSTRIAAUSTRIA

UKUK

HONG KONGHONG KONG

AUSTRALIAAUSTRALIA

ISRAELISRAELMALESIAMALESIA

INDIAINDIA

THAILANDIATHAILANDIA

USAUSADDEENNMARKMARK

Hutchison

July 18, 2002 |


UMTS

!new mobile communication technology:

"Multimedia, interactivity and high speed

→Data transfer speed: 10 times faster than GPRS (300kbit/s vs 30kbit/s)

→Not just communication, but also content provider...

Mkt 1-2,5 G Mkt 3G

2002 2003 2004 2005

44,1

1,4

40,9

6,26,2

35,5

13,713,7

28,0

22,022,0

45,5 47,1 49,2 50,0

0

10

20

30

40

50Mln SIM

Forecasts

July 18, 2002 |


statisticiansH3G is a start upH3G has no customers (not yet!) H3G has not yet data!So why do they need data miners and statisticians?Which is the role of statisticians in this company?

BusinessDivisions

Sales

CRM

Planning & Control

Technical Divisions

Information Technology

Services

Network

Logistic

Marketing

Communication

Statistics

Analysis

Data mining

???

Where?

Orientation to Product: th F

Orientation to Market / C t

July 18, 2002 |


statisticiansH3G is a start up.H3G has no customers (not yet!)H3G has not yet data!So why do they need data miners and statisticians?Which is the role of statisticians in this company?

Business Divisions

Sales

CRM

Planning & Control

Technical Divisions

Information Technology

Services

Network

Logistic

Marketing

Communication

Statistics

Analysis

Data mining

•closer to the core problems of the company•give suggestions and use analysis to change and move business on needs of the customers•living in a behaviour in which very different experiences and competences are present: marketing, budgeting, actions, customer management, communication...•...but it is fontamental to work with IT peolple

Orientation to Product: th F

Orientation to Market / C t

July 18, 2002 |


statisticians

different level of possible involvement of a statistician in making business

commercial and

technological driver

make analisys and propose ideas: he is the motor of mktg ideas

support to marketing

men: quantitative help to mktg

ideas

statistician: “reports” and

basic data tables

focus on product management

focus on customer management

ReportsTables Simple indicatorSimple models

Tools: Statistical modelsData miningReportsTables Indicators

Tables Indicators/rateModelsData mining

ReportsTables Simple indicators H3G

beforeh

launc

July 18, 2002 |


statisticians

!takle the UMTS problems using the past experience:

"cellular telecom world (GSM, TACS...)

"internet world (ISP, ASP, e-commerce...)

!what kind of business questions we will have?

!what kind of data we will need to answer the new business

questions?

!„design of experiment“ based on the past experiences:

"prepare data

"prepare tools / packages

"prepare people

July 18, 2002 |


data!customer data

"general

→socio demographics

→activation data (e.g. Subscription date,

Type of contract, Plan subscribed, ...)

"billing data

"telco data

→telco traffic data

→VAS & Killer Applications data

→other value added services (both internal

and M-sites) data

"external researches and data

"market research

"costs data

"internet data

→web traffic data

→page provisioning/ Web contents (e.g.

page views, unique visitors, ...)

→e-mail data

→data gathered through the web (both

questionnaires and customer behaviour)

→community data

"customer base management data

→operational CRM data

→campaign Management & Marketing

contact history data

→loyalty programs data

"H3G mistakes data (e.g. billing errors, ...)

"…

July 18, 2002 |


TLC: some problems!customer acquisition

"prospecting

→achive a minimal thereshold of customers ASAP

→find and attract the right customers: how much to spend for each

different customer?

"detecting subscription fraud

→determine the likelihood of a fraudulent application

!customer profitability

"Customer value

"dormancy and share of wallet

"risk monitoring and managemet

→determine and optimize risk parameters

July 18, 2002 |


TLC: some problems

#Customer loyalty

"predictive approach→predictive models of churn

"attrition and retention

→model and determine main drivers

"actions: loyalty programs/campaign/

up sell-cross sell

"customer relationship

→personalization of the care and of the

contacts

!Customer profiling

"who are customers

"what each customer wants

"how to contact each customer

#actions evaluation

"no case-control experiment

→customers are autoselected

"evaluate a posteriori some action

→estimate the effect of the action conditionally to the effecof all the others variables

July 18, 2002 |


churn: GAM

!Goal: find for each customer a score of propensity to churn

"Understand wich variables have effect on the decision to churn and which is the effect of each variable considered between the others

!It is more important to understand effects than the accuracy of prediction

"A GAM model has been fitted to a random sample of (balanced) data

"Results have been evaluated on the entire customer base -2

.5-1

.00.

0

N Y

Pro

pens

ione

alla

dis

attiv

azio

ne Internet Mailbox

-1.5

-0.5

0.5

N Y

Pro

pens

ione

alla

dis

attiv

azio

ne Opzione Segreteria

-1.0

0.0

1.0

1 2 3 4

Pro

pens

ione

alla

dis

attiv

azio

ne Zona di Attivazione

Traffico ougoing

Pro

pens

ione

alla

dis

attiv

azio

ne

0 20000 40000 60000 80000

-1.0

0.0

1.0

Traffico incoming Tariffa 'Picco'

Pro

pens

ione

alla

dis

attiv

azio

ne

0 5000 10000 15000

-1.0

0.0

1.0

Traffico incoming Tariffa 'Ordinaria'

Pro

pens

ione

alla

dis

attiv

azio

ne

0 2000 4000 6000 8000

-1.0

0.0

1.0

Altre Sim Attive

Pro

pens

ione

alla

dis

attiv

azio

ne

0 1 2 3 4

-1.0

0.0

1.0

Data di Attivazione

Pro

pens

ione

alla

dis

attiv

azio

ne

-10

12

3

1995 1996 1997 1997 1998 1998

-0.5

0.0

0.5

1.0

15 - Busin

ess Tim

e

20 - Gold

50 - Valore

51 - Valore 50

52 - Valore 25

Prop

ensi

one

alla

dis

attiv

azio

ne

Piano Tariffario

-0.5

0.0

0.5

1.0

Bollettino Posta

le

Carta di C

redito

Domiciliazio

ne Bancaria

Prop

ensi

one

alla

dis

attiv

azio

ne

Metodo di Pagamento

-20

24

Cellular Promoters

Franchise

e

GDO Dealers

Indipendent Dealers

Major Acco

unts

OPI Stores

Office Automatio

n

Special C

hannels

Prop

ensi

one

alla

dis

attiv

azio

ne

Canale di Vendita

-0.5

0.5

1.5

N Y

Pro

pens

ione

alla

dis

attiv

azio

ne Programma Affari

Variabile A Variabile B Variabile C

Variabile D Variabile E

SI NO

Variabile I

a b c d e

Variabile F

a b c d e f g h

Variabile H

a b c

Variabile I

Variabile H Variabile M Variabile N

Pro

pens

ione

alla

dis

attiv

azio

ne

Pro

pens

ione

alla

dis

attiv

azio

ne

Pro

pens

ione

alla

dis

attiv

azio

ne

July 18, 2002 |


churn: CART

! Prepaid and “traditional” customer:

"the prepaid customer IS disconnected when he has not recharged for 12 months

"the “traditional” customer ASKS to be disconnected whenever he wants

! The prepaid customer has migrated to the competitors well before the “technical” deactivation

! We want to anticipate another event: a SIGNALSIGNAL of the decision not to recharge

July 18, 2002 |


churn: CART!The SIGNALSIGNAL should be

"“Intuitive” and “simple”

"Related with the choice of the customer to leave

"Accurate

→determine attrition classes

→use models to classify existing customers

→Determine

!It will be obtained using

"Incoming trafic (little)

"Outgoing trafic (little)

!Errors of the SIGNAL:

"Activ sims with signal on = 16%

"Churner sims with signal off = 15%

July 18, 2002 |


churn: CART

!A CART model is fitted predicting the “signal”

!The choice and the order of variables to be used in the model is defined a priori by the reserch team. The criteria of decision are basedon

"Knowledge of the business

"Actionability

"Previous models estimated

"Preliminary analysis and univariate estimates

!Gini index has been used as split rule

!The Estimate of an index of propensity to churn for each node has been obtained by an “evaluation” data set which has been separated from the estimate dataset at the beginning of the analysis

July 18, 2002 |


churn: CART

Activation before xxx"churn"=4.6%

“B” service user"churn"=2.5%

More than k e-mail received"churn"=2.6%

less than k e-mail received"churn"=6.8%

More than n outbound calls"churn"=3.1%

Age: more than z years"churn"=14.0%

company"churn"=16.0%

female"churn"=17.8%

Living in the regions a,b,c,d,e"churn"=19.2%

Living in the others regionspercentuale della customer base=1.8

"churn"=21.1%

male"churn"=20.3%

Age: less than z years"churn"=19.6%

less than n outbound calls"churn"=16.1%

“B” service not user"churn"=7.2%

Activation after xxx"churn"=6.0%

“A” service subscribers"churn"=5.5%

“A” service NOT subscribers"churn"=1.4%

Customer Base"churn"=1.8%

Percentile of the population

1% 5% 10% 20%

Lift 4.39 4.14 3.63 2.86

!For each node and leaf of the tree a different index of propensity to churn (signal). The colour of the leaves of the tree are related with therisk levels: $ low, $ medium, $ high

!To estimate the accuracy of the fit and the qualityof the model one of the most used measure in data mining is the lift

!!Lift:Lift: in a node is the ratio between the proportionof predicted churners in the node of the model over the total population in that node and theproportion of churners in all the custoemr base

July 18, 2002 |


churn: CART

pure data mining Example: Old model

%%DecreaseDecrease ofof trafictrafic

BBlack boxlack box solution wheresolution where the the software (IT) software (IT) selectselect thethevariablesvariables,, choosechoose andand fitfit the the model in amodel in a completely completely automaticautomatic wayway

Example: Model CART I

%%Trafic characteristicTrafic characteristic(e(exx. . HighHigh usageusage in the pickin the pick hourshours))

%%Services usageServices usage

%%DecreaseDecrease ofof trafictrafic

%%ComplaintsComplaints

data mining leaded

TheThe statistician statistician leadlead thetheanalysisanalysis, the, the choiceschoices (at(at least least partialpartial) of the) of the variablesvariables, decide, decidehow to fit modelshow to fit models andand useuse datadatamining models as analysis toolsmining models as analysis tools

non actionable!non actionable!

Marketing Marketing andandCRMCRM actionsactions

July 18, 2002 |


action evaluation! A particular Value Added Service has been prepared and sold in the last year

"All customers can subscribe it

"Question: has the subscription of this VAS some effect on churn rates?

→It is not possible to use a case-control experiment (autoselection)

→It is important to estimate the effect taking in account of the conjoint

effect of other correlated variables

→Use of data mining model to estimate the specific effect of the presence

of that VAS

→Explanatory variables are both continuous and discrete

→A GAM model has been fitted to a random sample of data

July 18, 2002 |


action evaluation! The effect is significant: te subscribers are more loyal than theothers customers VAS presence

NO YESre

lative

effe

ct

July 18, 2002 |


customer profiling

0%

0%

CU

STO

MER

VA

LUE

CU

STO

MER

VA

LUE

100%

100%

0% 0% CUSTOMER LOYALTY CUSTOMER LOYALTY 100%100%

MAXIMIZEMAXIMIZEVALUEVALUE

%%PERSONAL CC REPPERSONAL CC REP%%NEW VASNEW VAS%%MEMBER GET MEMBERMEMBER GET MEMBER%%CARE & DELIGHTINGCARE & DELIGHTING

NO/LOWNO/LOW CCOST OST ACTIONSACTIONS

%%NO ADDITIONAL COST FOR NO ADDITIONAL COST FOR RELATIONSHIPRELATIONSHIP%%“MASS” ACTIONS“MASS” ACTIONS

INCREASEINCREASEVALUEVALUE

%%UP / CROSS SELL ACTIONSUP / CROSS SELL ACTIONS%%STIMULATE USAGESTIMULATE USAGE%%STIMULATE VAS SUBSCRIPTIONSTIMULATE VAS SUBSCRIPTION

BUILDBUILDLOYALTYLOYALTY

%%SERVICE DIFFERENTIATIONSERVICE DIFFERENTIATIONINBOUND & OUTBOUNDINBOUND & OUTBOUND

%%LOYALTY SCHEME (Incentives)LOYALTY SCHEME (Incentives)%%LOCKLOCK--IN (Disincentives)IN (Disincentives)

!The customers and the company

"plot each customer on the map

"understand his own behaviour

"analyze deeply the costumer values

"apply the rules to define actions

"deploy and evaluate actions

"obtain customer lifetime value

July 18, 2002 |


conclusions

!business needs good statisticians!

!problems are many and interesting

!databases are huge, but we need people that look insidethem, not just packages that perform algorithms

!Insigth tools

"Non standard problems

→not always the solution is already in the tool

"commercial software proclame they solve every problem justclicking a button

"not always start with huge datasets means to analyze alldata and needs very fast (paarallelized) algorithm (and tools)

July 18, 2002 |


Bruno ScarpaCustomer Intelligence

Thank’s to

Laurence Jacobs (Kdlabs

and all the people (marketingIT, statisticians...

I worked with in these years

Customer Profiling, Segmentation and Marketing Strategies ...homes.stat.unipd.it/sites/homes.stat.unipd.it... · Mobile multimedia operator of Hutchison Whampoa Limited. H3G. H3G

Documents