1. Churn is about dealing with risk
The risk of a customer to Churn to another company
Hugo Cisternas
Director innovandis
2. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Risk
• The customer has made a Promise to continue using the service
– But the future is not predictable with certainty
– Not all the customers will continue using the service as promised
• Conclusion: Make an imperfect prediction
– Estimate the degree of risk involved in every individual case
– Define a risk level that is “acceptable”
4. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Example 1
• In a casino, John and Mary are gambling with dices...
– Mary is about to draw a dice of 6 sides. If he gets a 6 he wins $5,000 in
any other case he doesn’t get anything
– John is about to draw a dice of 10 sides. If he gets a 10 he wins $5,000 in
any other case he doesn’t get anything
• Who is facing more risk?
6. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
So, what is risk?
Risk is exposition to uncertainty
• If there is certainty,
– there is no risk
• If there is uncertainty, but you are not exposed to the results,
– there is no risk
Note: in colloquial terms, risk is used to refer to the possibility of occurrence of a hazardous
event, so if a bad event looks more possible to happen, people say the event is more
risky… this use of risk makes it more difficult to understand the technical definition of risk
above.
7. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
What is risk?
• Example: A person jumps on a parachute.
– Will the parachute open? If the parachute fails, this person will suffer the
consequences (probably he will die), then, he is assuming a risk
– But a spectator in the ground is subject to the same uncertainty abut
the parachute, but he will not suffer the consequences of a failure, then
he/she is not assuming risks **
** Unless the jumper owes money to the spectator, or is a relative. In those cases, this
spectators will suffer the consequences, emotionally or financially if the parachute
does not open, then they are assuming some risk.
10. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Risk in service contracts
Risk in service contracts:
• There is uncertainty in the capacity or willingness to pay the service in the
future (involuntary churn).
• There is uncertainty in the capacity or willingness to continue with the
service in the future (voluntary churn)
• The exposition is the debt left by a defaulting customer or the lose of the
future flows of incomes from a voluntarily churned customer if they have
continued with the service.
• Note that serious and continuous internal operating problems of the company, like service
problems, network capacity, invoicing problems, etc. are not risk factors! It is almost
certain that customers will switch to a better operator. This are “hygienic factors” to be
controlled. And if those factors are in place for long time, the churn model based on this
data will have a very short term life.
11. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Churn modeling
• A set of tools to Lower (or control) the risk
– At the moment of application (application scoring)
– At renewal time
– When upgrading the service
– At collections (if the customer is defaulting, will he/she repay de amount
due?)
– To prevent voluntary churn or attrition (behavior scoring)
15. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
What is a good predictive model?
• For marketers and management, a predictive model is not the objective,
it is a medium to reach an objective
• The objective in this case is to reduce churn
– To make customers to stay longer (and continue paying)
• To reduce churn, you have to know the actionable factors related to churn,
and act to prevent or change those factors.
• If you make a good job acting on the factors related to churn, the churn
prediction model will become obsolete.
• The best churn model will include this actionable factors as components of the
model, to be able to manage the churn prevention programs.
• Summary:
– The best churn model is not the one with best statistical precision.
– The best churn model is the one that provide best insights to further prevent churn
behavior
16. First law or prediction in marketing:
“A good churn model should have in its
development , the seed of it own
obsolescence”
Hugo Cisternas, 2001
Corollary:
A successful churn prevention program will require to
constantly rebuild the predictive model finding new
factors that drive churn behavior
19. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Dataset for analysis
• Establish a baseline
• Get knowing data at baseline time and back
• Flag the outcome at the end of the predicting frame
Baseline
Predictingframe
Still active
Churned
6 month
9 month
12 month
¿?
Back
20. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Which data is relevant?
• Every data may be relevant
• The analysis should omit prejudices about a data item
• Not all data items will become equally important
• There are interactions between some data
– Older people will, generally, have long time employment
– Singles will, in general, not own a house
• At the beginning the degree of interaction is unknown
21. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
First: Just count
• For each data item
– In each category
• How many churners are?
• How many active are?
– Example:
• How many home owners are churners? Active?
• How many singles are churners? Active?
• How many married are churners? Active?
– Better…
• How many frequent callers are churners? Active?
– Even better…
• How many frequent callers from previous 6 months who have increased calls
in the next quarter are churners? Active?
Please note: examples are simplified and variables are chosen for exposition
purposes, no real data s being used
22. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Example: Home ownership
• sample of 1.000 active and 1.000 churners
Please note: examples are simplified and variables are chosen for exposition
purposes, no real data s being used
Active Churners
# % # %
Owners 600 60% 300 30%
Rent 300 30% 600 60%
Other 100 10% 100 10%
24. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Example: Home ownership
• Calculate the “odds” of being active or churner
– Odds are 2 to 1 that an owner will be active
– Odds are 1/2 to 1 that a renter will churn
Please note: examples are simplified and variables are chosen for exposition
purposes, no real data s being used
Active Churner Odds
# % # % Of being active
Owner 600 60% 300 30% 2:1
Rent 300 30% 600 60% 1:2 or .5:1
Other 100 10% 100 10% 1:1
27. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Second: Define the samples
For which population will the churn model be build ?
• Have enough history
• Stability ( are there significant changes in economy, service, sales
promotions, competitors we have to take into account?)
• What will be predicted
– Involuntary churn (default, delinquent patterns)
– Voluntary churn (hard, soft)
– On application
– From behavior
• Have enough data
Note: When you look at your data, a lot of fields will be missing, data may look like
garbage (and it may be)… but you will always find useful data.
28. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Third: Define active and churn
• Define with no ambiguity which should be an active customer and a
churning (or churned) customer
• This definition should be operative, it will be used to select and classify
the data records
• Organize and align cohorts
29. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Fourth: Timeframe of prediction
• With how much time in advance will be made the prediction?
– At the time of the prediction, al customers should still be active
– At the predicted time, some will remain active and some will default to
churn
30. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Fifth: Get the data
• Get history behavior records
– Traffic
– Payment
– Complaints
– Contract changes (upgrades, downgrades, etc)
– Anything available… don’t’ overlook any data.
• Get demographic, psychographic data
• Get data from application
32. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Seventh: Data transformations
• Some data will require transformation
• Dates:
– Birth date becomes age
– Application date become tenure
– etc
• Amounts
– Is it necessary to correct inflation?
Note: take care of cohorts.. Time frame alignment is critical for transformations.
34. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Characteristic analysis
• Calculate percentages and odds
Active Churn % Active % Churn Odds
4 5 0.2% 0.3% 0.8/1
1106 1467 55.3% 73.4% 0.75/1
806 443 40.3% 22.2% 1.81/1
27 29 1.4% 1.5% 0.93/1
16 20 0.8% 1.0% 0.8/1
16 12 0.8% 0.6% 1.33/1
25 24 1.3% 1.2% 1.04/1
Not available
Rent
Own
From parents
From relatives
Government
Other
35. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Re categorization
• Some categories may have not enough cases.
– Join cases in one category.
Rent 1106 1467 55.3% 73.4% 0.75/1
Own 806 451 40.3% 22.6% 1.78/1
All other 84 85 4.2% 4.3% 0.98/1
Not available 4 5 0.2% 0.3% 0.8/1
Active Churn % Active % Churn Odds
38. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Getting the Score
• CHAID, CART, etc.
• Logistic regression
• Neural networks **
** Neural networks are a fascinating tool, but have an operative problem for
marketers: they are black boxes, there is very difficult (if not impossible) to
understand the underlying factors that explain the churn. If you cannot know
the factors, you have no insight to build your churning prevention strategy…
and you don’t want your model to change or adapt constantly without your
involvement.
Neural networks are excellent for reactive strategies to control imminent churning
customers.
50. Churn is about dealing with risk
The risk of a customer to Churn to another company
Hugo Cisternas
Director innovandis
51. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Hugo Cisternas
DIRECTOR INNOVANDIS
Database Marketing / Strategic Planning / Market Research
Contact: hcisternas@innovandis.org
With more than 25 years experience in Database, Information
Architecture and Statistical Analysis, had the risponsability of
Database Marketing and Planning for Wunderman clients between
1999 and 2010.
Wile at Wunderman, he led the team of Database Marketing and
Planning of the agency, both in the areas of direct marketing, internal
marketing, sales, marketing B-to-B and branding, including consulting,
design, implementation and campaign management, database
marketing and CRM.
Currently developing specialized consulting work, applying technology
and innovation to the demanding business and marketing needs that
the companies have today. It also makes classes and lectures.
52. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
HUGO CISTERNAS
Involved in projects like::
Database Marketing for Financiera ATLAS of Citibank, CMR Falabella,
Johnson’s, Codigas, Enagas, Isapre Consalud, Entel S.A., Entel PCS,
Seguros Cruz del Sur, Transbank, LanPass, Soprole, Caja de Compensación
Los Héroes, Ripley, Larraín Vial stokbrokers,
CRM consultancy for VTR Cable, Euroamérica life insurance, Torre, Larraín Vial
Direct Marketing for Citibank and Atlas, CMR Falabella, Multiopción credit card
from Johnson’s, Codigas, Enagas, Isapre Consalud, Entel S.A., Entel PCS,
Cruz del Sur insurance, Transbank, Caja de Compensación Los Héroes,
Metrogas, etc.
Branding and Planning: ATLAS Citibank, Johnson’s, Isapre Consalud,
Transbank, Caja de Compensación Los Héroes, Ripley, Aguas Andinas,
Mademsa, Cousiño Macul, Toblerone , etc
Internal marketing: ING, Metrogas, EntelPCS, Aguas Andinas.
IT and database projects: Servicio de Impuestos Internos (Internal revenue
services), National Library of Congress Telefónica CTC, Mutual de Seguridad,
CTC Celular (Movistar), Movistar (Argentina), TelCel (Venezuela), Ministerio
de Agricultura, Ministerio de Justicia, Ministerio de Relaciones Exteriores,
Canal 13 TV channel…
53. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
Hugo Cisternas
DIRECTOR INNOVANDIS
Database Marketing / Planificación Estratégica / Market Research
Contacto: hcisternas@innovandis.org
Con más de 25 años de experiencia en Bases de Datos, Arquitectura de
Información y Análisis Estadísticos, tiene la responsabilidad de los
servicios de Database Marketing y Planificación Estratégica de
Marketing para los clientes de Wunderman entre 1999 y 2010
Durante este período ha dirigido al equipo de Planning y de Database
Marketing en la planificación estratégica requerida por los clientes de
la agencia, tanto en las áreas de marketing directo, marketing interno,
promociones, marketing B-to-B y posicionamiento de marca, como en
la asesoría, diseño, implementación y administración de campañas,
database marketing y CRM.
Actualmente desarrolla trabajos de consultoría especializada, aplicando
tecnología e innovación a las exigentes necesidades comerciales y de
marketing que tiene la empresa de hoy. Además hace clases y dicta
conferencias.
54. Hugo Cisternas
Director innovandis
hcisternas@innovandis.org
HUGO CISTERNAS
Ha participado en proyectos destacados como:
Database Marketing para Financiera ATLAS de Citibank, CMR Falabella,
Johnson’s, Codigas, Enagas, Isapre Consalud, Entel S.A., Entel PCS,
Seguros Cruz del Sur, Transbank, LanPass, Soprole, Caja de Compensación
Los Héroes, Ripley, Larraín Vial corredores de bolsa,
Consultorías CRM para VTR Cable, Euroamérica Seguros, Torre, Larraín Vial
Marketing Directo para Citibank y Atlas, CMR Falabella, Tarjeta Multiopción de
Johnson’s, Codigas, Enagas, Isapre Consalud, Entel S.A., Entel PCS,
Seguros Cruz del Sur, Transbank, Caja de Compensación Los Héroes,
Metrogas, etc.
Posicionamiento y gestión estratégica de marcas como: ATLAS Citibank,
Johnson’s, Isapre Consalud, Transbank, Caja de Compensación Los Héroes,
Ripley, Aguas Andinas, Mademsa, Cousiño Macul, Toblerone , entre otras
Planificación y desarrollo de marketing interno para empresas como ING,
Metrogas, EntelPCS, Aguas Andinas.
Participación en proyectos tecnológicos y de bases de datos de gran
envergadura como por ejemplo: Servicio de Impuestos Internos, Biblioteca
del Congreso Nacional, Telefónica CTC, Mutual de Seguridad, CTC Celular
(Movistar), Movistar (Argentina), TelCel (Venezuela), Ministerio de Agricultura,
Ministerio de Justicia, Ministerio de Relaciones Exteriores, Canal 13 de
Televisión
Editor's Notes
CHAID: Basado en Chi Cuadrado, tiene debilidades
Regresión Logística: Mejor
Redes Neurales: Puede ser muy exacto, pero poco transparente