How to apply CRM using data mining techniques.

Understanding Customers Who Stop Gordon S. Linoff Founder Data Miners, Inc. [email_address]

What to Expect from this Talk . . . Understanding Customers Who Stop

Learn about different reasons for stopping

Voluntary and Involuntary Churn

Learn different technical approaches

Predictive (“response”) modeling approach

Survival Analysis Approach

Learn about different applications

Planning marketing campaigns

Measuring the impact of business efforts

Estimating number of customers in the future

Who Am I? Gordon S. Linoff

Author or Co-author (with Michael Berry) of four books on Data Mining

Data Mining Techniques for Marketing, Sales, and Customer Support, Second Edition (1998, 2004)

Mastering Data Mining (2000)

Mining the Web (2002)

Data Analysis with SQL and Excel (2008)

Translated into Chinese, Japanese, French, and Italian

But not (yet) Spanish or Portuguese

Founded Data Miners, Inc. with Michael Berry in 1998

Worked with many leading companies, including Pfizer, New York Times, T-Mobile, Bank of America

Born and raised in Miami

Born here We are here

Why Do Customers Stop?

It’s those devils of temptation

Don’t Pay the Bill! Stick around Just Quit! Upgrade!

Ways to Address Voluntary Attrition

Allow customers who are not valuable to leave.

a cheap way to get rid of not-valuable customers

Offer incentives to remain for a particular amount of time.

teaser rates on credit cards, free weekend airtime, and so on

Offer incentives to valuable customers.

special features accessible through the Web and so on

Stimulate usage.

miles for minutes, donations to charities, and so on

Predicting Voluntary Attrition Can Be Dangerous

Voluntary attrition often resembles the other types.

Confusing voluntary and expected attrition means that you waste your marketing dollar.

If you confuse voluntary and forced attrition you lose twice:

once by wasting your marketing budget

again by increasing write-offs.

Ways to Address Forced Attrition

Stop marketing to the customer.

no more catalogs, billing inserts, or other usage stimulation

Reduce credit lines; increase minimum payments.

Accelerate charge-off timeline.

Increase interest rates to increase customer value while they continue to pay.

Predicting Forced Attrition Can Be Dangerous

Forced attrition often examines good customers.

They have high balances.

Confusing forced attrition with good customers may encourage them to leave.

Confusing forced attrition with voluntary attrition may hamper win-back efforts.

Visualizing Customers, The Calendar Timeline Customer Starts Customer Stops Customer Still Active

Predictive Response Modeling Assigns A Churn Score

Predictions for some fixed period in the future

Who will churn in the next month or the next six months?

Data to build the model comes from the past

Data to be scored (assigned a churn score) comes from the present

Models built using data mining and statistical techniques

decision trees, neural networks, logistic regression, regression, and so on

Jan Feb Mar Apr May Jun Jul Aug Sep Model Set Score Set 4 3 2 1 +1 4 3 2 1 +1

Case Study: Churn Prediction

Competitive market

6 carriers

Incumbent cellular carrier with 5 million customers

Churn rate of about 1%/month

Business goal: incorporate propensity-to-churn score as part of business processes

Churn modeling done in coordination with building data warehouse

Chose SAS Enterprise Miner for the project

“Assign a churn score to all customers”

This work took place over the course of two months

during the first month, we solved a simpler problem

during the second month, we built a more general churn model

Solving the Right Business Problem

Was: “Assign a churn score to all customers.”

Issues

recent customers with little call history

telephones? individuals? families?

voluntary churn vs. involuntary churn

how will the results be used?

Became: “By 24 September, provide a list of the 10,000 Elite customers who are most likely to churn in October.”

The new problem is much more actionable !!!

Technical Aspects of the Problem

Data

Customer information, such as billing plan, zip code, additional services, activation date and so on

Handset information, such as manufacturer, and so on

Dealer information, such as marketing region and size

Historical churn information by demographics, handset type

Billing history

Technique

Decision Trees

Tools

Oracle data warehouse

SAS Enterprise Miner

A Real Life Decision Tree Example: Predicting Churn

The training set is twice as large as the test set in this example.

Both have about the same percentage of churners (the numbers in black), about 13.5%.

As we build the tree, we want the performance on the training set and the test set to be similar.

13.5% 13.8% 86.5% 86.2% 39,628 19,814 This side is the training set This side is the test set

After the first split

After looking at all potential splits for all variables, the algorithm has chosen Handset Churn Rate for the first split.

In fact, the tree has a 3-way split, for values less than 0.7%, between 0.7% and 3.8%, and greater than 3.8%.

Notice that the performance on the training set and the test set is about the same.

13.5% 13.8% 86.5% 86.2% 39,628 19,814 3.5% 3.0% 96.5% 97.0% 11,112 5,678 14.9% 15.6% 85.1% 84.4% 23,361 11,529 28.7% 29.3% 71.3% 70.7% 5,155 2,607 < 0.7% < 3.8% Handset Churn Rate >= 3.8% Handset churn rate provides the best split The lift for this rule is 29.3%/13.8% or 2.12

As the tree gets bigger, more nodes are added 13.5% 13.8% 86.5% 86.2% 39,628 19,814 3.5% 3.0% 96.5% 97.0% 11,112 5,678 14.9% 15.6% 85.1% 84.4% 23,361 11,529 28.7% 29.3% 71.3% 70.7% 5,155 2,607 < 0.7% < 3.8% Handset Churn Rate >= 3.8% CALL0/CALL3 58.0% 56.6% 42.0% 43.4% 219 99 39.2% 40.4% 60.8% 59.6% 148 57 27.0% 27.9% 73.0% 72.1% 440 218 < 0.0056 < 0.18 >= 0.18 67.3% 66.0% 32.7% 34.0% 110 47 70.9% 52.0% 29.1% 48.0% 55 25 25.9% 44.4% 74.1% 55.6% 54 27 < 4855 < 88455 >= 88455 Total Amt Overdue

Learning from the Decision Tree

If someone has . . .

no other handsets on the account,

and a basic family price plan,

and a handset churn rate on the higher side, then

then they are very likely to churn (60.1%).

Some existing customers use the family plan to get discounts on new handsets.

Churn Model Success

Verified that initial group of 10,000 customers were candidates for retention offers

Extended model to all groups of customers

Built churn management system

And They Lived Happily Ever After

Understanding Customers Who Stop is More Than Churn Modeling

Churn modeling is quite powerful and quite useful

Under the right circumstances

There are other things we want to know . . .

What is the relationship of tenure and other covariates to stopping?

How much money are we gaining by keeping customers from stopping?

How many customers are going to stop in the future?

And hence, Survival Analysis

Survival Data Mining Adds the Element of When Things Happen

Time-to-event analysis

Terminology comes from the medical world

Estimating how long patients survive based on which patients recover, which patients do not

Can measure effects of variables (initial covariates or time-dependent covariates) on survival time

Natural for understanding customers

Can be used to quantify marketing efforts

Customers on the Tenure Timeline, Rather Than the Calendar Timeline Known Customer Start Known tenure when customers stop Unknown tenure for active customers (censored)

How Long Will A Customer Survive?

Survival always starts at 100% and declines over time

If everyone in the model set stopped, then survival goes to 0; otherwise it is always greater than 0

Survival is useful for comparing different groups

Use Censored Data to Calculate Hazard Probabilities

Hazard, h(t), at time t is the probability that a customer who has survived to time t will not survive to time t+1

h(t) =

Value of hazard depends on units of time – days, weeks, months, years

Differs from traditional definition because time is discrete – hazard probability not hazard rate

# customers who stop at exactly time t # customers at risk of stopping at time t

Bathtub Hazard (Risk of Dying by Age)

Bathtub hazard starts high, goes down, and then increases again

Example from US mortality tables shows probability of dying at a given age

Hazards Are Like An X-Ray Into Customers Peak here is for non-payment and end of initial promotion Bumpiness is due to within week variation

Hazards Can Show Interesting Features of the Customer Lifecycle 0% 1% 2% 3% 4% 5% 6% 7% 8% 0 30 60 90 120 150 180 210 240 270 300 330 360 390 420 450 480 510 540 570 600 630 660 690 720 750 780 810 840 870 900 930 960 990 Tenure (Days After Start) Daily Churn Hazard Peak here is for non-payment after one year Peak here is for non-payment after two years

How Long Will A Customer Survive?

Survival analysis answers the question by rephrasing it slightly:

What proportion of customers survive to time t?

Survival at time t, S(t), is the probability that a customer will survive exactly to time t

Calculation:

S(t) = S(t – 1) * (1 – h(t – 1))

S(0) = 100%

Given a set of hazards by time, survival can be easily calculated

Survival is Similar to Retention Retention is jagged. It is calculated on customers who started exactly w weeks ago. Survival is smooth. It is calculated using customers who started up to w weeks ago.

Average Tenure Is The Area Under The Curves

Survival to Quantify Marketing Efforts

A company has a loyalty marketing effort designed to keep customers

This effort costs money

What is the value of the effort as measured in increased customer lifetimes?

SOLUTION: survival analysis

How much longer to customers survive after accepting the offer?

How to quantify this in dollars and cents?

We Can Use Area to Quantify Results

Increase in survival is given by the area between the curves.

For the first year, area of triangle is a good enough estimate

93.4% 85.6% Note: there are easy ways to calculate the exact value 65% 70% 75% 80% 85% 90% 95% 100% 0 2 4 6 8 10 12 14 16 18 Months After Intervention Percent Retained Group 1 Group 2

Loyalty-Responsive Customers Are Doing Better Than Others

Survival for first year after loyalty intervention

Group 1: 93.4%

Group 2: 85.6%

Increase for Group 1: 7.8%

Average increase in lifetime for first year is 3.9% (assuming the 7.8% would have stayed, on average, 6 months)

Assuming revenue of $400/year, loyalty responsive contribute an additional revenue of $15.60 during the first year

This actually compares favorably to cost of loyalty program

Competing Risks: Customers May Leave for Many Reasons

Customers may cancel

Voluntarily

Involuntarily

Migration

Survival Analysis Can Show Competing Risks

overall S(t) is the product of Sr(t) for all risks

What happens to our customers over time?

A Better Way to Look at This

Customer-Centric Forecasting Using Survival Analysis

Forecasting customer numbers is important in many businesses

Survival analysis makes it possible to do the forecasts at the finest level of granularity – the customer level

Survival analysis makes it possible to incorporate customer-specific factors

Can be used to estimate restarts as well as stops

The Forecasting Solution is a Bit Complicated

Using Survival for Customer Centric Forecasting 0 10 20 30 40 50 60 70 80 90 100 110 120 130 Day of Month 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 Number Actual Predicted

Understanding Customers Who Stop