How to apply CRM using data mining techniques.


Published on

Gordon Linoff, Co-Founder, Data Miners, Inc. Como aplicar CRM usando minería de datos. How to apply CRM using data mining techniques.

Published in: Business, Technology
No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide
  • How to apply CRM using data mining techniques.

    1. 1. Understanding Customers Who Stop Gordon S. Linoff Founder Data Miners, Inc. [email_address]
    2. 2. What to Expect from this Talk . . . Understanding Customers Who Stop <ul><li>Learn about different reasons for stopping </li></ul><ul><ul><li>Voluntary and Involuntary Churn </li></ul></ul><ul><li>Learn different technical approaches </li></ul><ul><ul><li>Predictive (“response”) modeling approach </li></ul></ul><ul><ul><li>Survival Analysis Approach </li></ul></ul><ul><li>Learn about different applications </li></ul><ul><ul><li>Planning marketing campaigns </li></ul></ul><ul><ul><li>Measuring the impact of business efforts </li></ul></ul><ul><ul><li>Estimating number of customers in the future </li></ul></ul>
    3. 3. Who Am I? Gordon S. Linoff <ul><li>Author or Co-author (with Michael Berry) of four books on Data Mining </li></ul><ul><ul><li>Data Mining Techniques for Marketing, Sales, and Customer Support, Second Edition (1998, 2004) </li></ul></ul><ul><ul><li>Mastering Data Mining (2000) </li></ul></ul><ul><ul><li>Mining the Web (2002) </li></ul></ul><ul><ul><li>Data Analysis with SQL and Excel (2008) </li></ul></ul><ul><li>Translated into Chinese, Japanese, French, and Italian </li></ul><ul><ul><li>But not (yet) Spanish or Portuguese </li></ul></ul><ul><li>Founded Data Miners, Inc. with Michael Berry in 1998 </li></ul><ul><li>Worked with many leading companies, including Pfizer, New York Times, T-Mobile, Bank of America </li></ul><ul><li>Born and raised in Miami </li></ul>Born here We are here
    4. 4. Why Do Customers Stop? <ul><li>It’s those devils of temptation </li></ul>Don’t Pay the Bill! Stick around Just Quit! Upgrade!
    5. 5. Ways to Address Voluntary Attrition <ul><li>Allow customers who are not valuable to leave. </li></ul><ul><ul><li>a cheap way to get rid of not-valuable customers </li></ul></ul><ul><li>Offer incentives to remain for a particular amount of time. </li></ul><ul><ul><li>teaser rates on credit cards, free weekend airtime, and so on </li></ul></ul><ul><li>Offer incentives to valuable customers. </li></ul><ul><ul><li>special features accessible through the Web and so on </li></ul></ul><ul><li>Stimulate usage. </li></ul><ul><ul><li>miles for minutes, donations to charities, and so on </li></ul></ul>
    6. 6. Predicting Voluntary Attrition Can Be Dangerous <ul><li>Voluntary attrition often resembles the other types. </li></ul><ul><li>Confusing voluntary and expected attrition means that you waste your marketing dollar. </li></ul><ul><li>If you confuse voluntary and forced attrition you lose twice: </li></ul><ul><ul><li>once by wasting your marketing budget </li></ul></ul><ul><ul><li>again by increasing write-offs. </li></ul></ul>
    7. 7. Ways to Address Forced Attrition <ul><li>Stop marketing to the customer. </li></ul><ul><ul><li>no more catalogs, billing inserts, or other usage stimulation </li></ul></ul><ul><li>Reduce credit lines; increase minimum payments. </li></ul><ul><li>Accelerate charge-off timeline. </li></ul><ul><li>Increase interest rates to increase customer value while they continue to pay. </li></ul>
    8. 8. Predicting Forced Attrition Can Be Dangerous <ul><li>Forced attrition often examines good customers. </li></ul><ul><ul><li>They have high balances. </li></ul></ul><ul><li>Confusing forced attrition with good customers may encourage them to leave. </li></ul><ul><li>Confusing forced attrition with voluntary attrition may hamper win-back efforts. </li></ul>
    9. 9. Visualizing Customers, The Calendar Timeline Customer Starts Customer Stops Customer Still Active
    10. 10. Predictive Response Modeling Assigns A Churn Score <ul><li>Predictions for some fixed period in the future </li></ul><ul><ul><li>Who will churn in the next month or the next six months? </li></ul></ul><ul><li>Data to build the model comes from the past </li></ul><ul><ul><li>Data to be scored (assigned a churn score) comes from the present </li></ul></ul><ul><li>Models built using data mining and statistical techniques </li></ul><ul><ul><li>decision trees, neural networks, logistic regression, regression, and so on </li></ul></ul>Jan Feb Mar Apr May Jun Jul Aug Sep Model Set Score Set 4 3 2 1 +1 4 3 2 1 +1
    11. 11. Case Study: Churn Prediction <ul><li>Competitive market </li></ul><ul><ul><li>6 carriers </li></ul></ul><ul><li>Incumbent cellular carrier with 5 million customers </li></ul><ul><li>Churn rate of about 1%/month </li></ul><ul><li>Business goal: incorporate propensity-to-churn score as part of business processes </li></ul><ul><ul><li>Churn modeling done in coordination with building data warehouse </li></ul></ul><ul><li>Chose SAS Enterprise Miner for the project </li></ul>
    12. 12. “Assign a churn score to all customers” <ul><li>This work took place over the course of two months </li></ul><ul><ul><li>during the first month, we solved a simpler problem </li></ul></ul><ul><ul><li>during the second month, we built a more general churn model </li></ul></ul>
    13. 13. Solving the Right Business Problem <ul><li>Was: “Assign a churn score to all customers.” </li></ul><ul><li>Issues </li></ul><ul><ul><li>recent customers with little call history </li></ul></ul><ul><ul><li>telephones? individuals? families? </li></ul></ul><ul><ul><li>voluntary churn vs. involuntary churn </li></ul></ul><ul><ul><li>how will the results be used? </li></ul></ul><ul><li>Became: “By 24 September, provide a list of the 10,000 Elite customers who are most likely to churn in October.” </li></ul>The new problem is much more actionable !!!
    14. 14. Technical Aspects of the Problem <ul><li>Data </li></ul><ul><ul><li>Customer information, such as billing plan, zip code, additional services, activation date and so on </li></ul></ul><ul><ul><li>Handset information, such as manufacturer, and so on </li></ul></ul><ul><ul><li>Dealer information, such as marketing region and size </li></ul></ul><ul><ul><li>Historical churn information by demographics, handset type </li></ul></ul><ul><ul><li>Billing history </li></ul></ul><ul><li>Technique </li></ul><ul><ul><li>Decision Trees </li></ul></ul><ul><li>Tools </li></ul><ul><ul><li>Oracle data warehouse </li></ul></ul><ul><ul><li>SAS Enterprise Miner </li></ul></ul>
    15. 15. A Real Life Decision Tree Example: Predicting Churn <ul><li>The training set is twice as large as the test set in this example. </li></ul><ul><li>Both have about the same percentage of churners (the numbers in black), about 13.5%. </li></ul><ul><li>As we build the tree, we want the performance on the training set and the test set to be similar. </li></ul>13.5% 13.8% 86.5% 86.2% 39,628 19,814 This side is the training set This side is the test set
    16. 16. After the first split <ul><li>After looking at all potential splits for all variables, the algorithm has chosen Handset Churn Rate for the first split. </li></ul><ul><li>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%. </li></ul><ul><li>Notice that the performance on the training set and the test set is about the same. </li></ul>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
    17. 17. 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
    18. 18. Learning from the Decision Tree <ul><li>If someone has . . . </li></ul><ul><li>no other handsets on the account, </li></ul><ul><li>and a basic family price plan, </li></ul><ul><li>and a handset churn rate on the higher side, then </li></ul><ul><li>then they are very likely to churn (60.1%). </li></ul>Some existing customers use the family plan to get discounts on new handsets.
    19. 19. Churn Model Success <ul><li>Verified that initial group of 10,000 customers were candidates for retention offers </li></ul><ul><li>Extended model to all groups of customers </li></ul><ul><li>Built churn management system </li></ul>And They Lived Happily Ever After
    20. 20. Understanding Customers Who Stop is More Than Churn Modeling <ul><li>Churn modeling is quite powerful and quite useful </li></ul><ul><ul><li>Under the right circumstances </li></ul></ul><ul><li>There are other things we want to know . . . </li></ul><ul><ul><li>What is the relationship of tenure and other covariates to stopping? </li></ul></ul><ul><ul><li>How much money are we gaining by keeping customers from stopping? </li></ul></ul><ul><ul><li>How many customers are going to stop in the future? </li></ul></ul>And hence, Survival Analysis
    21. 21. Survival Data Mining Adds the Element of When Things Happen <ul><li>Time-to-event analysis </li></ul><ul><li>Terminology comes from the medical world </li></ul><ul><ul><li>Estimating how long patients survive based on which patients recover, which patients do not </li></ul></ul><ul><li>Can measure effects of variables (initial covariates or time-dependent covariates) on survival time </li></ul><ul><li>Natural for understanding customers </li></ul><ul><li>Can be used to quantify marketing efforts </li></ul>
    22. 22. Customers on the Tenure Timeline, Rather Than the Calendar Timeline Known Customer Start Known tenure when customers stop Unknown tenure for active customers (censored)
    23. 23. How Long Will A Customer Survive? <ul><li>Survival always starts at 100% and declines over time </li></ul><ul><li>If everyone in the model set stopped, then survival goes to 0; otherwise it is always greater than 0 </li></ul><ul><li>Survival is useful for comparing different groups </li></ul>
    24. 24. Use Censored Data to Calculate Hazard Probabilities <ul><li>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 </li></ul><ul><li>h(t) = </li></ul><ul><li>Value of hazard depends on units of time – days, weeks, months, years </li></ul><ul><li>Differs from traditional definition because time is discrete – hazard probability not hazard rate </li></ul># customers who stop at exactly time t # customers at risk of stopping at time t
    25. 25. Bathtub Hazard (Risk of Dying by Age) <ul><li>Bathtub hazard starts high, goes down, and then increases again </li></ul><ul><li>Example from US mortality tables shows probability of dying at a given age </li></ul>
    26. 26. 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
    27. 27. 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
    28. 28. How Long Will A Customer Survive? <ul><li>Survival analysis answers the question by rephrasing it slightly: </li></ul><ul><ul><li>What proportion of customers survive to time t? </li></ul></ul><ul><li>Survival at time t, S(t), is the probability that a customer will survive exactly to time t </li></ul><ul><li>Calculation: </li></ul><ul><ul><li>S(t) = S(t – 1) * (1 – h(t – 1)) </li></ul></ul><ul><ul><li>S(0) = 100% </li></ul></ul><ul><li>Given a set of hazards by time, survival can be easily calculated </li></ul>
    29. 29. 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.
    30. 30. Average Tenure Is The Area Under The Curves
    31. 31. Survival to Quantify Marketing Efforts <ul><li>A company has a loyalty marketing effort designed to keep customers </li></ul><ul><li>This effort costs money </li></ul><ul><li>What is the value of the effort as measured in increased customer lifetimes? </li></ul><ul><li>SOLUTION: survival analysis </li></ul><ul><ul><li>How much longer to customers survive after accepting the offer? </li></ul></ul><ul><ul><li>How to quantify this in dollars and cents? </li></ul></ul>
    32. 32. We Can Use Area to Quantify Results <ul><li>Increase in survival is given by the area between the curves. </li></ul><ul><li>For the first year, area of triangle is a good enough estimate </li></ul>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
    33. 33. Loyalty-Responsive Customers Are Doing Better Than Others <ul><li>Survival for first year after loyalty intervention </li></ul><ul><ul><li>Group 1: 93.4% </li></ul></ul><ul><ul><li>Group 2: 85.6% </li></ul></ul><ul><ul><li>Increase for Group 1: 7.8% </li></ul></ul><ul><li>Average increase in lifetime for first year is 3.9% (assuming the 7.8% would have stayed, on average, 6 months) </li></ul><ul><li>Assuming revenue of $400/year, loyalty responsive contribute an additional revenue of $15.60 during the first year </li></ul><ul><li>This actually compares favorably to cost of loyalty program </li></ul>
    34. 34. Competing Risks: Customers May Leave for Many Reasons <ul><li>Customers may cancel </li></ul><ul><ul><li>Voluntarily </li></ul></ul><ul><ul><li>Involuntarily </li></ul></ul><ul><ul><li>Migration </li></ul></ul><ul><li>Survival Analysis Can Show Competing Risks </li></ul><ul><ul><li>overall S(t) is the product of Sr(t) for all risks </li></ul></ul><ul><li>What happens to our customers over time? </li></ul>
    35. 35. A Better Way to Look at This
    36. 36. Customer-Centric Forecasting Using Survival Analysis <ul><li>Forecasting customer numbers is important in many businesses </li></ul><ul><li>Survival analysis makes it possible to do the forecasts at the finest level of granularity – the customer level </li></ul><ul><li>Survival analysis makes it possible to incorporate customer-specific factors </li></ul><ul><li>Can be used to estimate restarts as well as stops </li></ul>
    37. 37. The Forecasting Solution is a Bit Complicated
    38. 38. 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
    39. 39. Understanding Customers Who Stop <ul><li>Is churn modeling </li></ul><ul><ul><li>To determine who is going to leave during particular periods of time and to plan marketing campaigns </li></ul></ul><ul><li>Is survival modeling </li></ul><ul><ul><li>To understand the impact of tenure and to see how business processes affect churn; and </li></ul></ul><ul><ul><li>To quantify the effects of customers stopping or not stopping </li></ul></ul><ul><li>Is forecasting </li></ul><ul><ul><li>To determine what happens to the customers over time </li></ul></ul>