Replies to This Discussion

Permalink Reply by Ralph Winters on November 17, 2010 at 5:28pm

BR, No there shouldn't be. If there are, I would also be interested in seeing them. These are all different statistical techniques that exist only in the purest sense. They are not tied to specific applications. Often the choice is up to the modeler, and what you have expertise in doing. It is also dependent upon the number of degrees of freedom in each model, and the model assumptions you are willing to make. Typically for a churn problem I would use a logistic model, and support it with decision tree analysis. But a lot of people nowadays are doing churn using survival analysis.

-Ralph Winters

• ▶ Reply

Permalink Reply by Tom Wolfer on November 18, 2010 at 5:41am

BR. One point to follow-up on what Vincent mentioned. Does the company for whom you are doing this analysis have a customer service department? If so, you may be able to develop a bit of picture of why customers are leaving based on comments that they are giving. Even better, if your customer service reps (or sales reps) aren't asking people the reason for their leaving (either when they leave, or, as a follow-up after they do), then they should be. If churn is a major concern, as it sure is here in Canada with all the new telco entrants, then collecting this information on the fly should be a priority and will greatly help you in your modeling efforts.

Yes, I would also agree with Ralph in that logistic regression supported by an inductive decision tree to verify the model findings is the right course of action. The tree will also be valuable in helping those for whom you are doing the model to interpret the results in plain English. One more note about the modeling exercise: ensure that the historical data that is used to predict churn (however you define this) takes into consideration any lag time that there is between when data is collected, cleaned, stored, and ready for use at any given point in time for modeling.

• ▶ Reply

Permalink Reply by Idielle Walters on November 18, 2010 at 8:19am

For involuntary I would say analyze the billing data and for voluntary churn make sure you include callcenter data. Performing text mining on call center data will definately give you a lot of insight, but unfortunately this data is not always captured with a high level of accuracy by the call center agents. You should also take note of new promotions by competitors and analyze the influence of these promotions on your revenue and churn.

• ▶ Reply

Permalink Reply by Tom Wolfer on November 18, 2010 at 1:44pm

Idielle does make a great point. However, at the very least, you can look at the call centre data much the same as you would the feedback from a focus group: not representative, but an excellent guide with regards to key concepts that should me measured/reported on in customer surveys or a more rigorous data capture program. This data would then be used for modeling.

• ▶ Reply

Permalink Reply by Diana Mathew on November 19, 2010 at 3:21am

Hi! This is such a great article and I am sure a lot of money saving enthusiasts are going to benefit from this. Keep it up! I am Diana Mathew, an Australian Entrepreneur, ebook author (The Money Tree by Diana Mathew) and a Saving Money guru.
If you have time, maybe you can visit me too:
http://www.mymoneytree.com.au

• ▶ Reply

Permalink Reply by Tom Wolfer on November 19, 2010 at 6:20am

Hello Diana. Thank you for this. Does your publication address ways in which to measure churn, or does it cover practical ways in which churn may be reduced on an operational level?

• ▶ Reply

• ‹ Previous
• 1
• 2
• Next ›
• Page