Drilling Down Newsletter - July 2001 -
Customer LifeCycles, Data Mining

In this issue:
# Best Customer Retention Articles
# Tracking the Customer LifeCycle
# Practice What You Preach: Online Advertising
   Effectiveness?  Tell Me About It... (Part 3)
# Questions from Fellow Drillers
-----------------------------
Hi again folks, Jim Novo here.  This month we've got great customer retention article links and a look at customer LifeCycle measurement and tracking.  We also get deeper into measuring the true payback of online advertising, and take a popular question - what exactly goes on in data mining? 

Let's do some Drillin'!

Best of the Best Customer Retention Articles
====================
This article is on a DM News web site and will move into their paid subscription archive 30 days after the date of publication listed below, so check it out soon!  The URLs are too long for the newsletter, so the following links take you to a page you can link directly to the article from. 

Note to web site visitors: These links may have expired by the time you read this.  You can get these "must read" links e-mailed to you every 2 weeks before they expire by subscribing to the newsletter.

How Much Should Be Spent on CRM
June 26, 2001   iMarketing News
This guy Hughes has got some nerve.  He's basically saying it is downright silly to go the CRM route for marketing when you can get most of the benefits at a fraction of the cost using plain vanilla database marketing - and he has a model to prove it.  I agree  with him; that makes two of us.  Anybody else? 

And, here are three other "must read" articles which have no expiration date:

Three Keys to Ensuring CRM Success
June 28, 2001   CRMCommunity.com
Now, here's a person who knows what they are talking about! Imagine, making sure you understand customer behavior before you get involved with CRM.  Wish I'd thought of it.  Oh, and by the way, if you are looking for a framework to accomplish this understanding, may I humbly suggest my book.  You might not even need "CRM," depending on what you are trying to accomplish (increasing profits in customer marketing does not require CRM). 

Discerning Distinctions in Buying Behavior
June 25, 2001   ClickZ
Don't look now folks, but the world is catching up to us.  Track customer behavior for best results, they say.  Of course, they don't tell you how, which is the information you need to know and will find on almost every page of my site.

Making Customer Relationship
Management Work
July 5, 2001   Knowledge@Wharton
"Focusing too closely (on customers) at the individual level is a mistake" says the always excellent Peter S. Fader, supporting my own Micro vs. Macro position.  Other gems such as why "firing" unprofitable customers is just ridiculous and the related need to study customer behavior over time are covered. 

Tracking the Customer LifeCycle
=====================

Based on a national survey, 50% of marketing managers do not know their customer defection rate, and the other 50% underestimate the true defection rate.  After reading this shocking statistic, I figured it was time to do a series on customer LifeCycles, which can be used not only to track customer defection, but also to define opportunities to retain customers before they defect.

If you understand the customer LifeCycle, you can *predict* the primary defection points and react to them before customers leave you.  This is the highest ROI marketing you can possibly do, because it's cheaper than win-back (response is much higher) and preserves the investment and profits you have in the customer. 

So we're going to take a little tour though LifeCycle-based marketing land the next few issues of the newsletter.  If you're new to our Drilling Down crew, or if you have not seen the articles on the Drilling Down web site describing customer LifeCycles, you might want to brush up.  See the article on Customer LifeCycles or the tutorial Comparing the Potential Value of Customer Groups for more information on Customer LifeCycles.

At the core of a LifeCycle-based marketing approach is (shocker) customer behavior.  Customers tend to behave in certain ways unique to your business and products, and if you can discover these patterns, you can use them to predict customer behavior.  If you can predict customer behavior, you can make a ton of money marketing to your customers, because you can anticipate their behavior and take appropriate steps to try and modify it. 

Many approaches to customer marketing rely on customer behavior "triggers."  For example, a win-back program is triggered when the customer defects.  Have you switched long  distance or cellular providers lately?  Did you get inundated with win-back calls begging you to reconsider?  "Jim, we just wanted you to know we have lowered our rates."  Yea, well, thanks for telling me after over-charging me.  But could they have known I was about to switch?

Sure.  If they had looked at the calling patterns of defected customers like me, they would have seen a common thread in the behavior.  These patterns create the "trigger points" for initiating high ROI marketing campaigns before the defection.  The proper profit maximizing approach is to wait until I *look like* I'm going to defect, and then call me and offer a lower rate *before* I defect.

I would humbly submit marketing to the customer *after* they defect is a sub-optimal approach; the decision has already been made.  If you can market to them when they appear *likely* to defect, you optimize your marketing resources by not applying them too soon or too late in the customer LifeCycle.

An easy to implement and proven powerful LifeCycle trigger is called latency.  Latency refers to the average time between customer activity events, for example, making a purchase, calling the help desk, or visiting a  web site.  All you have to do is calculate the average time elapsed (latency) between the two events, and use this metric as a guide for anti-defection campaigns.  

When you see a particular customer's behavior diverge from the average customer behavior, you get a triggering event.  Since the calculation of latency is very simple, and the diverging behavior is easy to spot, this type of anti-defection campaign is an ideal candidate for "lights-out" or automated rules-based customer retention campaigns.

As an example, let's take purchase behavior in a retail scenario.  If you were to examine your customers, and find the average time between the second purchase and the third purchase was 2 months, you have found "third purchase latency."  Any customer who goes more than 2 months after the second purchase without making a third purchase is diverging from the norm, and a likely defection candidate. 

It's simple logic.  If the average customer makes a third purchase within 2 months of the second purchase, and a particular customer breaks this pattern, they are not acting like the average customer.  Something has changed.  This particular customer's LifeCycle has become out of synch with the average customer LifeCycle, and this condition is a trigger point for high ROI customer marketing.

On average, if you divert marketing resources away from customers who have made a 3rd purchase within 2 months after the second, and apply these resources to customers who are "crossing over" the 2 month LifeCycle trigger point without making a third purchase, you will end up spending less money and generating higher profits for any given marketing budget.  You are applying your limited resources right at the time in the customer LifeCycle when they create the most powerful impact - at the point of likely customer defection.

Now, will all these customers respond?  No, of course not.  But the ones that do become  active, loyal customers again, and those that don't are probably not going to be good customers in the future.  The behavior of the rest of your customers tells you so.  These non-responding customers may not be worth spending money on to "win-back," and in fact, will have much lower response rates to a win-back campaign.  They have already demonstrated their lack of interest with their behavior, and you could be better off financially by just letting them go and focusing on more responsive, more profitable customers.

The above example is a relatively crude approach to latency.  As you might suspect, different customer segments will have different latencies, and the more you fine-tune a latency campaign, the more profitable it will become.

For example, let's say you execute the latency campaign described above, and succeed in retaining 30% of the defecting customers, making a tidy profit.  But you really have two major product lines, software and hardware, each 50% of sales.  Could the latency be different between software and hardware customers?  You betcha.  Upon further analysis, you find third purchase latency for software is really one month, and for hardware it's three months.  The *average* is 2 months.  So you bust the two groups apart, and run separate latency-based campaigns, one for each product line.  

In your original third purchase latency campaign, you promoted to customers who did not make a third purchase within 2 months of the second purchase.  This means you were "late" for software (because the average latency is really 1 month) and early for hardware (because the average latency is really 3 months).  When you realign the timing based on the line of merchandise, you find instead of retaining 30% of customers, you retain 50% of the customers, because you have synched-up the marketing effort with the true customer LifeCycle more tightly.

And that, folks, is what LifeCycle-based marketing is all about - using your own customer's behavior to telegraph to you the most important (and profitable) time to market to them.  The customer, through their behavior, raises a hand and asks you to take action.  If you synch up your marketing efforts with the natural customer LifeCycle, you can't help but being more successful.

-------------------------
If you'd like to see more on LifeCycle-based marketing in future newsletters, be sure and let me know. 
-------------------------

Practice What You Preach: Online Advertising
Effectiveness?  Tell Me About It #3
=====================

OK, is Jim getting ripped off on his online advertising or not?  The only advertising I buy is highly targeted to search terms, primarily through GoTo and the Google AdWords program.  This means I get two kinds of traffic from the same search engine - paid and unpaid - for the same search terms! 

Last month, we looked at a chart comparing the value of these visitors for my top 3 search terms (relationship marketing, customer retention, customer loyalty) and did a break out of visitor value by source - paid ad or "free" search. 

By the way, in many cases both paid and free links are displayed at the same time (if I rank high enough for the search term involved).  Visitors from paid ads are clearly of better quality - higher rates of downloading, bookmarking, and newsletter subscription.  Paid ad visitors also stay twice as long on the web site.

This is a monster change from the previous analysis, which showed when looking at *all* search terms (not just the top 3), paid versus unpaid, the *free* visitors appeared to be of higher value based on their behavior. 

The implication of the above shift: there is variability in the quality of visitor generated according to the *search phrase*, and this may account for some or all the difference between the quality of a pay versus free visitor.  Intuitively, this makes sense to me, because I only pay for relevant search terms, and "free visitors" may be arriving as a result of a non-relevant search.  This is tremendously important to know, especially in light of the general industry commentary that paid search listings result in poorer search quality for users.  Hmm... 

So, let's take a closer look at search term quality by busting up the aggregate "paid" search results above by search term, and see what we get.  The following table compares each search term individually with the total site statistics, where  RM = Relationship Marketing, CR = Customer Retention, CL = Customer Loyalty, and TS = Total Site statistics.

Metric___________RM___CR___CL___TS
Avg. Visit Length     8.49   8.44   6.87   8.21
% 1 Page Visits      24%    22%   20%   43%
% Downloading     8.2%   6.1%   3.7%  3.1%
% Bookmarking    9.6%   7.6%  12.2%  5.9%
% Subscribing       4.5%   4.5%   2.4%   3.2%

Clearly, the paid ads on average generate a higher quality visitor, and there is substantial variability even among the top 3 search terms in visitor quality.  The term Customer Loyalty generates visitors with a shorter visit length and lower newsletter subscribe rate than the overall site!  But at the same time, they bookmark at much higher rates.  A bit puzzling, and whenever a behavioral marketer sees data sets with potentially conflicting indicators such as seen in the term Customer Loyalty, we know there is probably something else going on we need to find out about.

So find out we will, by Drilling Down yet another level in the next newsletter.

---------------------------
If you'd like to see more on web log analysis in future newsletters, be sure and let me know.
-------------------------------

Questions from Fellow Drillers
=====================

Q: Hi Jim,

I'm interested in the different algorithms used by the various analytical CRM vendors you mentioned on your web site. Are you, or someone you might recommend, well versed in their differences, applications, successes?

A:  Well, that's a big topic.  The vendors pretty much use the same algorithms - there are only so many approaches, and once they're coded, anybody can use them.  They may have minor differences in the way they are implemented, but particularly among the "data mining" type algorithms, they are all very closely related.  New ones come along once in awhile (the latest is called "genetic modeling") but they are all parts of the same family; each has strengths and weaknesses, depending on the data and ultimate goal.

There are really two big camps - so called "top down" modeling, where a human creates a hypothesis and tests it by building a model, and "bottoms up," which is data mining or machine learning, where the machine looks for patterns and tries to make sense of them. 

Top down models include all the pure statistical approaches, like nearest neighbor, clustering, regression, and so on.  Machine learning includes neural networks, fuzzy logic, case-based reasoning, genetic modeling, and so forth.  Algorithms like CHAID and CART are something in between; they evolved out of statistics but are also the basis for many machine learning models.

If you are really interested in descriptions of what all these things are and how they work, try these two books:

Data Mining Your Website - Jesus Mensa 

Building Data Mining Applications for CRM - co-authored by Berson, Smith, Thearling

The first is a tighter, more practical book.  The second is a monster and ties data mining more directly to CRM.  Both also briefly describe the top down or statistical approaches, and talk about some of the reasons you would use one instead of another (particularly the Mensa book, which doesn't "worship" data mining as much as the other).

Generally, to be *incrementally* more successful than a traditional statistical approach, data mining requires very clean data, a long period to "train" the application, customer records with 100's of variables, and a few Ph.D. stats people hanging around to interpret the machine language.  The training thing consists of the machine spitting out improbable answers over and over until you train it to spit out the right ones.  How do you know which are the right ones?  Frequently, from "top down" analysis done by humans. 

So there is a somewhat circular argument for data mining, and it is really best used when you have gone though all the statistical top down work first, and are looking for the "next level" of an answer.  Otherwise, you don't know if what you got out of the miner makes any sense, unless you know your business very well from a stats standpoint. 

Of course, the reason data mining became so popular as a concept in the past couple of years is the software was going to tell you everything and like magic run your business for you - no humans needed.  Turns out not to be the case, it seems.

Good top down modelers have a number of different modeling approaches at hand, and will "test" statistically to see which modeling approach provides the best "fit," that is, is the most stable over time and doesn't under or over predict an outcome.  The most common stats packages, some of which have evolved to the point where you can pretty much run them out yourself without a stats background, are from SPSS and SAS, which have naturally made forays into data mining as well.  In fact, many of the CRM packages that "deliver" data mining capabilities really deliver SPSS or SAS.

Which brings us to RFM, the original behavioral model and the topic of my site and book.  When you run all this stuff above looking for response or future value models, Recency and Frequency will always factor highly into the result, no matter whether you go with top down stats or bottoms up machine analysis.

The Recency and Frequency variables are so embedded into human nature they always end up in any model predicting behavior.  What you get with all these other models, top down or bottom up, is the 3rd, 4th, 5th etc. most powerful predictors, with diminishing predictive power at each level.  Said another way, Recency and Frequency will give you an 80% accurate model.  Add a 3rd variable and you get 85%.  Add a 4th and you get 88%. Add a 5th and you get 90%, and so on. 

That's why I tell people who have never done any customer modeling before this data mining stuff is like trying to get a Ph.D. without ever going to high school.  It's overkill in this situation, and is most useful only if you have gone through basic stat models first.  In fact, the result of a regression model or other statistical approach is one of the best data sets you can feed a mining engine.  Likewise, any good human stat person, before they build you a regression model, will ask for the results of RFM testing to help build the model. 

It's like a pyramid, with RFM at the bottom, statistics in the middle level and data mining on the top.  The most powerful, most broadly  applicable models providing the highest immediate impact, the so-called "low hanging fruit," are RFM-based.  Statistics adds further refinement, targeting even finer shades of behavior.  Data Mining figures out if there is anything left to predict; it can produce segments you would never have found otherwise, although there may only be 1/8 of 1% of your customers in these segments, **particularly** if you have gone through RFM and stats first.

The original RFM is a static model, predicting behavior at a point in time.  In my book, you learn how to convert the model to look at behavior over time, a much more powerful, LifeCycle-oriented tool.  Any company considering adding customer analytics, whether CRM-related or not, should go through the process of scoring customers using simple behavior-based models and trying their hand at high ROI marketing programs first.  If you do this before shelling out the money for a data mining package, you'll be more likely able to justify the ROI and figure out if data mining will really help you. 

Want to do a quick test for potential ROI on customer analytics?  See this article:

http://www.jimnovo.com/Pre-CRM.htm

Will any CRM consultant or software vendor tell you all this?  Nah.  That's why I wrote the book, don't you know...

---------------------------

That's it for this month's edition of the Drilling Down newsletter.  If you like the newsletter, please forward it to a friend!  Subscription instructions are at the top and bottom. 

Any comments on the newsletter (it's too long, too short, topic suggestions, etc.) please send them right along to me, along with any other questions on customer Valuation, Retention, Loyalty, and Defection, to me. 

'Til next time, keep Drilling Down!

Jim Novo

Copyright 2001, The Drilling Down Project by Jim Novo.  All rights reserved.  You are free to use material from this newsletter in whole or in part as long as you include complete attribution, including live web site link and/or e-mail link. Please tell me where & when the material will appear.