Predicting Customer Lifetime Value
Skip to content
Marketing Feb 1, 2008

Pre­dict­ing Cus­tomer Life­time Value

When is it sen­si­ble to give perks to customers?

Based on the research of

Edward Malthouse

Robert C. Blattberg

Pre­dic­tion is very dif­fi­cult, espe­cial­ly about the future.”

Add Insight
to your inbox.

We’ll send you one email a week with content you actually want to read, curated by the Insight team.

While Nobel Lau­re­ate physi­cist Niels Bohr prob­a­bly had sub­atom­ic, quan­tum mechan­i­cal phe­nom­e­na in mind when he made that state­ment, the same could be said for cus­tomer behav­iors. Or so thought Robert Blat­tberg, Pro­fes­sor of Mar­ket­ing at the Kel­logg School of Man­age­ment, and Edward Malt­house, Asso­ciate Pro­fes­sor of Inte­grat­ed Mar­ket­ing Com­mu­ni­ca­tions at Northwestern’s Medill School of Jour­nal­ism, in their Jour­nal of Inter­ac­tive Mar­ket­ing paper describ­ing the sur­pris­ing uncer­tain­ty inher­ent in dif­fer­en­tial­ly mar­ket­ing to cus­tomers based upon their past performance.

As Blat­tberg put it, There’s just more serendip­i­ty in behavior.”

For their efforts to bet­ter under­stand long-term cus­tomer behav­ior and val­ue, Malt­house and Blat­tberg were award­ed the journal’s Best Paper of 2005.”

Ulti­mate­ly, I’m inter­est­ed in dri­vers of long-term cus­tomer behav­ior,” said Blat­tberg, the Polk Broth­ers Dis­tin­guished Pro­fes­sor of Retail­ing, and Direc­tor of the Cen­ter for Retail Man­age­ment. I study this by look­ing at sales data­bas­es, not by study­ing it in a lab.”

Mar­keters attempt to pin­point the best cus­tomers with the help of tools such as RFM analysis.

Malt­house and Blat­tberg are espe­cial­ly inter­est­ed in behav­ior that deter­mines cus­tomers’ long-term val­ue — that is, the expect­ed ben­e­fits to a com­pa­ny after tak­ing into account the expect­ed costs of main­tain­ing a rela­tion­ship with a cus­tomer over an extend­ed peri­od of time. A valu­able rela­tion­ship, like any oth­er com­pa­ny asset, mer­its prop­er invest­ment and man­age­ment. Many com­pa­nies divert sig­nif­i­cant atten­tion and resources to mar­ket­ing tech­niques that seek to cul­ti­vate loy­al­ty among the per­ceived best customers.

Com­pa­nies are focused on reward­ing the best cus­tomers, because they think they’ll con­tin­ue to be the best cus­tomers,” said Blattberg.

If you have been a great cus­tomer, cat­a­log com­pa­nies may send you more mail­ings, air­lines may give you pri­or­i­ty for upgrades, cred­it card com­pa­nies may waive late fees, and hotels may leave a bot­tle of caber­net in your suite. These are dis­cre­tionary mar­ket­ing invest­ments, bestowed upon unsus­pect­ing cus­tomers and intend­ed to cul­ti­vate con­tin­ued, prof­itable busi­ness among the best cus­tomers. At the heart of this prac­tice lies the belief that a good cus­tomer yes­ter­day will be a good cus­tomer tomorrow.

But is it pos­si­ble to pre­dict which cus­tomers will be best over the long haul, and to do so reli­ably enough to jus­ti­fy giv­ing them the white glove, five-star treat­ment? Malt­house and Blat­tberg set out to answer these ques­tions. Said Blat­tberg, This paper focused on whether past best cus­tomers remain future best cus­tomers: Are past heavy users of X future heavy users of X?”

Mar­keters attempt to pin­point the best cus­tomers with the help of tools such as RFM analy­sis. This tech­nique mea­sures how recent­ly cus­tomers made pur­chas­es (R, recen­cy), how often they made pur­chas­es (F, fre­quen­cy), and how much they spent (M, mon­e­tary val­ue). If the axiom stat­ing that 80 per­cent of your busi­ness comes from 20 per­cent of your cus­tomers” is true, then RFM analy­sis can help iden­ti­fy who the best 20 per­cent of cus­tomers have been. But does it allow com­pa­nies to pre­dict whether the top 20 per­cent from the past will bring long-term advan­tage in the future?

The lit­er­a­ture on cus­tomer data­bas­es focused on RFM analy­sis looks at cus­tomers’ propen­si­ty to make their very next pur­chase,” said Blat­tberg. But it doesn’t auto­mat­i­cal­ly fol­low that they’ll con­tin­ue as best cus­tomers over the long term.”

To mea­sure and attempt to pre­dict cus­tomers’ long-term val­ue, Malt­house and Blat­tberg eval­u­at­ed reams of sales data col­lect­ed by var­i­ous com­pa­nies over many years. The scope of their data was immense. They stud­ied over 136,000 donors to a not-for-prof­it orga­ni­za­tion dur­ing a two-year peri­od, over 71,000 cus­tomers of a ser­vice provider dur­ing a five-year span, over 41,000 cus­tomers of a spe­cial­ty cat­a­log com­pa­ny over a twelve-year peri­od, and over 24,000 small busi­ness­es that were cus­tomers of a larg­er com­pa­ny dur­ing a sev­en-year span. The data­bas­es described a wide range of cus­tomer behav­iors, such as length of con­tract, amount of ser­vice usage, and the date, price, type, and val­ue of pur­chas­es. How­ev­er, none of the com­pa­nies gave pref­er­en­tial perks to their cus­tomers, so Malt­house and Blat­tberg were able to gain a more objec­tive view of cus­tomers’ long-term behav­ior and value.

By arti­fi­cial­ly divid­ing the data into past, present, and future time peri­ods, Malt­house and Blat­tberg were able to build sta­tis­ti­cal mod­els and draw con­clu­sions that enabled them to accom­plish their cen­tral objec­tive: trans­form­ing past pur­chas­ing data into pow­er­ful pre­dic­tors of long-term cus­tomer pur­chas­ing behav­ior and val­ue to a company.

Blat­tberg and Malt­house picked points in time near the mid­dle of each data set and deemed these points to be now.” For exam­ple, using a data­base that spanned twelve years end­ing on July 31, 1995, they defined now” as August 11990.

Hav­ing estab­lished a now” in each data set, Malt­house and Blat­tberg in effect cre­at­ed pasts” (August 1, 1983 to July 31, 1990) and futures” (August 2, 1990, to July 311995).

By reori­ent­ing time, Malt­house and Blat­tberg could build, test, and extrap­o­late from sta­tis­ti­cal mod­els. Data on cus­tomers’ past behav­ior and val­ue were fed” into sta­tis­ti­cal regres­sion equa­tions. These equa­tions cal­cu­lat­ed which com­bi­na­tions of behav­ioral fea­tures were most relat­ed to val­ue. For exam­ple, a regres­sion equa­tion might have shown that cus­tomers’ past val­ues to the not-for-prof­it orga­ni­za­tion were most direct­ly affect­ed by the amount of the cus­tomers’ most recent dona­tions and the lengths of time that cus­tomers’ had been mem­bers of the organization.

Once the mod­els had been fine-tuned to relate past behav­ior to val­ue using the train­ing data,” they were fed hold­out” data from a dif­fer­ent set of cus­tomers of the orga­ni­za­tion. The mod­els drew on the rela­tion­ships they had estab­lished from the train­ing” data — link­ing behav­ior and val­ue — to pre­dict” future val­ue when giv­en only future behav­ior. Once the mod­els had pre­dict­ed cus­tomers’ future val­ues, Malt­house and Blat­tberg com­pared the pre­dict­ed val­ues with the val­ues that the cus­tomers actu­al­ly attained. The cus­tomers who were among the most valu­able 20 per­cent were con­sid­ered best.”

Since they knew the mod­els would not be capa­ble of per­fect­ly cat­e­go­riz­ing best and non-best future cus­tomers based on their past behav­ior, Malt­house and Blat­tberg eval­u­at­ed the strength of their mod­els by mea­sur­ing the two pos­si­ble types of clas­si­fi­ca­tion errors that they could make. To a mar­ket­ing man­ag­er, these errors trans­late to spend­ing mon­ey on cus­tomers you should not spend mon­ey on, and not spend­ing mon­ey on cus­tomers you should. To a sta­tis­ti­cian, these errors are called false pos­i­tives and false negatives.

Across all the mod­els and data sets, whether they pre­dict­ed cus­tomers’ val­ues from one to six years into the future, the pat­terns were remark­ably consistent.

A false pos­i­tive occurred when the mod­el pre­dict­ed that a cus­tomer would be a future best cus­tomer when in fact the cus­tomer was not. For exam­ple, a cus­tomer who used to con­duct quite a bit of busi­ness with a com­pa­ny was pre­dict­ed to con­tin­ue to be a great cus­tomer in the future. How­ev­er, if the cus­tomer lost her job, she might not have been able to spend as much, and thus did not actu­al­ly con­tin­ue to be among the best customers.

Sim­i­lar­ly, false neg­a­tives occurred when cus­tomers were not pre­dict­ed to be espe­cial­ly valu­able, even though their future pur­chas­ing behav­ior proved to be extreme­ly ben­e­fi­cial to the com­pa­ny. For exam­ple, a cus­tomer who did lit­tle or no busi­ness in the past due to lack of dis­pos­able income was pre­dict­ed to be a poor cus­tomer in the future. But if the cus­tomer took a new, lucra­tive job, he could actu­al­ly become a poten­tial gold mine for the company.

Malt­house and Blat­tberg were stunned by the pat­terns of false pos­i­tives and neg­a­tives that they observed. Across all the mod­els and data sets, whether they pre­dict­ed cus­tomers’ val­ues from one to six years into the future, the pat­terns were remark­ably con­sis­tent. So con­sis­tent, in fact, that Blat­tberg and Malt­house pro­posed two new, empir­i­cal rules.

They referred to the first insight as the 20 – 55 rule: of the actu­al top 20 per­cent of future cus­tomers, rough­ly 55 per­cent will be mis­clas­si­fied as poor or aver­age cus­tomers, and thus will not receive spe­cial treat­ment. This false neg­a­tive rate was strik­ing­ly con­sis­tent across dif­fer­ent mod­els and data sets, rang­ing from 51 per­cent to 55 percent.

They dubbed the oth­er rule the 80 – 15 rule: of the actu­al bot­tom 80 per­cent of future cus­tomers, rough­ly 15 per­cent will be mis­clas­si­fied and will receive spe­cial treat­ment. This false pos­i­tive rate was also remark­ably con­sis­tent across data sets, rang­ing from around 13 per­cent to 15 percent.

The ram­i­fi­ca­tions of this work were clear. Rough­ly a quar­ter of all cus­tomers were mis­clas­si­fied, false­ly pos­i­tive or neg­a­tive. A com­pa­ny that makes mar­ket­ing deci­sions that are mis­guid­ed one out of every four times could gain con­sid­er­ably by recon­sid­er­ing its tar­gets in a more sophis­ti­cat­ed way. That is, if it sur­vives long enough to do so.

We found that best cus­tomers con­tin­ued to be best cus­tomers at a much low­er rate than we expect­ed,” said Blat­tberg. If a sig­nif­i­cant pro­por­tion of future best cus­tomers comes from past poor cus­tomers, you risk los­ing them. As soon as you dif­fer­en­ti­ate cus­tomers you face this problem.”

Malt­house and Blat­tberg used these results to devise a fair­ly straight­for­ward for­mu­la that could help man­agers deter­mine when it would be sen­si­ble to give perks to cus­tomers. The for­mu­la depend­ed on four vari­ables: the cost of giv­ing a perk; the cost of alien­at­ing cus­tomers by not giv­ing perks to those who deserve them; the extra prof­it gained from loy­al cus­tomers to whom perks were giv­en; and the extra prof­it gained from cus­tomers who were pleas­ant­ly sur­prised to receive perks even though they did not actu­al­ly deserve them.

To illus­trate these find­ings, Blat­tberg turned to the skies (and tick­et coun­ters and con­gest­ed ter­mi­nals). Because of the reward struc­ture, low volume/​less good cus­tomers might nev­er become good cus­tomers,” he said. For exam­ple, I fly a lot. But since I live in Chica­go, I’m not going to fly Delta very often. So I’m nev­er going to be a good cus­tomer for them, and I’m going to get seats in the mid­dle, I’m going to have to wait in long lines. But what if I move to Atlanta? I’m still going to fly a lot, so I could poten­tial­ly become a great cus­tomer for Delta. But since they didn’t treat me too well for all those years in Chica­go, they risk los­ing me.”

Then con­sid­er South­west,” he con­tin­ued. They take a more egal­i­tar­i­an approach: We don’t care who you are, you get your board­ing pass based on when you check in. Which is bet­ter? You might not get treat­ed spe­cial by South­west, but that still might be bet­ter than get­ting treat­ed poor­ly by Delta.”

Malt­house and Blat­tberg posed a sim­ple solu­tion: allo­cate rewards based on actu­al future behav­ior rather than pre­dict­ed future behav­ior. The dis­tinc­tion was sub­tle but impor­tant. Rather than try­ing to guess which cus­tomers would be valu­able, Malt­house and Blat­tberg encour­aged com­pa­nies to wave a car­rot in front of all of their cus­tomers and reward those who behaved in the desired way. This is what air­lines do with miles pro­grams— any cus­tomer who flies a cer­tain num­ber of miles gets the reward. Cred­it cards and super­mar­kets that offer cash-back bonus­es also fol­low this approach.

This research would be less rel­e­vant for com­pa­nies that do not main­tain and use data­bas­es of cus­tomer infor­ma­tion — for exam­ple, sell­ers of can­dy bars, toi­let paper, and most oth­er con­sumer pack­aged goods. Nor would the study of such perks real­ly apply to com­pa­ny-cus­tomer rela­tion­ships in which future rewards are clear­ly stat­ed at the time of pur­chase. For exam­ple, rental car com­pa­nies may say up front, Rent four days, get the fifth free,” influ­enc­ing cus­tomers from the out­set with clear­ly defined incen­tives. How­ev­er, busi­ness­es that opt to sur­prise” their cus­tomers hop­ing to instill loy­al­ty have much to gain from this research, and from research that could follow.

There are lots of inter­est­ing future ques­tions,” said Blat­tberg. Could you cre­ate reward sys­tems that increase the odds that best cus­tomers con­tin­ue to be best cus­tomers? What kind of reward struc­ture would that be? Do seg­ments of cus­tomers become dor­mant and come back, or do they die forever?”

Were he alive today, might Niels Bohr have been drawn to mar­ket­ing strate­gies rather than elec­tron orbitals? It is dif­fi­cult to say. As dif­fi­cult, per­haps, as try­ing to wres­tle ran­dom quirks of human con­sump­tion habits into mechan­i­cal­ly pre­dictable busi­ness processes.

About the Writer

Dr. Brad Wible (Northwestern University, The Graduate School, 2004) is a Senior Program Associate with the Research Competitiveness Program, a Science and Policy Program at the American Association for the Advancement of Science. He lives in Washington, DC.

About the Research

Malthouse, Edward and Robert Blattberg (2005). “Can We Predict Customer Lifetime Value?” Journal of Interactive Marketing, 19(1): 2-16.

Suggested For You

Most Popular

Organizations

How Are Black – White Bira­cial Peo­ple Per­ceived in Terms of Race?

Under­stand­ing the answer — and why black and white Amer­i­cans’ respons­es may dif­fer — is increas­ing­ly impor­tant in a mul­tira­cial society.

Leadership

Why Warmth Is the Under­ap­pre­ci­at­ed Skill Lead­ers Need

The case for demon­strat­ing more than just competence.

Most Popular Podcasts

Careers

Pod­cast: Our Most Pop­u­lar Advice on Improv­ing Rela­tion­ships with Colleagues

Cowork­ers can make us crazy. Here’s how to han­dle tough situations.

Social Impact

Pod­cast: How You and Your Com­pa­ny Can Lend Exper­tise to a Non­prof­it in Need

Plus: Four ques­tions to con­sid­er before becom­ing a social-impact entrepreneur.

Careers

Pod­cast: Attract Rock­star Employ­ees — or Devel­op Your Own

Find­ing and nur­tur­ing high per­form­ers isn’t easy, but it pays off.

Marketing

Pod­cast: How Music Can Change Our Mood

A Broad­way song­writer and a mar­ket­ing pro­fes­sor dis­cuss the con­nec­tion between our favorite tunes and how they make us feel.