00:00:05
well one of the most established
00:00:06
practices in the field of marketing and
00:00:08
customer valuation is to summarize a
00:00:11
customer using what's called rfm
00:00:13
segmentation which means I take
00:00:15
everything I know about my customer and
00:00:17
I compute just three simple numbers how
00:00:19
recently did they buy when's meaning
00:00:21
when's the last time they bought how
00:00:23
frequently do they buy meaning the
00:00:25
number of time periods in which they
00:00:26
bought and when they buy how much money
00:00:28
do they spend that's called r FM
00:00:30
segmentation it's the basis of the way
00:00:32
in which most companies decide who are
00:00:34
the valuable customers and who are the
00:00:36
non-valuable customers and my research
00:00:38
is very simple it basically says that's
00:00:41
not a complete characterization of
00:00:43
customers you have to add one more
00:00:45
letter to rfm and I call that c which
00:00:47
means clumpiness which means some
00:00:49
customers do Buy in a regular pattern
00:00:52
and historically if you bought orange
00:00:54
juice if you bought diapers you bought
00:00:56
things in a regular pattern but
00:00:57
clumpiness refers to the fact that
00:00:59
people buy and burst and those burst
00:01:02
periods indicate something very
00:01:04
different about the customer and that
00:01:06
those customers could be extremely
00:01:12
valuable the key takeaways of my
00:01:14
research is very simple let's imagine
00:01:16
you want to build a simple what I call a
00:01:18
simple mathematical model you want to
00:01:20
predict who are going to be the valuable
00:01:22
customers in the future and you have
00:01:24
four things you can use to predict it as
00:01:26
I mentioned recency frequency monetary
00:01:29
value and let let's say the marketing
00:01:30
spend towards the customer those are the
00:01:33
classic ways in which companies build
00:01:35
what are called scoring models I'm
00:01:36
claiming you need to add one more number
00:01:39
and that's see how clumpy the customer
00:01:41
is this is no more difficult to compute
00:01:43
than rfm you can do it in Excel it's
00:01:46
very quick to compute you can compute it
00:01:47
for literally a 100 million customers in
00:01:50
a second and the findings of my research
00:01:52
suggest that higher clumpy customers are
00:01:56
worth more out of sample meaning in
00:01:58
their future value even after
00:02:01
controlling for rfm and marketing
00:02:03
expender which means we have found
00:02:06
another variable that firm should track
00:02:08
about customers and use it to predict
00:02:10
their worth in the
00:02:15
future two things surprised me about my
00:02:17
conclusions one is um I just figured
00:02:21
that this rfm based segmentation which
00:02:23
had been around for so long and is used
00:02:25
by so many firms had been validated in
00:02:28
the sense that there wasn't any anything
00:02:30
else simple out there that could help
00:02:32
explain customer value I mean you can do
00:02:34
all kinds of fancy web scraping and all
00:02:36
kinds of other variable construction but
00:02:38
clumpiness is so simple so first I was
00:02:41
surprised that that had been missed that
00:02:44
in other words hot and cold periods are
00:02:46
indicative of something about the
00:02:48
customer I think the second part that
00:02:50
surprised me is that at least the data
00:02:52
sets I've analyzed it's true for let's
00:02:55
call it digital and online consumption
00:02:57
Goods but it's not true for regular
00:03:01
consumer package Goods in other words
00:03:04
historical models I can see why they fit
00:03:07
fine because you buy toilet paper in a
00:03:09
regular pattern you buy orange juice in
00:03:12
a regular pattern but you don't consume
00:03:14
Hulu in a regular pattern you don't bid
00:03:16
on auctions at eBay at a regular pattern
00:03:18
you don't buy books at Amazon on a
00:03:20
regular based pattern so I think the two
00:03:22
things that surprised me is number one
00:03:24
that people had missed what seemed to me
00:03:26
to be something fairly obvious and
00:03:28
secondly that it applied
00:03:30
historically if you look at historically
00:03:32
purchased Goods clumpiness really isn't
00:03:35
there but if you look in the new wave
00:03:36
the new economy clumpiness is pervasive
00:03:39
in every data set I've
00:03:44
analyzed I think of all the research
00:03:46
I've done over my career which is now
00:03:48
it's hard for me to believe but it's
00:03:49
been a 20-year career I think this is
00:03:52
probably the most practical thing I've
00:03:53
done um the work I do tends to be what I
00:03:56
call fancy complex statistical modeling
00:03:59
and this is n about statistical modeling
00:04:01
this is about a number clumpiness that
00:04:04
firms can actually compute today they
00:04:06
don't need to collect any additional
00:04:08
data it's the same data they're using to
00:04:10
compute R F&M and customer lifetime
00:04:12
value and they can figure out how much
00:04:15
value it adds to predicting customer
00:04:17
value your rank ordering of customers
00:04:19
will change your decisions about which
00:04:21
customers are valuable to reactivate
00:04:24
imagine customers have churned well
00:04:26
which ones are valuable to reactivate my
00:04:28
claim is the clumpy ones even though
00:04:30
they've churned those are the ones to
00:04:32
reactivate cuz if you reactivate them
00:04:34
they'll come back and be clumpy again
00:04:35
and do a lot of stuff in the future so I
00:04:37
think it has huge practical value and
00:04:39
the beauty of it is if you go to my
00:04:42
website I have an Excel sheet there that
00:04:44
has worked out examples it actually has
00:04:46
an Excel sheet that you can just
00:04:48
download and you can start using
00:04:50
clumpiness
00:04:55
today a lot of people have today talked
00:04:57
about big data and I'm actually I love
00:05:01
big data but I'll tell you what I love
00:05:02
even more than big data I love data
00:05:04
compression and what I mean by data
00:05:06
compression is you can collect thousands
00:05:09
and thousands of variables on people now
00:05:11
you can track where they are you can
00:05:13
track what they bought what web pages
00:05:14
they looked at but that's not science
00:05:18
that's a that's data collection now the
00:05:20
question is which of that information is
00:05:22
actually useful for the business problem
00:05:24
at hand and that's what I call data
00:05:26
compression so the way I view clumpiness
00:05:29
as an addition to traditional variables
00:05:31
like rfm marketing activity and stuff
00:05:33
like that I Vis I view it as a form of
00:05:36
let's call it increased data compression
00:05:38
I'm just telling you you need to cover
00:05:41
you need to keep a little bit more data
00:05:43
you can't compress things down to three
00:05:44
numbers you got to compress it down to
00:05:46
four so I view This research that I'm
00:05:48
doing as kind of I view it in the data
00:05:50
compression world I love the problem of
00:05:53
taking big data and compressing it down
00:05:55
to small data and that's how I view
00:05:58
clumpiness
00:06:02
the part that's unknown to me right now
00:06:05
which is I've done a lot of work on
00:06:06
clumpiness I know it exists across
00:06:08
Industries I know it exists I know it
00:06:10
can be predict of predictive value
00:06:13
here's what I don't know what causes it
00:06:16
so what I do know is I've related
00:06:18
marketing activity to clumpiness so
00:06:20
firms can try to make you clumpy by
00:06:22
sending you an email by sending you a
00:06:23
catalog by targeting you Etc that much I
00:06:26
know but I haven't really studied yet
00:06:29
what's the optimal way in which firms
00:06:31
should targeting you target you knowing
00:06:33
that clumpiness exists I haven't looked
00:06:34
at like for example do you consume more
00:06:36
clumpy content if it's a series like
00:06:39
imagine watching Breaking Bad or Mad Men
00:06:41
or something like that or imagine you're
00:06:43
a firm and you're trying to sell a sweet
00:06:45
of products like you know a facial care
00:06:47
line and a you know moisturizer line and
00:06:50
all this other stuff should you package
00:06:52
it together and make it seem like people
00:06:54
are progressing towards a goal so here's
00:06:57
what I do know I know mathematically how
00:06:58
to compute it I know it's trivial for
00:07:01
firms to do I know it's predictive but
00:07:03
the part that's left unknown to me is
00:07:05
the psychology of why which is why I'm
00:07:07
partnering right now with a lot of my
00:07:09
more consumer psychology oriented
00:07:11
colleagues we're going to start running
00:07:12
a lot of Behavioral experiments in the
00:07:14
lab to try to get to the under
00:07:16
underlying psychological underpinnings
00:07:19
of why people behave in a clumpy
00:07:24
fashion I'm not sure I've seen much
00:07:28
about clumpiness if except what you see
00:07:30
is you see stories in the news all the
00:07:32
time about people kind of binging on
00:07:35
content and so I like the word
00:07:37
clumpiness other people like the word
00:07:39
binging um the reason I like clumpiness
00:07:42
is that it refers to you know the
00:07:44
opposite which is non- clumpy which is
00:07:46
kind of equally spaced kind of arrivals
00:07:49
or equally spaced purchases so what I
00:07:52
would say is I don't think I've seen a
00:07:54
story about clumpiness but anytime you
00:07:56
see a story about people binging content
00:07:59
people consuming things you know a
00:08:01
student sat up for 18 hours watching
00:08:03
this it applies and the concept is so
00:08:07
pervasive and every time I talk to
00:08:08
whether it's managers students academics
00:08:11
about it everyone believe it exists the
00:08:13
part as you mentioned in your earlier
00:08:15
question that shocks people is that it's
00:08:18
actually predictive of customer
00:08:23
value I think it dispels the idea that
00:08:27
in some sense um customers can just be
00:08:30
categorized by a simple set of numbers
00:08:32
um you need to go a little bit beyond
00:08:34
that you need to go a little bit beyond
00:08:36
what I would call Simple theories of how
00:08:39
people behave and what clumpiness if you
00:08:41
actually think about what clumpiness
00:08:42
says is if you look at recency frequency
00:08:45
monetary value which is kind of the
00:08:47
historical basis of consumer Behavior it
00:08:50
basically ignores what I call the inter
00:08:52
Ral times it basically says I can take
00:08:55
all the data like if a two-day P it was
00:08:57
a two-day window and then a 4-day window
00:08:59
then a a three-day window then a sixday
00:09:00
window I can throw all of that away and
00:09:02
all I need to know is when's the last
00:09:04
time you came and how many times did you
00:09:06
come but what this dispels is that the
00:09:09
arrival pattern of people is
00:09:12
uninformative it's very informative
00:09:14
people that come and burst then go away
00:09:16
and then come back and burst and then go
00:09:18
away it's just I think those are just
00:09:20
different types of people I think those
00:09:22
people are fundamentally different I
00:09:25
personally believe there are clumpy type
00:09:27
people and non- clumpy type people
00:09:29
however what we've also shown is it
00:09:31
varies by product categories so we found
00:09:34
for example that women tend to be more
00:09:35
clumpy than men we found that younger
00:09:38
people tend to be more clumpy in their
00:09:39
consumption than older people so I think
00:09:41
what's really going to dispel I think
00:09:43
what the myth that we're going to dispel
00:09:45
is that like not only are all people
00:09:47
created equal but that there are simple
00:09:49
ways to just categorize all people into
00:09:52
a certain
00:09:57
type I think what everyone's done is
00:09:59
there's a whole class of mathematical
00:10:01
models that have been popularized
00:10:03
although they've been around for 50
00:10:04
years but have been popularized over the
00:10:06
last 10 years called hidden Markov
00:10:08
models the idea is very simple let's
00:10:10
imagine there are two states of the
00:10:12
world you're in a hot state or a cold
00:10:14
State and you rotate back and forth
00:10:16
between a hot and a cold state that
00:10:18
mathematical model is clumpiness you're
00:10:21
hot you do a lot of stuff you're cold
00:10:23
you don't hot cold hot cold what TP what
00:10:26
separates this work is the work I'm
00:10:28
doing isn't Ivory Tower mathematics it's
00:10:32
a simple number that someone can compute
00:10:36
so I fit hidden Mark of models to data
00:10:39
what I wanted to do was to bring to the
00:10:41
practitioner a way that they could
00:10:43
compute a simple number it's a statistic
00:10:46
it's not a statistics paper it's a paper
00:10:49
about a number a statistic as we call it
00:10:51
you just compute the number and then do
00:10:54
what you want with it you could try to
00:10:55
use it to predict customer value you
00:10:57
could use it to see are men more than
00:10:59
women you could use it to segment people
00:11:02
that's what typifies and separates this
00:11:04
work is that it's a simple metric based
00:11:08
approach that practitioners can use it's
00:11:11
not a fancy modeling based approach but
00:11:13
they're both trying to cover the same
00:11:19
problem what we've studied so far with
00:11:21
reaching clumpy customers is we've
00:11:23
studied whether email cataloges
00:11:26
different types of marketing channels
00:11:28
are more effective and what we found not
00:11:30
surprisingly is email has more of a
00:11:32
short-term effect as you would expect
00:11:34
catalog has more of a longer term effect
00:11:37
what we've yet to really understand is
00:11:40
are there certain words in a email or
00:11:43
catalog or you know video campaign that
00:11:46
will engage or you know if you'd like
00:11:48
cause people to be more clumpy are there
00:11:50
certain topics that are more clumpy more
00:11:52
some product categories that will
00:11:54
necessarily be more clumpy all we've
00:11:56
done so far is I think I've established
00:11:58
the phenomenon on exists I know it
00:12:00
exists across lots of Industries I know
00:12:03
certain types of people tend to be more
00:12:05
clumpy the part that I haven't done
00:12:06
which is shocking because I'm a
00:12:07
professor of marketing is talk about the
00:12:10
marketing implications of it yet that's
00:12:12
going to require bigger and newer data
00:12:14
sets that allow me to link things about
00:12:17
marketing campaigns to people's clumpy
00:12:19
Behavior I know how to do it it's just I
00:12:21
need richer and better data to do
00:12:27
it I'm thinking about three different
00:12:29
streams to follow up this clumpy
00:12:31
research first of all I'd be thrilled to
00:12:34
just analyze more data sets and prove
00:12:36
how pervasive the clumpiness measure is
00:12:38
so I've analyzed data sets from Amazon
00:12:41
from CD Now from eBay from Hulu from
00:12:44
YouTube and also from some traditional
00:12:47
consumer package Goods companies look if
00:12:49
PNG wants to contact me tomorrow I'm
00:12:51
happy to apply clumpiness to their work
00:12:53
if Google wants to contact me tomorrow
00:12:55
if Goldman Sachs wants to contact me
00:12:58
tomorrow if fizer wants to contact me
00:13:00
tomorrow I'm happy to work with their
00:13:03
data and understand clumpiness and how
00:13:04
it can predict customer value so that's
00:13:06
one area I just want to apply it to new
00:13:07
data sets the second is I want to
00:13:10
understand the psychological processes
00:13:11
why are people behaving in a clumpy
00:13:13
fashion the third and final pieces I
00:13:15
want to relate marketing activity to
00:13:17
clumpiness now that's going to require
00:13:19
not just people's behaviors like what
00:13:21
did they do what websites did they visit
00:13:23
what did they purchase but information
00:13:25
about the marketing campaigns themselves
00:13:27
possibly even the copy of the marketing
00:13:29
campaign which channels they were sent
00:13:31
through and that's going to allow me to
00:13:33
come up with optimization ways for firms
00:13:36
to optimize their marketing campaigns to
00:13:39
activate clumpiness so those are the
00:13:40
three Avenues I'm going to be working on
00:13:46
[Music]
00:13:58
next
