Predictive Analytics + Constrained Optimization =
Efficient Marketing Paul%Maiste,%President,%Lityx%
www.lityx.com%|%[email protected]%March%18,%2015%
INFORMS%New%York%
2%
Abstract
PredicGve%modeling%has%become%more%mainstream%in%the%last%5%years%as%companies%add%more%advanced%capabiliGes%to%their%analyGc%tool%sets.%%There%has%also%been%more%talk%about%incorporaGng%concepts%related%to%opGmizaGon%(constrained%or%unconstrained)%into%business%decision%making.%%There%is%no%argument%that%each%can%be%quite%powerful%in%their%own%right. In%this%presentaGon,%the%speaker%will%take%it%to%the%next%level:%combining%predicGve%models,%predicGve%opGmizaGon,%and%constrained%opGmizaGon%techniques%to%provide%tools%that%allow%marketers%to%make%opGmal%decisions%and%dramaGcally%improve%efficiency.
! Background%! PredicGve%OpGmizaGon%Models%! Constrained%OpGmizaGon%! Case%Studies%
• Prospect%Offer%OpGmizaGon%• Print%Media%TargeGng%OpGmizaGon%• Retail%Direct%Mail%TargeGng%OpGmizaGon%
! Wrap%Up%
3%
Agenda
! Lityx%is%an%analyGc%soluGons%and%services%firm%with%a%proprietary%cloud\based%predicGve%modeling%and%opGmizaGon%pla]orm%LityxIQ.%
Lityx Background
! We%apply%deep%experGse%in%complex%analyGc%soluGons%with%a%focus%on%applicaGons%in%markeGng%analyGcs%and%customer%relaGonship%management.%
4%
5%
How%Can%%We%Make%It%%Happen?%
Business Value
Co
mp
lexi
ty
High Low
Low
High
What Happened?
Why Did It Happen?
What Will Happen?
How Can We Make It Happen?
Descriptive Analytics
Prescriptive Analytics
Predictive Analytics
Diagnostic Analytics
Gartner%framed%analyGc%progression%as%the%move%forward%from%InformaGon%to%OpGmizaGon%
Analytic Progression
6%
How%Can%%We%Make%It%%Happen?%
Business Value
Co
mp
lexi
ty
High Low
Low
High
What Happened?
Why Did It Happen?
What Will Happen?
How Can We Make It Happen?
Descriptive Analytics
Prescriptive Analytics
Predictive Analytics
Diagnostic Analytics
Gartner%framed%analyGc%progression%as%the%move%forward%from%InformaGon%to%OpGmizaGon%
Analytic Progression
Op#mal'Marke#ng'Strategy%
Scoring%and%ImplementaGon%
Constrained%OpGmizaGon%
PredicGve%(OpGmizaGon)%Modeling%
7%
Generic Concept
! Separate%techniques%work%together%
! We%can’t%forget%the%execuGonal%piece%of%it%(scoring/%implementaGon)%
! The%global%problem%(opGmal%markeGng)%is%solved%with%the%combinaGon%
! Where%does%OpGmizaGon%appear%in%this%approach?%• PredicGve%models%are%naturally%an%opGmizaGon%process%(nothing%new%here).%
• “PredicGve%OpGmizaGon%Models”%incorporate%markeGng%tacGcs%into%the%predicGve%model%inputs.%o Offer%type%o Offer%value%o Discounts%o Messaging,%sequencing%o Channel%
• Constrained%OpGmizaGon%incorporates%hard%constraints%into%the%overall%soluGon%o Business%or%markeGng%constraints%o Contractual%constraints%o Channel%constraints%
8%
Optimization Components
%%%A%couple%of%predicGve%opGmizaGon%modeling%examples%
9%
10%
Offer Optimization
$ OID HLF 0MP P
EEDO RD1HPBMRL
EEDO
EEDO )EEDO (
2TNDB DCOMEH AH H U
( ) +
N H MEEDO HP EEDO ( CHPBMRL
OMEHAD
OMNMPHHML
LNOMEHAD
OMNMPHHML
2TNDB DC 0RP M DOOMEH AH H U -
3 0=5 3
=O LP B HML 4HP MOU 6MU U 1D MFO N HBP
NNDLCDC 1 OID HLF 2 D DL P
OID HLF 6DSDOP$ EEDO =UND$ EEDO RD
Offer'Elas#city'Curves'
Modeled%Behaviors%
MCD DC /D SHMOP$ DPNMLPD D$ NDLC% RD
MarkeGng%Levers%
11%
Actual Offer Elasticity Curves
12%
Cadence Optimization
.905 .629 .528 .489
.796 .644 .593 .480
.606 .564 .469 .402
.361 .543 .515 .390
Note:%Gray%=%low%sample%size%
Not%targeted%in%last%two%months%
Targeted%two%
months%ago%
Targeted%last%month%
Targeted%both%of%last%
two%months%
Last'Two'Months'Ac#vity'
Not%targeted%3%or%4%months%ago%
Three'an
d'four'm
onths'a
go'
Targeted%4%months%ago%
Targeted%3%months%ago%
Targeted%both%3%and%4%months%ago%
Lowest%values%here%
Best%values%here%
We'can'incorporate'cadence'into'our'predic#ve'op#miza#on'model.''
'Recency'and'frequency'can'be'categorized'in'a'number'of'different'ways'
Cells'show'average'observed'response'rates'
More%“Freshness”%
More%“Freshness”%
%%%
What%about%constraints?%
13%
! Constraints%make%the%overall%problem%harder%to%solve,%and%(even%worse)%give%a%less%than%globally%opGmal%soluGon.%
! But%we%need%to%account%for%them.%! Number%and%complexity%of%constraints%determine%the%soluGon%approach.%
14%
Business/Marketing Constraints
Dimensions' Constraints' Solu#on' Example'
0% 1%(total%budget)% Rank%order%all% Overall%markeGng%budget%
1%(channel)% 1%(total%budget)% Rank%order%by%dim% MarkeGng%budget%by%channel%
1%(channel)% 2%(channel%budget)% Rank%order%by%dim% Channel%max%spend%and%min%spend%
2+% 2+% Harder!%%Need%tools%
Channel%max%spend%with%product%min%performance%+%more%
! In%our%offer%opGmizaGon%example,%if%there%was%a%policy%to%never%offer%more%than%$100,%we%would%not%realize%global%opGmality.%
15%
Sub-optimality
%%%
Prospect%Offer%OpGmizaGon%
16%
Case Study 1
! We%want%to%maximize%%• Expected%profitability%OR%• Response%rate%OR%• Expected%revenue%
! We%have%predicGve%opGmizaGon%models%that%predict%• Spend%by%prospect%• Response%rate%by%prospect%
! Constraints%• Limits%on%total%amount%of%offers%• Limits%on%freebies%(e.g.,%hotel%rooms%capacity%constraints)%• Experimental%tesGng%and%cross\decile%sampling%• Segment%level%markeGng%plan%
17%
Offer Optimization Problem
18%
Data View
19%
Outputs Number'to'Contact'by'Offer'
Budget'by'Offer'
! For%this%industry,%prospects%are%placed%into%an%offer%matrix%for%markeGng%execuGon.%
20%
Implementation
%%High'Offer'Value' Low'Offer'Value'
Offer%1% Offer%2% Offer%3% Offer%1% Offer%2% Offer%3%Predicted'Value''<$75%%
3506% 3962% 3331% 763% 2095% 1428%
Predicted'Value''
$75R$299'%%
5556% 4572% 6074% 358% 612% 641%
Predicted'Value''
$300R$749'%%%
1642% 384% 394% 38% 12% 17%
Predicted'Value''$750+%
475% 12% 14% 10% 0% 0%
%%%
%Print%Media%TargeGng%OpGmizaGon%(Cadence)%
21%
Case Study 2
! MulGple%Channels%• Newspaper%inserts%• Shared%mail%(mulGple%programs)%
! Many%choices%• 30,000+%zip%codes%• 60,000+%targeGng%zones%
! Many%constraints%! PredicGve%opGmizaGon%models%predict%zip\level%response%rate%given%prior%history,%cadence,%demographics%
! Decision%• On%a%monthly%basis,%how%do%we%best%allocate%our%fixed%budget?%%(i.e.,%what%zips/zones%should%be%targeted)%
22%
Print Media Targeting Problem
23%
It’s a complex process
Aggregation of program/zip level actual historical response rates
Zip/zone targeting history (cadence)
Historical response rates
Media Data Business Rules and Constraints
Marketing constraints
Mathematical Optimization
Model
• Minimum insert restrictions
• Fixed setup costs • Target/Do not target
rules
• Cross-channel budget min/max percent
• Max papers • Restricted papers
Ranked Zone List - Ordered by Decreasing
Expected CPO
Final Optimization Result
• Zip/zone maps • CPM • Circulation
Predictive Model of Newspaper Zone Response Rates
Predictive Model of Shared Mail Zone Response Rates Dynamic Data
(Processed Monthly)
Prediction of Response Rates at Zone Level
Static Models (updated approx yearly)
Census and Geo-demographic data at
Zip level
23%
Scoring
Somewhat Static Data (yearly)
! The%constraint%list%covers%a%wide%variety%of%metrics,%dimensions,%and%business%rules.%
! Without%these%accounted%for,%the%problem%could%not%be%solved%to%the%saGsfacGon%of%the%business.%
24%
Constraints
25%
Outputs
26%
Output
%%%
%Retail%Direct%Mail%TargeGng%OpGmizaGon%
27%
Case Study 3
! Maximize%overall%response%rate%! Each%prospect%has%a%predicGve%model%score%based%on:%
• Demographics%(age,%income,%educaGon)%• LocaGon%(distance,%driveGme%to%closest%store)%• Appended%behaviors%(credit%card%usage,%internet%usage)%
! Constraints%• Meet%store%minimum%quanGty%needs%• Cross\decile%tesGng%and%metrics%• Prior%Gmes%targeted%
o Cadence%is%(currently)%controlled%in%this%case%study%through%constraints%–%not%incorporated%yet%into%a%modeling%paradigm.%
28%
Retail Direct Mail Problem
29%
Outputs Mail'Qty'by'Decile'
Mail'Qty'by'Prior'Times'Targeted'
Mail'Qty'by'Store'Loca#on'
! ImplemenGng%modeling%and%constraints%has%helped%generate%incremental%sales%consistently%since%the%program%begain%mid\July%2014%
30%
Results
! Incorporate%predicGve%opGmizaGon%modeling%• Cadence%/%prior%contacts%• Direct%mail%package/message%
o What%is%opGmal%way%to%communicate%with%each%prospect?%
31%
Future Objectives
! Purng%it%all%together%raises%overall%complexity%! But%the%reward%can%be%significant%
• Improve%key%business%metrics%• Ease%of%markeGng%execuGon%
! Tools%can%help%
32%
Summary