An Introduction to Predictive Analytics Final

8/6/2019 An Introduction to Predictive Analytics Final

1/31

This presentation cannot be reproduced or shared without FICOs express consent.

2009 FICO Corporation.1

An Introduction to Predictive AnalyticsFor Business Rule Developers

Mac BelniakPrincipal Sales Consultant, Model Builder

Monday, August 24, 2009


2/31


2009 FICO Corporation..2

Agenda

Decisions & Uncertainty

Predictive Analytics

Analytics & Rules

Model Builder Demonstration

Decision Management

Q&A


3/31


2009 FICO Corporation..3

Agenda



Analytics & Rules


Decision Management

Q&A


4/31


Uncertainty clouds decision making

Approve/DeclineLoan

DelinquencyHistory

CurrentLimit

CurrentBalance

Profit/

Loss

Default?

DecisionTimeline

Odds ofDefault


5/31


Predicting if an event will occur parts the clouds

Binary outcome models

What are the odds an event will occur in a futuretime period?

E.g. We cant know if a borrower will not miss apayment next year.

But we can predict the odds that they will not miss apayment next year!

10:1

The prediction can help inform future decisions.

Examples What are the odds that a prospect will respond to a

campaign?

What is the likelihood that this transaction is fraudulent? What are the odds that a customer will attrite if I lower

their credit limit?


6/31


Predicting amounts brings even more clarity

Continuous outcome models

What will an amount be in a future time period?

E.g. We cant know how much we will lose if a

borrower defaults. But we can predict the amount that will be lost!

$106,000

The forecast can help inform future decisions.

Examples How many dollars will the prospect spend if they respondto the campaign?

How much will a cardholders utilization change if Iincrease their credit limit?


7/31


Agenda



Analytics & Rules


Decision Management

Q&A

2009 FICO Corporation..


8/31


Predictive models have a natural lifecycle

Access Prepare Explore

Data

Discover Approve Deploy

Model

Rapidscorecard

development Streamlined

model reviewFast migrationinto decisionapplication


9/31


Historical data is a key input for predictive modeling

P

erformance

De

velopment

3Months

24Months

History Behavior Measurement

Tenure?

Months since account opened.

Delinquency?

30 days delinquent in last year.

Search for Credit?

# Inquiries last 6 months.Utilization?

Balance / Credit Limit.

Target:Good or bad?

Was the customer 30, 60, or 90 daysdelinquent or charged off during the

performance period?

2007

2008

2009


10/31


Predictive models learn from historical patterns

TargetVariable

Customer

ID

Monthssinceaccountopened

30daysdelinquentinlastyear

#inquirie

slast6months

Balance/

CreditLimit

GoodorB

ad

000000001 9 0 1 30% 0

000000002 2 0 0 64% 0

000000003 11 0 2 100% 0000000004 7 1 1 84% 1

000000005 8 0 1 30% 0

000000006 17 0 0 37% 0

000000007 3 0 1 83% 0

000000008 11 0 1 41% 0

000000009 23 0 2 73% 0

InputVariablesHistorical Data

ScorecardAlgorithm

ModelDevelopment

Predictive Model


11/31


For each variable,understand if and how it predicts historical outcomes

Odd

sofGood

10:1

5:1

1:1

15:1

20:1

Total Portfolio Odds

# inquires in the last 6 months

None 1 2 3 4


12/31


Find the most informative set of variables

Odds

ofGood

# inquires in the last 6 months

Odds

ofGood

Years at current address


13/31


Variable Weight

# inquires last 6 months

None 53

1 34

2 10 24

# times 30 days delinquent last year

None 53

1 24

2 63

3 Max 33

Months since account opened

0 36 45

37 Max 24

Variable Weight

# inquires last 6 months

None 64

1 55

2 10 48


None 74

1 60

3 53

4 Max 45


0 36 51

37 Max 75

Train the predictive model

TargetVariable

CustomerID

Monthssinceacco

untopened

30daysdelinquen

tinlastyear

#inquirieslast6m

onths

Balance/CreditLimit

GoodorBad

000000001 9 0 1 30% 0000000002 2 0 0 64% 0

000000003 11 0 2 100% 0

000000004 7 1 1 84% 1

000000005 8 0 1 30% 0

000000006 17 0 0 37% 0

000000007 3 0 1 83% 0000000008 11 0 1 41% 0

000000009 23 0 2 73% 0

InputVariablesScorecardAlgorithm


14/31


Generate scores and rank order the customers

CustomerID

#inquirieslast6m

onths

30daysdelinquen

tinlastyear

Monthssinceacco

untopened

000112354 1 0 9

InputVariables Variable Weight# inquires last 6 months

None 64

1 55

2 10 48


None 74

1 60

2 533 Max 45


0 36 51

37 Max 75

55

74

+

51

+

= 180

150 175 200 225

GoodBad

Score


15/31


Evaluate the quality of the ranking

Customers Sorted by Scores

10%

20%

%BadCusto

mersBelowth

eScore

30%

40%

50%

60%

70%

80%

90%

100%

0%

WorstCustomers

BestCustomers

10%

20%

30%40%

50%

60%

70%

80%

90%

100%

0%

The 20% of customers withthe lowest score cover over

70% of the bads!


16/31


Agenda



Analytics & Rules


Decision Management

Q&A



17/31


Measuring benefitof analytics with champion/challenger

Champion: The rules and models that currently automate decisions. Challenger: Alternative rules and models that are likely more effective. Champion/Challenger: Running the challenger on a small fraction of

the population to verify that it is better.

IF account_number ends in 0-8 Champion

IF behavior_score > 650

&& utilization > 70%THEN up limit by 10%

Champion$ Profit / Account

Challenger$ Profit / Account

ELSE

Challenger IF behavior_score_1 > 750

&& utilization > 80%

THEN up limit by 5%


18/31


Agenda



Analytics & Rules Model Builder Demonstration

Decision Management

Q&A



19/31


Model Builder supports the entire analytic lifecycle

Access Prepare Explore Model Deploy

SAS Datasets Special Values

Teradata Oracle SQL Server Fixed-Width &

Delimited Text;ASCII & EBCDIC

MB Native Files

Replace Missing Sample

Stratified Random Partition Train/Test Sort Join/Merge Append Filter/Where

Variable Creation Arrays, RegularExpressions, etc.

View Table Statistics

Summary Frequency Correlation By-Variable

DatasetComparison

Linear Regression Logistic

Regression Neural Networks Reason Codes K-Means & Bang

Clustering PCA Evaluation

Reports ROC, GINI, KS,etc.

Services-Oriented-Architecture

Batch Transactional PMML

Blaze Advisor Teradata

Scorecard Tracking Char. Analysis Population Stability


20/31


Historically, deployment has been expensive

Traditional technique

Document model

Ask IT to recode

Lengthy testing

How many $ lost per day?

Time To Deploy Predictive ModelsSelf-Reported Measures, Global Financial Companies

PredictiveModelSpecs

CompiledExecutableIT Software

Development

SQL inDatabase


21/31


Quickly deploy via Blaze Advisor

Imported models are available to ruledevelopers and authorized businessusers can see and modify them

PMML integration

ImportDM

RepositoryRule Service

.NET

Rule Service

Java

Code Gen

COBOL

Model Builder


22/31


Model Buildersupports scorecard tracking to trigger model refresh

Easily evaluatepopulation stability

Track shifts in score distribution

Analyzecharacteristic-level changes


23/31

2009 FICO Corporation.23 2009 FICO Corporation.23

Demonstration

Behavior Risk Scorecard Import Data from SAS

Check the Data Find Predictive Patterns Select Best Predictors Evaluate Scorecard Quality

Export to Blaze Advisor


24/31


Agenda



Analytics & Rules Model Builder Demonstration

Decision Management

Q&A



25/31


Fast, Fast, Fast

Stronger predictive models in less time

Accelerated analytic review & approval

Rapid deployment into production systems

Access Prepare Explore

Data

Discover Approve Deploy

Model

Minimizing Time to Better Decisions with Model Builder


26/31


Customer Lifecycle Solutions

ACQUIRE MANAGE PROTECT

Marketing Origination CustomerManagement

Collections &Recovery

Fraud

Decision

ManagementApplications

Marketing

Solutions

Capstone

Decision ManagerLiquidCredit

Service

TRIAD

adaptive controlsystem

Debt Manager

RecoveryManagement

System

Falcon

FraudManager

LifecycleAnalytics andServices

PreScore

ServiceOrigination andSmall Business

Scores

TransactionRisk Scores

Debt PlacementServices

Collection Scores

CardAlertService

Across the Lifecycle

Analytics Industry Standard and Custom

Credit Bureau Scores: FICOScore Global FICOScore FICOExpansion ScoreInsurance Scores FICO Credit Capacity Index MyFICO

Custom Analytics: Predictive Analytics Optimization Portfolio Analytics

DecisionManagementTools

For Application Developers and In-House Analytics TeamsBusiness Rules Management: FICO Blaze Advisor

Model Development: Model Builder

Optimization: Decision Optimizer Xpress-MP Suite

ProfessionalServices

Analytic Consulting

Business and Technology Consulting

FICO address all stages of the customer lifecycle


27/31


AnalyticCapabilities

Shared

Workflows

Data

Create value by connecting decisions

Decision ManagementApplication

Smarter DecisionDecisionLogic

Other Applicationsand

Lifecycle Areas


28/31


FICO is the leader in Decision Management

FICO is where the mathematical approach to problem-solvingthat is inherent in todays scores and analytics all began.

William Blair & Company

PIONEER Introduced many analytic breakthroughs, including credit scoring

Integrated predictive analytics with rules management

LEADER Products that serve whole industries at high scale 10 billion FICO scores sold a year

Reviewing 20 billion card transactions a year for fraud

PARTNER Expertise on best practices in Decision Management, includingoperational, legal and regulatory issues

Serving many of the worlds top businesses:

9 of the top 10 Fortune 500 companies

2/3 of the top 100 banks in world

2/3 of top US P&C insurers

1/2 of the top US retailers


29/31


30/31


Agenda



Model Builder Demonstration Analytics & Rules

Decision Management

Q&A



31/31

This presentation cannot be reproduced or shared without FICOs express consent.


THANK YOU

Mac [email protected]

An Introduction to Predictive Analytics Final

Documents