UBS Data Mining Workshop Introduction

Download Report

Transcript UBS Data Mining Workshop Introduction

Applications of Data Mining in
Banking
Maria Luisa Barja ([email protected])
Jesús Cerquides ([email protected])
Ubilab IT Laboratory
UBS AG
Zurich, Switzerland
1
Outline






5/11/98
Data Mining in Banking
Application Areas
Pitfalls in the Development of Data Mining Projects
An Alternative: A Data Mining Framework
Open Projects
Summary
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
2
Data Mining in Banking





5/11/98
Banks have many and huge databases
Valuable business information can be extracted
from these data stores
Unfeasible to support analysis and decision
making using traditional query languages
Human analysis breaks down with volume and
dimensionality
Traditional statistical methods do not scale and
require significant analysis expertise
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
3
Application Areas
Four main areas
 Marketing
 Credit Risk
 Operational Risk
 Data Cleansing
5/11/98
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
4
Applications: Marketing
Objective:
Improve marketing techniques and target customers
Traditional applications:
 Customer segmentation
Identify most likely respondents based on previous campaigns

Cross selling
Develop profile of profitable customers for a product

Predictive life cycle management:
Develop profile of profitable customers X years ago

Attrition analysis:
Alert in case of deviation from normal behaviour
5/11/98
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
5
Applications: Credit Risk
Objective:
Reduce risk in credit portfolio
Traditional applications:
 Default prediction
Reduce loan loses by predicting bad loans

High risk detection
Tune loan parameters ( e. g. interest rates, fees) in order to maximize profits

Profile of highly profitable loans
Understand characteristics of most profitable mortgage loans
5/11/98
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
6
Applications: Data Cleansing
Objective:
Detect outliers, duplicates, missing values,...
Traditional applications:
 Data quality control
Detect data values which do not follow the pattern

Missing values prediction
Predict values of fields based on previous values
5/11/98
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
8
Pitfalls in the Development of
Data Mining Projects



Data Mining is a process, not a package!
Expensive, difficult to justify in first instance
Having substantial parts in common, most data
mining projects provide custom solutions that:
–
–
–

5/11/98
Are more expensive
Take more time to develop
Have a higher risk of not being finished
Ideally, use more than one technique to get a full
view of the data
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
9
Proposed Alternative
Identify the common functionality used for
the development of data mining solutions
 Implement and pack this functionality in a
way that it can be:

–
–
–
5/11/98
Reused in many projects.
Customized to meet the needs of each project.
Extended, so it grows with its usage.
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
10
Object Oriented Frameworks

A framework is a reusable, “semi-complete”
application that can be specialized to
produce custom applications.
Framework design expertise
Programming language expertise
OO expertise
Domain expertise
5/11/98
Framework
Framework usage expertise
Programming language expertise
OO expertise
Ensemble
Coding expertise
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
11
Data Mining Framework:
Benefits





5/11/98
Reduces design and development efforts for building
concrete applications.
Lowers threshold for “proof of concept” data mining
applications to be developed.
Allows comparison of results across various methods.
Facilitates selection of best method(s) for particular
domains and business objectives.
Eases extensibility to new types of methods and
algorithms.
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
12
Data Mining Framework:
General Architecture
Project Management
Technique
implementation
Component structure
5/11/98
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
13
Data Mining Framework:
Component Structure
Project Management
Technique
implementation
Data
5/11/98
Process
Visualization
Component
Metadata
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
14
Data Mining Framework: Method
Implementation
Project Management
Data Understanding
Data Preparation
Modeling
Learning Data
Database Access
Component structure
5/11/98
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
15
Data Mining Framework:
Modeling
Prediction
Description
Classification
Clustering
Regression
Modeling roles
Learning Data
5/11/98
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
16
Data Mining Framework:
Open Projects

Design and development of:
–
–
–
–
–
5/11/98
A graphical user interface.
The prediction/description component (based
on bayesian networks).
The clustering component.
The project management component.
The preprocessing component.
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
17
Summary
Data Mining has emerged as an strategic
technology for a large bank
 Several business areas where it can be
applied
 Application development difficulties
 Proposed a solution based on OO
framework technology

5/11/98
©1998 M.L.Barja, J.Cerquides Ubilab IT Laboratory, UBS AG
18