Use of Data Mining and Machine Learning for Fraud Detection Ramazan Isik, CFE, CIA, CRISC, CRMA Chief Audit Executive DenizBank (Sberbank Group)
Welcome!
Use of Data Mining and Machine
Learning for Fraud Detection
CPE PIN Code: UKLK
Ramazan Isik, CFE, CIA, CRISC, CRMAChief Audit Executive
Denizbank (Sberbank Group)
Use of Data Mining and Machine
Learning for Fraud Detection
Ramazan Isik, CFE, CIA, CRISC, CRMAChief Audit Executive
DenizBank (Sberbank Group)
Служба внутреннего аудита
Ramazan Işık, CFE, CIA, CRISC, CRMA
Chief Audit Executive, DenizBankJanuary, 2017
Use of Data Mining /Machine
Learning for Fraud Detection
Sberbank at a Glance
EMPLOYEES COUNTRIES BRANCHES ATMs CUSTOMERS
321.000 22 17.500 90.000 138 Mio
The largest bank in Russia (by total assets,
loans and deposits.) & One of the top 30 banks in the world
137 mio Retail & 1.1 mio Corporate Clients
Presence in 22 Countries
16 territorial banks in Russia located across
11 time zones.
3
DenizBank at a Glance
The 5th largest private bank in Turkey according to consolidated assets at $41 bn.
One of the 2 private banks having branches in 81 Provinces in Turkey
Presence in 6 Countries (Turkey, Russia,
Austria, Germany, Cyprus, Bahrain)
1st at asset growth with 30% average in previous 10 years
EMPLOYEES CITIES BRANCHES ATMs CUSTOMERS
15.000 81 757 3.989 8.8 Mio
4
DenizBank’s Innovation Culture
5
fastPay
Mobile Loan
Banking
Twitter Loan
Most
Innovative
Application of
Technology
Grand Steve
Sales&
Customer
Service
Silver Stevie
Dijital Deniz
MPEAwards
Best
Payment
Solution
fastPay
Banking
Channel
Innovation
BAI 2014
MOST INNOVATIVE
BANK OF THE YEAR
DenizBank
EFMA ACCENTURE
INNOVATION
AWARDS
2015
DenizBankGLOBAL INNOVATOR
Financial Word
Innovation
Awards
BAI 2016
MOST INNOVATIVE BANK OF THE YEAR
DenizBank
Internal Audit Department
6
Head of Internal Audit
Branch Audit
Fraud Data Analytics
Head Office & Subsidiaries Audit
Examinations & Investigations
IT Audit
4 Vs of Big Data
7
Big Data
8
Data Mining
Discovering patterns in large data sets, involving methods at the
intersection of artificial intelligence, machine learning, statistics, and
database systems.
The overall goal of the data mining process is to extract information
from a data set and transform it into an understandable structure for
further use.
9
Machine Learning
Machine Learning is the study and
construction of algorithms and models that
can learn from and make predictions on
data. Generally operated by a model, for
example, inputs for future predictions.
Machine Learning
MODEL PREDICTION
10
Machine Learning in Daily Life
11
Face Recognition Virtual Personal Assistant
Self-Driving Cars
Road Traffic Monitoring and
Prediction
Personalization and Recommendation
Systems
Medical Diagnosis
Use of AI & Machine Learning in Financial Services
12
Use of AI & Machine Learning in Financial Services
Customer Lifetime
Value (CLV)
Campaign Analytics
Customer
Segmentation
Silent and
Proactive Churn
Social Media
Listening
Value at Risk
Calculation
Suspicious Activity
Reporting
Stress Testing
Pattern
Recognition and
Machine Learning
to Detect Fraud
Collateral Analysis
Collection
Delinquency
Simulations to
Predict Default
Risk
Central Limit
Management
Determining
Regulatory Capital
13
Is It Possible not Working with BD/ML?
• BD/ML is like the Internet in the
early 2000s.
Using BD/ML will provide a competitive
advantage in the near future (2-5
years).
Soon (after 5 years), not using BD/ML
will be a competitive disadvantage.
Available technologies will boom and
experts will not be sufficient at all.
Companies that have begun to implement and use a DB / ML technology will
have the advantage of understanding and applying new technologies.
14
Business Analytics Practices in DenizBank
Business AnalyticsPractices in DenizBank
15
Analytical Journey of DenizBank
Mar 2013
2005
Sept 2005
Jul 2008
KK CRMCRM
Fraud
Operational Efficiency
Marketing Models
Churn
PD, EAD, LGD
Campaign Optimization
Risk
Customer Lifetime Value(Revenue Prediction)
External Fraud
Jan 2008
ATM Cash Optimization ATM Location
Selection
Internal Fraud Detection
Overdue Debt Collection Optimization
Mar 2010 Sep 2012 Dec 2012
Jul 2014Mar 2013
Dec 2014
Apr 2011
Aplication
Scorecard
Jul 2015
Borrower’s Income
Prediction
16
Data Analytics Practices in the Internal Audit Department
Data Analytics Practicesin Internal Audit Department
17
Use of Data Analytics for Fraud Detection
Use of Data Analytics for Fraud Detection
18
Why We Need Machine Learning to Detect Fraud?
59 Rule-Based Scenarios
19.500 Scenario Results Per Day
Millions of Transactions-Customers
Different Employee Profiles
High Number of False Positive
Scenario Results
19
Objectives of the Inter-Fraud Tool Project
Easy interface integrated with
the core banking platform
Very easy to compose a
scenario without an IT
development and coding
Decreasing false positive results
for making more accurate sampling
Analyzing links and relations
among customers and
employees
Making quick queries for past
transactions
20
Infrastructure Inter-Fraud Tool
Datamarts (customer, employee,
transaction, relation)
The model was developed with “IBM
SPSS Modeler” and coded into the
core banking platform
Data mining technique: Decision tree
Anomaly identification in terms
of employee, customer, relation, and
transaction
21
Infrastructure Inter-Fraud Tool
Learning through newly
encountered types of fraud cases
Easy-to-use interface
Link analysis: Diagrams showing
both financial and non-financial
relations
Determining employee risk
22
3 Pillars of the Inter-Fraud Tool
Data Mining
&Learning
Algorithm
Detection
of Suspicious
Transactions
Investigation of
Suspicious Transactions
23
Data Mining & Learning Algorithm
Data Mining
&Learning
Algorithm
Detection
of Suspicious
Transactions
Investigation of
Suspicious Transactions
24
State-of-the-Art Machine Learning
Machine learning algorithm;
Learning: Similar patterns offraud cases by using real datafrom 2012, 2013, 2014, and2015.
Prediction: Using the model,rules, and the input to reachapproximate or definite resultswhich are similar to previousfraud cases.
25
Analytical Dimension of Inter-Fraud Tool
Transaction Datamart (69)
Trx Amount Currency Code Trx Channel Order User Cash Cover Balance/Limit Debit/Credit Applicant
Beneficiary ....
Customer Age Customer Type Channel Usage Transaction Types Transaction
Frequency Transaction
Amount RFM ....
Employee Age Mission Name Off-Hours
Transactions Number of
Trx/Customer Customer
Monitoring Logs Performance ...
Related customer/ other/employee
Relationship Type(78)
(family, financial, business,neighbor...)
Relation RFM Score
...
CustomerDatamart (380)
EmployeeDatamart (420)
Transaction
AnomalyScore
Transaction Risk Score
Relation Datamart
(45)
CustomerDatamart (380)
EmployeeDatamart (420)
CustomerAnomaly
Score
Employee Anomaly
Score
RelationAnomaly
Score
26
Customer Behavior Anomalies
Frequency of financial transactions
Channel usage habits
Type of transactions, time of transactions, amount of transactions
Employee Behavior Anomalies
Work hours
Query types and frequencies (frequency of customer signature, id,… queries)
Transaction types, amounts, frequencies
Relation Anomalies
What is the relation type? (Relative, sister, brother, neighbor, same school….) Link Analysis
Transaction Anomalies
Recency analysis
Frequency Analysis
Monetary Analysis
Anomaly Identification
27
Detection of Suspicious Transactions
Data Mining
&Learning
Algorithm
Detection
of Suspicious
Transactions
Investigation of
Suspicious Transactions
28
Easy to Compose a Scenario
29
Easy to Compose a Scenario
Pre-defined best practice scenarios
Composing a scenario from a data source or combining two
scenarios without IT development
Parametric structure for more effectivescenarios
Exception lists
30
Easy-to-Use Inter-face
31
Easy-to-Use Inter-face
Easy filtering scenario results
Advanced data mining and transaction risk score to focus on riskier cases
Easy access to investigation tools
32
Investigation of Suspicious Transactions
Data Mining
&Learning
Algorithm
Detection
of Suspicious
Transactions
Investigation of
Suspicious Transactions
33
Powerful Investigation Tools
34
Powerful Investigation Tools
Link analyses function Financial relations Non-financial relations
Powerful investigation modules Customer monitoring Employee monitoring Case history Employee-customer
Relations monitoring
Simulation screen for past query
35
Duration of
working
period with
other
colleagues
Working period in
the recent role
Past
disciplinary
penalties of
the employee
Credit history of
the employee
Vacation behaviors (not
taking annual
leave/working while on
annual leave)
Job
classification
of the
employee
Employee
risk score
Main Structure of Employee-Risk Matrix
Benefits of the Inter-Fraud Tool Compared to Previous Tool
03
01
02
04
05
90% More Efficient (1 hour to 6 min.)
89% Time Saving(9 hours to 1 hours)
89% Time Saving(9 hours to 1 hours)
38% Decreased(354 high risk decreased to 220 high risk)
70% Decreased
(10 min. to 3 min.)
37
Significant Cases Detected by Inter-Fraud Tool
38
Branch employee’s transactions on his father's account for
his benefit
High volume commercial activities in Branch employee’s
account
Branch employee’s usage of his relative’s credit card for
his/her own benefit
Money withdrawal transactions made by an unauthorized
person from 94-year-old branch customer’s account
Questions ?
Ramazan Işık, CFE, CIA, CRISC, CRMA
Chief Audit Executive, DenizBank
linkedin.com/in/ramazanisik
Welcome!
Use of Data Mining and Machine
Learning for Fraud Detection
CPE PIN Code: UKLK
Ramazan Isik, CFE, CIA, CRISC, CRMAChief Audit Executive
Denizbank (Sberbank Group)
Use of Data Mining and Machine
Learning for Fraud Detection
Ramazan Isik, CFE, CIA, CRISC, CRMAChief Audit Executive
DenizBank (Sberbank Group)