Page 59 - CMA Journal (July-August 2025)

Focus Section

a) Enhanced Precision: By examining broader ranges Unsupervised learning detects irregularities in
of characteristics and behavioral data, ML models uncategorized data, making it effective for identifying
better differentiate between genuinely suspicious new fraud types.
transactions and unusual but legitimate activities 3) Reinforcement Learning
(such as vacation purchases in new locations).
Reinforcement learning can be used when there isn’t a lot
b) Risk Assessment: Rather than simple “fraud/
of training data available, the ideal end state cannot be
legitimate” classifications, ML models typically clearly defined, or the only way to learn about the
provide risk ratings. This allows financial institutions
environment is to interact with it.
to make more sophisticated decisions and focus
resources on investigating the highest-risk A. Primary Machine Learning Methods
transactions. Method Purpose
(5) Diverse Financial Service Applications Decision Trees & Transaction type
Random Forests classification
Machine learning is applied across a wide range of
financial fraud scenarios: Logistic Regression Binary fraud/legitimate
classification
• Credit Card Fraud: ML examines transaction details
such as amount, location, merchant, and timing to Neural Networks Complex fraud
promptly identify and flag suspicious activities, pattern recognition
preventing unauthorized usage or withdrawals.
K-Means Clustering Grouping similar behaviors to
• Financial Statements: ML can code accounting identify outliers
entries, enhancing the accuracy of rules-based
Autoencoders Input reconstruction to
approaches and enabling greater automation, while detect unusual variations
predictive models can also forecast revenues.
B. Data-Related Obstacles in Fraud Detection
• Account Security: ML monitors user access patterns
to detect unusual activities, such as repeated failed Several data challenges complicate ML fraud detection:
login attempts or access from unfamiliar devices or • Class imbalance: Fraudulent cases occur less
locations. frequently than legitimate ones.
• Documentation and Internal Fraud: ML analyzes • Data privacy: Sharing financial information for
invoices and related documents for inconsistencies, model development creates legal issues.
including duplicate billing or questionable vendor • Feature development: Extracting valuable features
information.
from raw data is essential.
• Insurance Fraud: ML identifies fraudulent claims, • Classification difficulties: Accurate and proper data
billing irregularities, and excessive service utilization.
labeling can be lengthy and prone to errors/
Standard Machine Learning Methods in mistakes.
Fraud Detection
C. Real-World Examples and Industry Uses
Various algorithms support fraud detection and are Multiple organizations have successfully implemented
typically divided into supervised, unsupervised, and
ML for fraud detection:
reinforcement learning approaches.
• PayPal: Utilizes deep learning and combined
1) Supervised Learning
methods to identify suspicious transactions.
Supervised learning uses categorized data (transactions • Mastercard: Applies decision intelligence for
already classified as fraudulent or legitimate) to train
monitoring billions of transactions.
models for classifying new transactions. Common
algorithms include: • Banks and FinTechs: Employ ML for immediate AML
screening, transaction monitoring, and customer
• Logistic Regression: A straightforward yet effective verification.
classification method for binary outcomes
(fraud/legitimate). About the Author: The writer is a Fellow member of ICMA and an
ACCA member with over 22 years of experience in public and private
• Decision Trees and Random Forests: These create sectors, including a tenure as acting CEO in a public sector company.
rule sets for data classification, often providing clear He is currently CFO at Craft Industrial Co., Saudi Arabia. His expertise
interpretations. spans strategic financial planning, regulatory compliance, business
development, and international financial reporting. He also mentors
• Support Vector Machines (SVM): This method emerging professionals and serves as an ACCA IPSAS Trainer and
identifies optimal boundaries separating fraudulent Corporate Trainer at the Pakistan Audit & Accounts Academy.
from legitimate transactions.
2) Unsupervised Learning

ICMA’s Chartered Management Accountant, Jul-Aug 2025 57