Fraud Detection and risk analysis

This project presents an exploratory data analysis and anomaly detection approach for identifying suspicious invoice activity within a financial services context. The goal is to uncover patterns of potential fraud or credit abuse in customer transaction data following a system integrity issue.

📌 Project Overview

A system error led to a period where standard risk checks—such as internal scoring, external credit ratings, and fraud detection mechanisms—were bypassed. As a result, a large number of transactions were approved without validation. This repository contains a structured investigation into invoice-level transaction data to identify unusual patterns and behaviors that may suggest fraudulent activity.

Since no confirmed fraud cases exist, this task focuses on unsupervised anomaly detection, domain-driven insights, and exploratory methods to assess the extent of potential risk exposure.

🔍 Objectives

• Analyze invoice transactions to identify irregularities or high-risk behaviors.
• Define and apply custom anomaly detection criteria based on domain knowledge.
• Detect potential fraud signals, such as:
  • Unusual purchase volumes
  • Irregular address or SSN usage
  • Repeated or coordinated transactions across multiple identities
  • Signs of identity misuse or synthetic identities
• model development and deployment

🧰 Tools & Technologies

• Python
• Pandas, NumPy for data processing
• Matplotlib, Seaborn for data visualization
• Scikit-learn, Isolation Forest, DBSCAN for unsupervised anomaly detection
• Jupyter Notebook for reproducible and interactive analysis

📁 Project Structure

. ├── data/ │ └── invoices.csv # Sample anonymized invoice data ├── notebooks/ │ └── fraud_detection_analysis.ipynb # Main analysis notebook ├── src/ │ └── preprocessing.py # Data cleaning and feature engineering │ └── anomaly_detection.py # Custom anomaly detection functions ├── requirements.txt # Python dependencies └── README.md # Project documentation

📊 Data Description

Each row represents a single invoice sent to a customer. Below are selected features included in the dataset:

Column Name	Description
invoice_no	Unique invoice ID
email	Anonymized customer email
social_security_number	Pseudonymized social security number / ID
customer_no	Internal customer ID
principal_amount	Original invoice amount
is_debt_collection_stopped	Binary flag indicating whether collections were stopped
last_event	Last status or activity on the invoice
last_event_date	Date of the last event
period_start, period_end	Invoice period start and end dates
born_year	Extracted year of birth (ambiguous 2-digit year)
Address fields	Includes city, postal code, street, and house number
firstname, lastname	Pseudonymized customer names

🧠 Methodology

• Exploratory Analysis: Summary statistics, customer segmentation, time-based trends
• Rule-Based Filters: Manually crafted signals such as:
  • High purchase amounts in a short period
  • Duplicate SSNs or addresses across accounts
  • Unusual frequency of orders from the same city or postal code
• Unsupervised Learning: Applied Isolation Forest and DBSCAN to detect outliers in multidimensional feature space
• Visualization: Highlighted anomalous points and distribution shifts during the risk system outage

🚀 Key Insights

• Detection of customer groups showing bulk purchase behavior shortly after the validation checks failed.
• Identification of duplicate or suspiciously similar SSNs and email patterns.
• Clustered activities around certain geographic locations potentially linked to organized behavior.

✅ How to Run

1. Clone the repository

   git clone https://github.com/RozaAbolghasemi/Fraud-detection/blob/main/Fraud_Detection.ipynb
   
   

2. Install the requirements pip install -r requirements.txt

3. Run the analysis Open the Jupyter notebook in the notebooks/ directory and follow the step-by-step analysis.

🧾 Disclaimer This project is a demonstration of anomaly detection and exploratory techniques on synthetic, anonymized data. It is intended for illustrative and educational purposes in a financial risk analysis context. No real individuals can be identified, and no actual customer data is used.