💳 Financial Fraud Detection using Machine Learning

📌 Project Overview

This project focuses on proactive fraud detection for financial transactions using machine learning. The objective is to build a predictive model that identifies fraudulent transactions and provides actionable business insights to improve fraud prevention infrastructure.

The dataset contains 6.3+ million transactions simulated over 30 days.

---

🎯 Business Objective

• Detect fraudulent transactions (isFraud)
• Improve detection beyond the existing rule-based system
• Reduce financial losses due to fraud
• Provide actionable recommendations for infrastructure upgrade

---

📊 Dataset Description

The dataset includes the following key features:

Feature	Description
step	Time unit (1 step = 1 hour)
type	Transaction type (CASH-IN, CASH-OUT, DEBIT, PAYMENT, TRANSFER)
amount	Transaction amount
oldbalanceOrg	Sender balance before transaction
newbalanceOrig	Sender balance after transaction
oldbalanceDest	Receiver balance before transaction
newbalanceDest	Receiver balance after transaction
isFraud	Target variable (1 = Fraud)
isFlaggedFraud	Rule-based fraud flag

⚠️ Note: Destination balances are not available for merchant accounts (IDs starting with "M").

---

🧹 Data Cleaning & Preprocessing

✔ Handling Structural Missing Values

Merchant accounts do not have destination balance information. We created a flag:

isMerchant = nameDest.startswith('M')

Destination balances for merchants were treated as missing values.

---

✔ Feature Engineering

Created new features:

• orig_balance_diff = oldbalanceOrg − newbalanceOrig
• dest_balance_diff = newbalanceDest − oldbalanceDest
• hour = step % 24
• isMerchant

These features help detect account-draining patterns.

---

✔ Handling Imbalanced Data

Fraud cases represent a very small percentage of transactions. We used:

class_weight='balanced'

in Random Forest to address imbalance.

---

🤖 Model Used

🔹 Random Forest Classifier

Why Random Forest?

• Handles nonlinear patterns
• Robust to outliers
• Handles imbalance
• Provides feature importance
• High performance on tabular data

---

📈 Model Evaluation

Since the dataset is highly imbalanced, we evaluated using:

• Confusion Matrix
• Precision
• Recall
• F1 Score
• ROC-AUC Score

🎯 Key Priority:

Maximize Recall for Fraud Class (Missing fraud = financial loss)

---

🔍 Key Predictors of Fraud

Top important features:

1. orig_balance_diff
2. amount
3. type_TRANSFER
4. type_CASH_OUT
5. oldbalanceOrg
6. hour

Fraudsters typically:

• Transfer funds
• Drain accounts fully
• Perform CASH-OUT transactions

These patterns align with business logic.

---

🏢 Business Insights

🔹 Existing Rule-Based System Limitation

Current system flags only transfers > 200,000. It misses many fraudulent behaviors.

🔹 Recommended Solution

Replace static rules with:

• ML-based fraud probability scoring
• Real-time risk assessment
• Multi-level transaction verification

---

🚀 Infrastructure Recommendations

1. Deploy model as real-time scoring API
2. Introduce risk-based authentication
3. Implement monitoring dashboard
4. Retrain model periodically
5. Conduct A/B testing before full deployment

---

📊 Measuring Success

After implementation, success can be measured using:

• Increase in fraud detection rate
• Reduction in financial loss
• Controlled false positive rate
• Improved operational efficiency

---

🛠 Tech Stack

• Python
• Pandas
• NumPy
• Scikit-learn
• Matplotlib
• Seaborn

---

📌 Conclusion

• Fraud primarily occurs in TRANSFER and CASH-OUT transactions.
• Fraudsters often empty accounts completely.
• ML-based detection significantly outperforms rule-based detection.
• Proactive fraud scoring improves risk management and financial protection.

---