🤖 Financial Crime Detection ML Model

Production-grade financial fraud detection pipeline — XGBoost + Random Forest trained on a PaySim-inspired synthetic dataset with SMOTE oversampling for class imbalance. Achieves ROC-AUC 1.00, PR-AUC 1.00, and zero false negatives on a 1.12% fraud-rate dataset.

---

📌 Project Overview

This project builds an end-to-end ML pipeline for detecting financial crime in payment transactions — directly mirroring the models used by fraud analytics teams at banks, payment platforms, and fintech companies. The dataset is modelled on the PaySim simulation framework (IEEE-CIS-style), with realistic class imbalance, transaction balance mechanics, and fraud typologies.

Relevance: Directly applicable to Data Scientist II, Data Analyst (ML), and Financial Crime Analytics roles at UST Global, Wissen Technology, Oracle, and BFSI analytics teams.

---

🏗️ Project Structure

fincrime-ml-model/
│
├── generate_data.py          # PaySim-inspired synthetic data generator (10,000 txns)
├── train_model.py            # XGBoost + Random Forest + SMOTE training pipeline
├── fincrime_dashboard.html   # Interactive ML dashboard (Chart.js)
│
├── paysim_data.csv           # Raw synthetic payment transactions
├── paysim_scored.csv         # Enriched with ML fraud scores and tiers
├── model_results.json        # Full metrics, feature importances, ROC data
│
└── README.md

---

📊 Dataset — PaySim-Inspired Synthetic Data

Property	Value
Total Transactions	10,000
Simulation Window	30 days (720 hours)
Fraud Cases	112 (1.12% — realistic class imbalance)
Fraud Transaction Types	CASH_OUT, TRANSFER only (PaySim pattern)
Total Fraud Amount	₹2.87 Crore
Avg Fraud Amount	₹2.56 Lakh
Avg Legit Amount	₹15,448

---

⚖️ Class Imbalance — Why SMOTE Matters

At 1.12% fraud rate, a naïve model predicts all-legit and achieves 98.88% accuracy while catching zero fraud. This is the classic imbalanced-class trap in financial crime ML.

Solution: SMOTE (Synthetic Minority Oversampling Technique) from imblearn — generates synthetic fraud samples via k-nearest-neighbours interpolation:

Stage	Fraud Samples	Legit Samples	Ratio
Before SMOTE	90 (train)	7,910	1 : 88
After SMOTE	7,910	7,910	1 : 1

This forces the model to learn genuine fraud decision boundaries rather than exploiting the majority-class shortcut.

---

🔧 Feature Engineering

Feature	Description	Importance
dest_balance_zeroed	Destination balance drained to zero post-txn	45.25%
balance_diff_dest	Destination account balance change	30.40%
error_balance_dest	Accounting error: old_dest + amount ≠ new_dest	24.30%
log_amount	Log-transformed transaction amount	0.20%
old_balance_orig	Originator account balance before txn	0.20%
surp_orig_zeroed	Originator account drained to zero	Derived
balance_drain_pct	Amount / originator balance	Derived
is_transfer_cashout	TRANSFER or CASH_OUT binary flag	Derived
night_txn_flag	Transaction between 22:00–06:00	Derived
error_balance_orig	Accounting discrepancy on originator side	Derived

The top 3 features are balance mechanics — fraudsters drain destination accounts to zero immediately, leaving detectable accounting inconsistencies. This mirrors real-world fraud ML findings from published PaySim research.

---

🤖 ML Models

Primary: XGBoost Classifier (300 estimators, depth 6, lr 0.05, SMOTE-balanced)
Baseline: Random Forest (200 estimators, balanced class weights, 5-fold CV)

Performance

Metric	XGBoost	Random Forest
Accuracy	1.00	1.00
Precision	1.00	1.00
Recall	1.00	1.00
F1 Score	1.00	1.00
ROC-AUC	1.00	1.00
PR-AUC	1.00	1.00
5-Fold CV AUC	—	1.00 ± 0.00

PR-AUC (Precision-Recall AUC) is the gold standard metric for fraud detection on imbalanced datasets — unlike ROC-AUC, it is sensitive to the minority class performance and does not inflate under class imbalance.

Confusion Matrix (Test Set — 2,000 transactions)

	Predicted Legit	Predicted Fraud
Actual Legit	1,978 ✓	0
Actual Fraud	0	22 ✓

---

📊 Dashboard Features

• 8 KPI cards — Txns, fraud cases, fraud amount, ROC-AUC, PR-AUC, F1, SMOTE ratio, false negatives
• Fraud vs Legit by Type — stacked bar (PAYMENT/TRANSFER/CASH_OUT/DEBIT/CASH_IN)
• Fraud distribution by hour — time-series line chart
• Alert tier distribution — doughnut
• Feature importance — colour-coded bars with importance percentages
• SMOTE explainer panel — before/after oversampling with rationale
• Model comparison — XGBoost vs Random Forest metrics + confusion matrix
• Fraud alert queue — 112 cases filterable by type, tier, and flag

---

🛠️ Tech Stack

Layer	Technology
Data Generation	Python · Pandas · NumPy (PaySim-inspired)
Class Imbalance	imblearn · SMOTE
ML Models	XGBoost · scikit-learn (Random Forest)
Feature Engineering	Balance error detection, log transforms, derived flags
Cross-Validation	StratifiedKFold (5-fold)
Evaluation	ROC-AUC · PR-AUC · F1 · Confusion Matrix
Visualisation	Chart.js · HTML/CSS · Fira Code font

---

▶️ How to Run

git clone https://github.com/ukishore33/fincrime-ml-model.git
cd fincrime-ml-model
pip install pandas numpy scikit-learn xgboost imbalanced-learn
python generate_data.py
python train_model.py
open fincrime_dashboard.html

Live Demo

---

👤 Author

Kishore U.
AML/KYC Compliance Analyst | Data Analytics
📱 6303308133 | Bengaluru, Karnataka | Immediate Joiner
🔗 LinkedIn · GitHub

---

📜 Disclaimer

All data is 100% synthetic — generated programmatically with no real financial or personal data. PaySim-inspired dataset structure based on published academic research. Built purely for portfolio demonstration.