Credit Card Fraud Detection

A machine learning pipeline for detecting fraudulent bank transactions — covering SQL-based investigation, class imbalance handling (SMOTE), model comparison, and an interactive Tableau dashboard.

Built by: Xintian (Cindy) Shi · LinkedIn · GitHub

---

Problem

Card fraud costs the global banking industry over $30 billion annually. Detecting it is hard: fraud is rare (0.17% of transactions), patterns shift constantly, and false positives — blocking legitimate customers — erode trust. This project builds a detection pipeline that balances precision and recall with that tradeoff in mind.

---

Dataset

Credit Card Fraud Detection — ULB Machine Learning Group

• 284,807 real European bank transactions over two days
• 492 confirmed fraud cases (0.172% of total)
• Features: Time, Amount, 28 PCA-anonymized components (V1–V28), Class

---

Key Findings (SQL + EDA)

• Fraud peaks at 2am–4am — 2am has a 1.71% fraud rate vs. 0.17% average; consistent with automated card-testing attacks when human review is minimal
• 13.8% of fraud transactions are under $1.00 — a known card-testing pattern where fraudsters verify stolen cards with micro-charges before making large purchases
• Average fraud amount: $122.21 vs. $88.29 for legitimate — fraudsters target higher-value transactions once a card is confirmed active
• V14, V10, V12 are the strongest fraud signal features in the PCA space (confirmed by both correlation analysis and Random Forest feature importance)

---

Approach

Step	Method
Storage & investigation	SQLite + 7 analytical SQL queries
EDA	pandas, NumPy, matplotlib, seaborn
Class imbalance	SMOTE — training set balanced to 50/50 (228K rows)
Baseline model	Logistic Regression
Main model	Random Forest (100 estimators)
Evaluation	Precision, Recall, F1, AUC-ROC, Precision-Recall Curve
Dashboard	Tableau Public

---

Model Results

Model	Precision	Recall	F1	AUC-ROC	PR-AUC
Logistic Regression	0.056	0.918	0.105	0.971	0.728
Random Forest (tuned)	0.842	0.816	0.829	0.967	0.870
XGBoost	0.697	0.867	0.773	0.979	0.871

Logistic Regression catches nearly all fraud (91.8% recall) but floods analysts with false positives — 94% of its fraud flags are wrong. Random Forest achieves the best precision (84.2%) — 5 out of 6 flagged transactions are genuine fraud. XGBoost leads on AUC-ROC and PR-AUC, making it the strongest overall ranker.

Top predictive features: V14, V4, V10, V17, V12 — consistent across both tree models

---

Business Impact

At a 95% precision threshold on the held-out test set (56,962 transactions, XGBoost):

	Value
Transactions flagged	82
True fraud caught	78 (79.6% recall)
False alarms	4
Fraud missed	20
Estimated loss prevented	$9,532.38

At 95% precision the model blocks only 4 legitimate cards to catch 78 fraud cases — a 19.5:1 true-positive ratio. Lowering the threshold to 80% precision catches more fraud at the cost of more false alarms; the Precision-Recall curve in 03_models.ipynb shows the full tradeoff.

---

Dashboard

View on Tableau Public →

---

How to Run

# 1. Clone repo and set up environment
git clone https://github.com/SXT2918/fraud-detection.git
cd fraud-detection
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 2. Download dataset
# Go to kaggle.com/datasets/mlg-ulb/creditcardfraud
# Unzip archive.zip and place creditcard.csv inside data/

# 3. Build SQLite database
python src/database.py

# 4. Run SQL investigation
sqlite3 data/fraud.db < sql/explore.sql

# 5. Open notebooks in order
jupyter notebook notebooks/

---

Project Structure

fraud-detection/
├── data/               # creditcard.csv + fraud.db (gitignored)
├── sql/
│   └── explore.sql     # 7 investigative SQL queries
├── notebooks/
│   ├── 01_eda.ipynb    # EDA: class imbalance, amount, time, correlations
│   ├── 02_features.ipynb  # SMOTE, scaling, train/test split
│   └── 03_models.ipynb    # Logistic Regression, Random Forest, PR curves
├── src/
│   └── database.py     # Loads CSV into SQLite, creates hourly_summary view
├── tableau/            # Exported PNGs + CSVs for dashboard
└── requirements.txt

---

Tech Stack

Python · pandas · NumPy · scikit-learn · imbalanced-learn · SQLite · matplotlib · seaborn · Tableau Public · Jupyter