💳 Credit Card Fraud Detection — Cost-Aware Pipeline

A practical fraud detection workflow with time-aware evaluation, probability calibration, and explicit threshold policies for controlling false positives vs false negatives.

Case study: CASE_STUDY.md

---

What this repository includes

• Notebook: Credit-Card-Fraud-Detection-A-Pipeline-Journey.ipynb
• Exported artifacts under ./artifacts/ (models + thresholds + run metadata)
• Scoring script: scripts/score_csv.py (optional)

---

Dataset

Source: Kaggle “Credit Card Fraud Detection” dataset (creditcard.csv).

Expected columns:

• Time, V1…V28, Amount, Class (0 = normal, 1 = fraud)

Local (recommended)

1. Download the dataset CSV.
2. Place it at: data/raw/creditcard.csv

Kaggle

The notebook also supports the common Kaggle input path: /kaggle/input/creditcardfraud/creditcard.csv

---

Getting started

1) Install

python -m venv .venv
# Windows: .\.venv\Scripts\activate
# macOS/Linux: source .venv/bin/activate

pip install -r requirements.txt

2) Run the notebook

Open and run:

• Credit-Card-Fraud-Detection-A-Pipeline-Journey.ipynb

The notebook will:

• train baseline + stronger models
• calibrate probabilities
• select threshold policies
• export artifacts to ./artifacts/

---

Artifacts

Artifacts are written to ./artifacts/ (models, thresholds, and run metadata). See artifacts/README.md for the expected files produced by the notebook.

---

Score a CSV (optional)

After exporting artifacts, you can score any CSV with the same feature columns:

python scripts/score_csv.py --csv data/raw/creditcard.csv --out artifacts/scored.csv --model xgb --policy min_cost

The output adds:

• fraud_proba
• fraud_pred (0/1)

---

Methodology notes

• Leak-safe evaluation via time-based train/test windows.
• Calibration produces probabilities suitable for threshold policies.
• Threshold policies define operating points (e.g., min expected cost).

---

License

MIT (code). Dataset licensing depends on the dataset source where you download it.