FraudShield - Fraud Detection System

About Me & This System

Portfolio profile up top, then a full technical walkthrough of the fraud pipeline, design decisions, and core references.

Links

🌐Portfolio LinkedIn GitHub ✉️Email

Don't hesitate to reach out.

System Architecture

How raw transaction input moves through feature engineering, model inference, decision logic, and impact analysis.

Pipeline Flow

1) Data + Feature Layer

Builds behavior, geo, temporal, and account-risk features (distance, unusual time, prior fraud, velocity, international mismatch).

2) Model Benchmarking

Compares LR, RF, XGBoost and other candidates with stratified CV; best model selected using imbalance-aware metrics.

3) Decision Layer

Combines ML probability with rule triggers to output practical actions: APPROVE, REVIEW, or BLOCK with full traceability.

4) Explainability

Uses SHAP (global + local) so each score is auditable and understandable for stakeholders and reviewers.

5) Business Impact Layer

Converts precision/recall trade-offs into operational cost, savings, and threshold-sensitive impact projections.

Feature Design Rationale

Behavioral drift features capture deviations from personal norms (frequency and amount spikes).
Geographic inconsistency (home vs transaction location) surfaces high-friction fraud patterns.
Merchant novelty and prior fraud history improve early-risk signal strength.
Temporal cues (late-night activity, weekend patterns) improve separability in edge cases.
Rules intentionally complement the model for high-severity scenarios where recall matters most.

How To Read This Dashboard

Exploratory: Understand fraud hotspots and pattern concentration.
Model Performance: Compare metrics, calibration, confusion matrix, and threshold tuning.
Live Scorer: Score one transaction with explanation, decision trace, and risk signals.
Business Impact: Translate model behavior into FP/FN costs and projected savings.
Goal: Prioritize explainable, decision-ready fraud triage over raw model accuracy alone.

Reference Papers

Davis & Goadrich (2006) — Why Precision-Recall curves are preferred for imbalanced classification.
Breiman (2001) — Random Forest methodology and variance reduction.
Chen & Guestrin (2016) — XGBoost system design for scalable gradient boosting.
Lundberg & Lee (2017) — SHAP unified framework for model explanations.
Niculescu-Mizil & Caruana (2005) — Probability calibration insights for supervised models.

Fraud Patterns at a Glance

Where, when, and how fraud occurs across all banking transactions.

Fraud Rate by Transaction Type

Fraud Rate by Merchant Category

Fraud Rate by City

Fraud Rate by Hour of Day

Amount Distribution — Fraud vs Normal

Fraud Rate by Prior Fraud History

Fraud Rate — International × New Merchant Combinations

Key Insights

🌍

International + New Merchant

The combined signal carries the highest fraud rate — strongest paired indicator in the dataset.

🕐

Late-Night Spike

Fraud peaks after 22:00. Unusual_Time_Transaction ranks top-3 across all models.

📍

Distance from Home

Fraudulent transactions occur on average 2× farther from home — a strong standalone risk signal.

📈

Prior Fraud = #1 Predictor

Even one prior fraud incident dramatically raises probability. Recidivism is the strongest individual feature.

Model Comparison & Evaluation

Classifier benchmark · 5-fold stratified CV · SHAP explainability · Decision threshold optimiser

Model Comparison

Loading…

ROC Curve

Precision-Recall Curve

Confusion Matrix — Best Model @ 0.5

Loading…

Feature Importance

Global SHAP — Mean |SHAP|

Calibration Curve

Threshold Optimiser

Fraud classifiers are never deployed at 0.5. Drag to see precision, recall, and F1 trade off in real time.

Threshold

0.40

—

Precision

—

Recall

—

False Pos

—

False Neg

Precision / Recall / F1 vs Threshold

Model Card

Loading…

Classification Report

Loading…

Live Fraud Scorer

Enter a transaction for instant fraud probability, SHAP explanation, and risk breakdown.

Transaction Details

Amount ($)

Account Balance ($)

Distance from Home (km) auto

Time (HH:MM)

Transaction Type

Merchant Category

Card Type

Transaction Location

Home Location

Distance and international status update automatically from the selected locations.

Behavioural Signals

Daily Tx

Weekly Tx

Avg Amount ($)

Max 24h ($)

Failed Tx

Prev Fraud

Risk Flags

International?

Auto-derived from the selected locations.

New Merchant?

Unusual Time?

🛡️

Fill in the transaction
details and click Analyse

Business Impact Analysis

Translating ML metrics into financial outcomes using a configurable cost matrix.

Cost Matrix Configuration

Cost per False Negative (Missed Fraud, $) Default = avg fraud transaction amount

Cost per False Positive (Blocked Legit, $) Investigation + customer churn overhead

Expected Monthly Transactions

🛡️ Fraud Caught / Mo

—

value recovered at optimal threshold

⚠️ Fraud Missed / Mo

—

escaped detection cost

🔍 FP Investigation Cost

—

false alarm review overhead

📈 Net Monthly Benefit

—

caught minus FP costs

Business Cost Curve vs Threshold

Monthly Savings vs No-Model Baseline

Year-End Projections Test set scaled to monthly volume × 12

💰 Est. Annual Savings

—

vs. no-model baseline

📊 Est. Annual Remaining Cost

—

FN + FP costs at optimal threshold

✅ Fraud Txns Caught / Month

—

true positives at optimal threshold

How To Read These Numbers

Cost-Optimal Threshold—

Expected Alerts / Month—

Precision At Optimum—

Fraud Capture Rate—

Miss Rate At Optimum—

Cost Reduction vs No-Model—

ASSUMPTIONS — Cost figures are illustrative and denominated in the same units as the dataset. FP/FN costs should be calibrated to your institution's actual cost structure. Projections scale linearly from the test set to monthly volume.