Banking Fraud Detection Dashboard

Interactive overview of fraud patterns, model performance, and risk segmentation — 11,903 transactions analysed

Fraud Rate
7.7%
916 of 11,903 transactions
Best Model AUC
0.95+
Gradient Boosting (temporal split)
Features Used
40+
Transaction + clickstream
Highest Risk
89.3%
International + Gift Card
Fraud Captured
85%+
High Risk tier

Transaction Fraud Patterns

How fraud distributes across transaction types, time, and flag combinations

Fraud Rate by Transaction Type Highest Risk Flags
Flag Interaction: International x Gift Card Critical
No Gift CardGift Card
Domestic 5.2% 36.8%
International 45.3% 89.3%
Key Finding
When both international and gift card flags are active, 89.3% of transactions are fraudulent. This single rule could catch a significant portion of fraud with minimal false positives.
Fraud Rate by Hour of Day Peak: 1-7 AM
Fraud Rate by Day of Week Weekend Effect

Clickstream Behavioural Signals

Browsing patterns that distinguish fraudulent sessions from legitimate ones

Sensitive Page Visits Strongest Signal
Fraud sessions visit 2.5x more sensitive pages (payment, address, password)
Dwell Time: Forgot Password Stolen Creds
Fraudsters spend 77s vs 32s — struggling with stolen credentials
Payment Page Dwell Time Pre-planned
Fraudsters are faster at payment (74s vs 92s) — pre-planned, not browsing
Navigation Entropy (Shannon)
Fraud sessions are more erratic: entropy 3.16 vs 2.99 for legitimate users
Direct-to-Payment Rate
6.2% of fraud sessions skip browsing entirely vs 3.4% legitimate

Model Performance

Four models trained with temporal validation and hyperparameter tuning

ROC-AUC Comparison (Tuned Models) All > 0.82
Model Summary
ModelAUCStrength
Gradient Boosting ~0.95 Best overall
Random Forest ~0.93 Robust ensemble
Logistic Regression ~0.92 Interpretable
Decision Tree ~0.82 Readable rules
Statistical Validation
Paired t-tests on 10-fold CV confirm AUC differences are statistically significant (p<0.05), not random noise.

Risk Tier Segmentation

Data-driven thresholds optimised for maximum fraud capture with minimum review burden

Transactions per Tier
Fraud Rate per Tier
% of Fraud Captured Target: 85%+

Recommendations

Immediate Actions
1. Flag International + Gift Card
Mandatory review for all transactions with both flags active. At 89.3% fraud rate, this is the highest-yield rule.
2. Deploy Scoring Model
Use the Gradient Boosting model for real-time risk scoring with the F1-optimised threshold.
3. Early Morning Staffing
Increase fraud review coverage during 1-7 AM when fraud rates peak.
Strategic Improvements
4. Integrate Clickstream Monitoring
Sensitive page counts, navigation entropy, and dwell times are hard for fraudsters to fake.
5. Monitor for Drift
Track AUC and probability distributions weekly. Retrain when AUC drops below 0.88.
6. A/B Test Thresholds
Validate risk tier boundaries on live data to confirm business impact before full deployment.