Machine Learning

Fraud Detection ML Models & Performance

Advanced ensemble models achieving 96%+ accuracy with continuous learning and drift detection

+1.2%
96.3%
Overall Accuracy
+0.8%
94.7%
Precision
+1.5%
95.9%
Recall
+1.1%
95.3%
F1 Score
Ensemble Model Architecture
Four-model ensemble with weighted voting

Random Forest

25% weight
Accuracy94.2%
Specs: 500 trees, max depth 15, min samples 20
Strength: Handles non-linear patterns, robust to outliers

XGBoost

30% weight
Accuracy95.1%
Specs: 1000 rounds, learning rate 0.05, early stopping
Strength: Best overall performance, handles missing values

Deep Neural Network

30% weight
Accuracy96.3%
Specs: 5 layers (256-128-64-32-1), dropout 0.3, Adam optimizer
Strength: Captures complex patterns, learns representations

LSTM Sequence Model

15% weight
Accuracy93.7%
Specs: Bidirectional, 30-day lookback, attention mechanism
Strength: Temporal pattern detection, velocity analysis
Voting Strategy
Weighted soft voting with confidence thresholds
Deep Neural Network (DNN)
30%
XGBoost
30%
Random Forest
25%
LSTM
15%

Final prediction uses weighted soft voting where each model contributes its confidence score multiplied by its weight. Threshold: 0.7 for fraud classification. If ensemble confidence < 0.85, transaction flagged for human review.