Model documentation

Model Description

Type: Supervised binary classifier

Algorithm: Gradient Boosted Decision Trees

Task: Fraud detection in financial transactions

Intended Use

Real-time fraud scoring

Risk assessment pipelines

Out-of-Scope Use

Legal or regulatory decision-making without human review

Ethical Considerations

Potential bias due to historical data

Regular fairness evaluations required

Training Documentation

Dataset Characteristics

Volume: ~50 million records

Feature count: ~120

Label distribution: Highly imbalanced

Training Methodology

Stratified sampling

Cross-validation

Hyperparameter optimization

Reproducibility

Fixed random seeds

Versioned datasets and code

Performance Metrics

Offline Metrics

AUC-ROC: 0.97

Precision at 90% Recall: 0.91

Online Metrics

Average latency: 42 ms

Error rate: < 0.1%