Back to Portfolio

Cost-Optimized Fraud Detection

Minimizing financial loss through business-aligned Machine Learning.

The 'Business First' Problem

In real-world fraud detection, accuracy is a misleading metric. A False Negative (missing a fraudulent transaction) can result in a $100 loss, while a False Positive (flagging a normal transaction) only incurs a $5 operational review cost.

The true objective of this project wasn't achieving the highest overall accuracy, but engineering a model that achieved the lowest Total Financial Loss.

Key Results

  • Precision 90%+
  • Cost Reduction 80%
  • Status Deployment Ready

Technical Workflow

Data Handling: Managed an extreme 0.17% class imbalance.

Leakage-Safe Pipeline: Utilized imblearn pipelines to ensure scaling and SMOTE resampling strictly occurred during cross-validation, preventing data leakage.

Modeling: Trained Random Forest and XGBoost algorithms, specifically optimized for Precision-Recall AUC (0.85).

The Breakthrough

The standard default probability threshold of 0.5 is arbitrary and mathematically agnostic to business logic. By calculating a custom cost matrix, I executed a Threshold Sweep to locate the exact mathematical point where financial bleeding stops.

Adjusting the probability cutoff to 0.69 drastically reduced False Positives without sacrificing critical True Positives, achieving an 80% reduction in total financial loss compared to the baseline.

Explainability & Transparency

Black-box models are a liability in financial institutions. I implemented SHAP (SHapley Additive exPlanations) values to completely decode the model's decision-making process.

By proving mathematically which specific latent features (such as V14 and V4) drove the underlying fraud signals, the model became fully transparent, trustworthy, and actionable for downstream bank analysts.

View Project on GitHub

Explore More Projects

Fraud Detection
Product Analytics