Predictive Financial Fraud Detection System

Context

Mobile Money payments move a large number of small transactions every day. Fraud is rare among them, which makes it hard to spot, whether by hand or with a poorly calibrated model.

Problem

The data is highly imbalanced: fraudulent transactions are a tiny share of the total. A model that answers “no fraud” everywhere gets a misleading accuracy while detecting nothing. The project needed a suitable metric and a model able to catch the rare class.

Solution

• Supervised classification pipeline with Scikit-Learn and Pandas.
• Random Forest and decision trees compared on the same pipeline.
• Evaluated with the F1-score rather than overall accuracy, to account for the imbalance.
• Results visualised with Matplotlib and Seaborn.

Results

• 96.4% F1-score on highly imbalanced data, with the Random Forest.
• Project completed and defended; not yet deployed to production.