Skip to content
All projects
Data Analytics+ AI/ML

Bank Transaction Fraud Analytics

Full EDA on a large-scale banking transaction dataset to surface spending patterns and behavioral anomalies, with supervised and unsupervised detectors benchmarked on precision, recall, and F1 rather than accuracy.

Key result

3 detectors— Benchmarked on precision, recall, and F1

01Problem

Fraudulent transactions are a tiny fraction of total volume, which makes them easy to miss and easy to fake success against — a model that predicts 'not fraud' every time still scores over 99% accuracy.

02Objective

Characterize fraud patterns through exploratory analysis, then compare supervised and unsupervised detectors on metrics that survive class imbalance.

03Architecture

  1. Transaction Data
  2. Exploratory Analysis
  3. Feature Engineering
  4. Detector Benchmarking
  5. Evaluation
Analysis and evaluation workflow

04Technology

  • Python
  • Pandas
  • scikit-learn
  • Random Forest
  • Isolation Forest
  • K-Means
  • Matplotlib

05Implementation

Key technical decisions

  1. 01Compared detectors on precision, recall, and F1 rather than accuracy, which is uninformative at this class imbalance.
  2. 02Included an unsupervised anomaly detector alongside the supervised model, since novel fraud has no label to learn from.
  3. 03Segmented customers by behavior to see whether flagged activity concentrates rather than spreading evenly.

What was built

  • Pandas pipeline for cleaning and per-customer aggregation.
  • Matplotlib distribution and correlation plots contrasting fraud against baseline behavior.
  • Random Forest as the supervised detector, Isolation Forest as the label-free comparison.
  • K-Means segmentation over behavioral features.
  • Precision, recall, and F1 reported for each detector.

06Data

Transaction dataset

  • Large-scale banking transaction records with a labeled fraud flag
  • Transaction amount, timestamp, channel, and merchant category
  • Customer identifiers supporting per-account aggregation
  • Severely imbalanced classes — fraud is a small minority of rows

07Results

3

Detectors benchmarked

P / R / F1

Evaluation metrics used

Segmented

Customers grouped by behavior

10Links

GitHub