Financial Transaction Pattern Analyzer

Detecting unusual spending behavior in business transaction data using feature engineering and unsupervised anomaly detection.

---

Overview

Problem: Given transaction data for businesses, can we identify unusual spending patterns that may indicate fraud, accounting errors, or behavior shifts?

This project explores that question through a realistic exploratory prototype — not a production system. The emphasis is on feature engineering and interpretability, not model complexity.

Notebook

Open transaction_analysis.ipynb to see the full analysis:

1. Synthetic dataset generation (6,400 transactions, 5 businesses, 12 months)
2. Exploratory analysis
3. Feature engineering — 10 contextual features
4. Three anomaly detection methods compared
5. Visualizations and results
6. Honest limitations and next steps
7. SQL implementation — core anomaly logic in SQLite

Key Idea

Raw transaction amounts are poor anomaly signals. A $10,000 payment is routine for one business and alarming for another. The interesting work is building context-aware features:

Feature	What it captures
amount_zscore_biz	Unusual for this account's history?
amount_zscore_vendor	Unusual for this vendor relationship?
is_new_vendor	First-ever payment to this payee?
rolling_7d_count	Transaction frequency spike this week?
is_near_duplicate	Same amount + same vendor within 48 hrs?

Methods

Method	Description
Rule-based scoring	Human-readable thresholds, fully auditable
Z-score composite	Statistical baseline, SQL-friendly
Isolation Forest	Tree-based unsupervised ML
SQL queries	Core anomaly logic in SQLite — production-ready

Stack

Python · pandas · NumPy · scikit-learn · matplotlib · SQLite

---

A portfolio project exploring anomaly detection in a fintech/business banking context.