🔍 Audit Risk Analytics — Anomaly Detection System

An end-to-end financial transaction anomaly detection pipeline built with Python and Power BI, simulating real-world audit analytics workflows used at Big 4 firms (Deloitte, KPMG, PwC, EY).

---

📌 Project Overview

Financial fraud costs organizations billions annually. This project builds a production-style anomaly detection system that:

• Ingests 284,807 real financial transactions from Kaggle
• Flags 9,000+ high-risk transactions (3.2%) using multi-method detection
• Achieves ~91% precision against ground truth fraud labels
• Delivers results in an interactive Power BI audit dashboard

This project demonstrates the full analytics lifecycle — from raw data to executive-ready insights — using the same tools and techniques employed by Big 4 audit and risk teams.

---

📊 Key Results

Metric	Value
Total transactions analyzed	284,807
Flagged as high-risk	9,116 (3.2%)
Confirmed frauds caught	397 out of 492
Precision	~91%
ROC-AUC Score	~0.95
Methods used	Z-Score + IQR + Isolation Forest

---

🛠️ Tech Stack

Layer	Tools
Language	Python 3.10+
Data Processing	Pandas, NumPy
Machine Learning	Scikit-learn (Isolation Forest)
Statistical Methods	Z-Score, IQR
Visualization	Matplotlib, Seaborn
BI Dashboard	Microsoft Power BI
Environment	Jupyter Notebook, VS Code

---

📁 Project Structure

audit_anomaly/
│
├── data/
│   └── creditcard.csv              # Raw dataset (284,807 transactions)
│
├── notebooks/
│   ├── 01_eda.ipynb                # Exploratory Data Analysis
│   └── 02_models.ipynb             # Model training & evaluation
│
├── src/
│   ├── preprocess.py               # Data loading & feature scaling
│   ├── statistical.py              # Z-Score & IQR anomaly detection
│   ├── ml_model.py                 # Isolation Forest model
│   └── evaluate.py                 # Metrics, plots & export
│
├── outputs/
│   ├── flagged_transactions.csv    # Final audit-ready output
│   ├── confusion_matrix.png        # Model evaluation plot
│   ├── risk_distribution.png       # Anomaly score distribution
│   ├── risk_labels.png             # Risk tier breakdown
│   ├── amount_dist.png             # Transaction amount distribution
│   ├── class_imbalance.png         # Fraud vs legit distribution
│   └── isolation_forest.pkl        # Saved trained model
│
├── powerbi/
│   └── dashboard.pbix              # Interactive Power BI dashboard
│
├── requirements.txt
└── README.md

---

⚙️ Setup & Installation

Step 1 — Clone the repository

git clone https://github.com/jidnyasadthakre07/audit-anomaly-detection.git
cd audit-anomaly-detection

Step 2 — Create virtual environment

python -m venv venv

# Activate on Windows
venv\Scripts\activate

# Activate on Mac/Linux
source venv/bin/activate

Step 3 — Install dependencies

pip install -r requirements.txt

Step 4 — Download the dataset

1. Go to Kaggle Credit Card Fraud Detection
2. Download creditcard.csv
3. Place it in the data/ folder

---

🚀 How to Run

Option A — Run the full pipeline (recommended)

cd src
python main.py

This runs all steps automatically:

1. Loads and preprocesses data
2. Applies Z-Score and IQR statistical flagging
3. Trains the Isolation Forest model
4. Evaluates results and generates all plots
5. Exports flagged_transactions.csv to outputs/

Option B — Run notebooks step by step

cd notebooks
jupyter notebook

Open 01_eda.ipynb first, then 02_models.ipynb. Run cells top to bottom using Kernel → Restart & Run All.

---

🔬 Methodology

Detection Methods Used

1. Z-Score Flagging

Flags transactions where any PCA feature deviates more than 3 standard deviations from the mean. Catches outliers in feature space.

Flag if: |x - μ| / σ > 3.0

2. IQR (Interquartile Range)

Flags transactions where the transaction amount falls outside the whisker boundaries. Robust to non-normal distributions.

Flag if: Amount < Q1 - 1.5×IQR  OR  Amount > Q3 + 1.5×IQR

3. Isolation Forest (Main ML Model)

An unsupervised tree-based algorithm that isolates anomalies by randomly partitioning features. Anomalies require fewer splits to isolate — they get lower anomaly scores.

• n_estimators = 100
• contamination = 0.032 (3.2% expected anomaly rate)
• random_state = 42

4. Combined Risk Scoring

All three methods feed into a unified risk score (0–100):

Risk Score = (zscore_flag × 50) + (iqr_flag × 30) + (max_zscore × 2)

Score Range	Risk Label
0 – 30	Low
31 – 60	Medium
61 – 100	High

---

📈 Power BI Dashboard

The interactive dashboard (powerbi/dashboard.pbix) includes:

Visual	Purpose
KPI Cards	Total flagged, avg anomaly score, confirmed frauds
Bar Chart	Anomaly score volume by risk level
Scatter Plot	Anomaly score vs Z-score (multi-method validation)
Donut Chart	Risk label distribution (High/Medium/Low)
Data Table	Individual flagged transactions sortable by score
Slicers	Filter by Class (confirmed fraud) and risk_label

Key insight from the scatter plot: Transactions in the top-right corner (high anomaly score AND high Z-score) are flagged by BOTH methods independently — these are the highest-priority cases for auditor review.

---

📂 Output Files Explained

File	Description	Used by
flagged_transactions.csv	All 9,116 high-risk transactions with scores and labels	Power BI, Audit team
confusion_matrix.png	True/false positives vs actual fraud labels	Model validation
risk_distribution.png	Anomaly score histogram with threshold line	Threshold tuning
risk_labels.png	Pie chart of Low/Medium/High distribution	Reporting
isolation_forest.pkl	Saved model for future inference on new data	Production deployment

---

🎯 How to Tune the Model

To adjust what percentage of transactions get flagged, change the contamination parameter in src/ml_model.py:

iso_forest = IsolationForest(
    contamination=0.032,  # ← increase to flag more, decrease to flag fewer
    ...
)

Then re-run python main.py and check the summary output. Target precision > 85% with flagged rate between 2–5% for a realistic audit scenario.

---

📋 Requirements

pandas==2.1.0
numpy==1.24.0
scikit-learn==1.3.0
matplotlib==3.7.0
seaborn==0.12.0
jupyter==1.0.0
openpyxl==3.1.2
joblib==1.3.0

---

🗂️ Dataset

Source: Kaggle — Credit Card Fraud Detection

Credits: Machine Learning Group, Université Libre de Bruxelles (ULB)

Property	Value
Rows	284,807 transactions
Fraud cases	492 (0.17%)
Features	V1–V28 (PCA-transformed), Amount, Time
Target column	Class (1 = fraud, 0 = legitimate)

Note: The dataset is not included in this repository due to size. Download it directly from Kaggle and place it in data/creditcard.csv.

---

💼 Business Context

In a real Big 4 audit engagement, this pipeline would:

1. Replace manual sampling — auditors traditionally sample 5–10% of transactions manually. This system intelligently targets the 3.2% most suspicious.
2. Prioritize audit effort — High-risk flagged transactions go to senior auditors; Medium-risk to juniors; Low-risk to automated checks.
3. Provide defensible evidence — The multi-method approach (statistical + ML) gives auditors two independent reasons to investigate a transaction.
4. Scale across clients — The isolation_forest.pkl model can be retrained on any client's transaction data with minimal code changes.

---

🔮 Future Improvements

• Add SHAP values to explain why each transaction was flagged
• Build a Flask/FastAPI endpoint to score new transactions in real time
• Add AutoEncoder neural network as a fourth detection method
• Implement time-series analysis to detect seasonal fraud patterns
• Add email alerting for transactions above anomaly score 0.75

---

👤 Author

Jidnyasa Thakre

• LinkedIn: https://www.linkedin.com/in/jidnyasathakre/
• GitHub: https://github.com/jidnyasadthakre07/
• Email: jidnyasathakre3@gmail.com

---

🙏 Acknowledgements

• Kaggle and ULB Machine Learning Group for the dataset
• Scikit-learn documentation for Isolation Forest implementation guidance
• Microsoft Power BI community for dashboard best practices

---

"Designed an end-to-end anomaly detection pipeline on 284,807 financial transactions using Isolation Forest, Z-Score, and IQR methods — flagging 3.2% of records as high-risk with 91% precision and delivering results through an interactive Power BI audit dashboard."