risk-engine 1.1.0

Explainable risk-based transaction anomaly detection engine with authentication and dashboard

Explainable Risk-Based Transaction Anomaly Engine

A secure, offline, explainable transaction anomaly detection engine designed for banking environments.

✨ New Features:

• 🔐 Email-based authentication for authorized users
• 📊 Interactive terminal dashboard with analysis history
• 🌐 Web-based visualization viewer for charts and results

🚀 Quick Start (For New Users)

Step 1: Install the Package

pip install risk-engine

Step 2: Register & Login

# Register a new user account
risk-engine register

# Login
risk-engine login

Step 3: Use the Dashboard

# Launch interactive dashboard
risk-engine

From the dashboard, you can:

• Run new analyses
• View analysis history
• Open web visualizations
• Track your statistics

Alternative: Direct CLI Usage

You can also run analyses directly without the dashboard:

risk-engine -i your_transactions.csv -o results/

View Results in Browser

risk-engine viewer -o results/

---

📚 Documentation

• Authentication & Dashboard Guide - Learn about user accounts, dashboard, and web viewer
• Publishing Guide - How to publish to PyPI

---

🚀 Quick Start (Classic Mode)

Step 1: Install the Package

pip install risk-engine

Step 2: Prepare Your Data

Your CSV file needs at minimum these columns:

• transaction_id - Unique ID for each transaction
• sender_account - The account making the transaction

For full functionality, also include: timestamp, amount, device_hash, ip_address, location

Step 3: Run the Analysis

risk-engine -i your_transactions.csv -o results/

Step 4: View Results

# See summary
cat results/summary.json

# View flagged transactions
head results/flagged_transactions.csv

# Check what triggered the flags
cat results/stats_reasons.csv

That's it! 🎉

---

Overview

The Risk Engine processes large transaction datasets in chunks, applies rule-based risk checks, and generates auditable anomaly alerts with full explainability for each flagged transaction.

Key Features

• Offline, Local Execution - No cloud dependencies, data stays on-premises
• Chunk-Based Processing - Handles 1GB+ CSV files with minimal memory footprint
• Explainable Risk Scoring - Each anomaly includes human-readable reasons
• Configurable Thresholds - Auto-calculated or manually overridden
• Velocity Simulation - Optional stress testing for rapid transaction patterns
• Multiple Output Formats - CSV and Parquet export

Installation

From PyPI (Recommended)

pip install risk-engine

From Source (Development)

git clone https://github.com/yourusername/risk-engine.git
cd risk-engine
pip install -e .

Usage

Basic Usage

risk-engine --input transactions.csv --output-dir outputs/

With Options

risk-engine \
  --input transactions.csv \
  --output-dir outputs/ \
  --threshold 4 \
  --simulation on \
  --chunk-size 100000

All CLI Options

Option	Short	Description	Default
--input	-i	Input CSV file (required)	-
--output-dir	-o	Output directory (required)	-
--threshold	-t	Risk score threshold	Auto-calculated
--simulation	-s	Velocity simulation (on/off)	off
--chunk-size	-c	Rows per processing chunk	500,000
--quiet	-q	Suppress progress output	False
--version		Show version	-

Python API

from risk_engine import run_engine, process_transactions

# Process a file
summary = run_engine(
    input_file="transactions.csv",
    output_dir="outputs/",
    threshold=4,
    simulation=True,
    chunk_size=100_000
)

# Or process a DataFrame directly
import pandas as pd
df = pd.read_csv("transactions.csv")
processed = process_transactions(df, simulation_mode=True)
flagged = processed[processed["final_is_anomalous"]]

Input Format

The engine expects a CSV with the following columns:

Required Columns

Column	Description
transaction_id	Unique transaction identifier
sender_account	Account identifier

Optional Feature Columns

Column	Description	Risk Check Enabled
timestamp	ISO8601 UTC timestamp	Off-hour detection
amount	Transaction amount	Z-score deviation
device_hash	Device fingerprint	New device detection
ip_address	IP address	New IP detection
location	Transaction location	Location change detection

The engine automatically adapts based on available columns - missing optional columns simply disable their corresponding risk checks.

Output Files

File	Description
flagged_transactions.csv	All anomalous transactions
flagged_transactions.parquet	Same data in Parquet format
stats_risk_scores.csv	Risk score distribution
stats_reasons.csv	Breakdown of anomaly reasons
summary.json	Processing summary

Risk Checks

Amount Deviation

Flags transactions where the amount is ≥1 standard deviation from the sender's mean.

New Device

Flags first-seen devices for each account.

New IP Address

Flags first-seen IP addresses for each account.

Location Change

Flags when transaction location differs from the previous transaction.

Off-Hour Activity

Flags transactions outside the sender's dominant transaction hour.

Velocity (Simulation Mode)

Groups transactions into 5-transaction sessions with 10-second intervals to detect rapid-fire patterns.

Threshold Calculation

The anomaly threshold is automatically calculated based on active risk checks:

Active Risks	Threshold	Ratio
1	1	100%
2	2	100%
3	3	100%
4	3	75%
5	4	80%
6+	5	83%

Use --threshold to override the automatic calculation.

Example Output

summary.json

{
  "total_transactions": 150000,
  "flagged_transactions": 7523,
  "anomaly_rate": 0.050153,
  "threshold": 4,
  "features_used": ["amount", "device_hash", "ip_address", "location", "timestamp"]
}

flagged_transactions.csv

transaction_id,sender_account,timestamp,amount,location,final_risk_score,final_reasons
TXN001,ACC123,2025-01-08T14:30:00Z,5000.00,Mumbai,4,"Unusual transaction amount; New device detected; Transaction location changed; Transaction at unusual time"

Requirements

• Python 3.9+
• pandas ≥ 2.0
• numpy ≥ 1.23
• pyarrow ≥ 14.0

---

📖 Interpreting Results

Understanding the Output Files

File	What It Contains	How to Use It
summary.json	Overall statistics	Quick health check - see total transactions, flagged count, anomaly rate
flagged_transactions.csv	All suspicious transactions	Main investigation file - review each flagged transaction
flagged_transactions.parquet	Same as above (binary format)	For loading into pandas/Spark for further analysis
stats_reasons.csv	Count of each risk type	Identify most common fraud patterns
stats_risk_scores.csv	Distribution of risk scores	Understand the risk profile of your data

Understanding Risk Scores

Each transaction gets a risk score from 0 to 6 (depending on active features):

Score	Meaning	Action
0	No risks detected	Normal transaction
1-2	Low risk	Usually normal, may warrant monitoring
3-4	Medium risk	Review recommended
5+	High risk	Flagged as anomalous - investigate

Understanding Anomaly Reasons

Each flagged transaction includes final_reasons explaining WHY it was flagged:

Reason	What It Means	Example
Unusual transaction amount	Amount differs significantly from user's normal spending	User typically spends ₹500-2000, but this transaction is ₹50,000
New device detected	First time this device is seen for this account	User logged in from a new phone
New IP address detected	First time this IP is seen for this account	Transaction from a new network/location
Transaction location changed	Different location from previous transaction	Previous: Mumbai, Current: Delhi
Transaction at unusual time	Transaction outside user's typical hours	User normally transacts at 10 AM, this is at 3 AM
Multiple transactions in short time	Rapid successive transactions (simulation mode)	5 transactions within 50 seconds

Example Analysis Workflow

# 1. Run the engine
risk-engine -i transactions.csv -o results/ --simulation on

# 2. Check summary
cat results/summary.json
# Output: {"total_transactions": 150000, "flagged_transactions": 304, "anomaly_rate": 0.002}

# 3. See most common fraud patterns
cat results/stats_reasons.csv
# Output:
# reason,count
# New device detected,304
# Transaction location changed,280
# Unusual transaction amount,150

# 4. Investigate flagged transactions
head -5 results/flagged_transactions.csv

What Does a Low/High Anomaly Rate Mean?

Anomaly Rate	Interpretation
< 0.1%	Very strict detection - only catching obvious anomalies
0.1% - 1%	Balanced detection - typical for production use
1% - 5%	Sensitive detection - may include false positives
> 5%	Very sensitive - consider raising threshold

Tip: Use --threshold to adjust sensitivity:

• Higher threshold (e.g., --threshold 5) = fewer flags, higher confidence
• Lower threshold (e.g., --threshold 3) = more flags, may catch more edge cases

---

🔧 Troubleshooting

Common Issues

"ModuleNotFoundError: No module named 'risk_engine'"

# Make sure you installed the package
cd /path/to/risk_engine
pip install -e .

"Error: Input file not found"

# Use absolute path
risk-engine -i /full/path/to/transactions.csv -o /full/path/to/output/

"Chunk size must be at least 1000 rows"

# Default chunk size is 500,000. For small files, just omit --chunk-size
risk-engine -i small_file.csv -o output/

High memory usage

# Reduce chunk size for large files
risk-engine -i huge_file.csv -o output/ --chunk-size 100000

---

License

MIT License - see LICENSE for details.