Real-time Risk Control Strategy Engine

🛡️ System Overview

Role: Algorithm Engineer & Strategist Tech Stack: Python, Scikit-learn, SMOTE, XGBoost (Implied), Pandas

This project is a comprehensive Fraud Detection System designed to identify high-risk transactions in real-time. Unlike standard classification tasks, this engine focuses on handling extreme class imbalance (1.1% fraud rate) and minimizing financial loss through a hybrid strategy of machine learning and rule-based heuristics.

---

🏗️ Architecture & Logic

1. Geospatial Intelligence Module

Enriches raw transaction streams with location context to detect "Impossible Travel" and "IP Hopping."

• Technical Challenge: High-latency IP lookups for 138k+ records.
• Optimization: Implemented a vectorized range-lookup algorithm (reducing lookup time by 90% vs iterative approach) to simulate low-latency production requirements.

2. Feature Engineering (The "Signal" Layer)

Constructed 20+ behavioral features to capture Fraud Patterns:

• Velocity Checks: interval_after_signup (Detects bot-farm "signup-and-buy" behavior).
• Device Fingerprinting: n_dev_shared (Identifies device farming rings).
• Temporal Analysis: High-risk time windows based on purchase_seconds_of_day.

3. Decision Engine (The "Brain")

A tiered decision funnel designed to balance Recall (Catching Fraud) vs. Precision (User Experience).

• Layer 1: Rules Engine (Block known bad IPs/Devices).
• Layer 2: ML Probability Score (Random Forest + SMOTE).
• Layer 3: Risk Scoring (0-10) for manual review queues.

---

📊 Performance & Business Impact

Model Strategy: Random Forest + SMOTE We prioritized Recall (catching fraud) over Precision because the cost of a Chargeback ($) is significantly higher than the cost of a manual review.

Metric	Result	Business Implication
Recall	High	Captures the majority of fraud attacks, minimizing direct financial loss.
ROC-AUC	Excellent	Strong separation capability between legitimate users and attackers.
Latency	<60ms	(Simulated) Optimized for real-time checkout flows.

---

💡 Key Fraud Insights (Behavioral Analysis)

• The "1-Second" Rule: 50% of fraudulent transactions occur within 1 second of signup, indicating automated bot attacks rather than human behavior.
• Device Velocity: Legitimate users rarely share devices with >2 accounts. Accounts with n_dev_shared > 3 showed a 95% fraud probability.

---

💻 Usage & Reproduction

Prerequisites

pip install -r requirements.txt

# Execute the full pipeline (ETL -> Feature Eng -> Training)
jupyter notebook run_strategy_engine.ipynb