🚀 BidWise – Real-Time Bidding (RTB) DSP Optimization Engine

AI-powered Demand Side Platform (DSP) engine for optimizing Real-Time Bidding under budget constraints using machine learning and expected value optimization.

---

📌 Overview

BidWise is a machine learning–driven Real-Time Bidding (RTB) engine designed to maximize:

Score = Clicks + N × Conversions

under strict constraints:

• 💰 Fixed advertiser budget
• ⚖ Second-price auction mechanism
• ⏱ Low-latency inference
• 🧠 Sequential decision making (no future request knowledge)
• 📦 Memory-efficient implementation

This project simulates a real-world ad-tech bidding system used in programmatic advertising platforms.

---

🧠 Core Concepts

🔹 CTR Prediction (Click-Through Rate)

Predicts probability that a user clicks an ad.

🔹 CVR Prediction (Conversion Rate)

Predicts probability that a user converts after clicking.

🔹 Expected Value Optimization

EV = P(click) + N × P(conversion)

🔹 Budget-Aware Dynamic Bidding

Bid = EV × BaseBid × BudgetFactor

Where:

BudgetFactor = RemainingBudget / InitialBudget

🔹 Second-Price Auction Logic

If:

Bid ≥ Market Price

• Win auction
• Pay second-highest price (market price)
• Deduct budget
• Update score

---

🛠 Tech Stack

Backend

• Python 3.9
• Pandas
• NumPy
• Scikit-learn (Logistic Regression)
• Joblib

Machine Learning

• CTR Logistic Regression Model
• CVR Logistic Regression Model
• ROC-AUC Validation
• Vectorized Inference Optimization

Frontend (Planned)

• Next.js
• Tailwind CSS
• Recharts / Chart.js
• Light/Dark Theme Toggle

---

🏗 Project Structure

BIDWISE_RTB/
│
├── backend/
│   ├── data/
│   │   └── train.csv
│   │
│   ├── models/
│   │   ├── ctr_model.pkl
│   │   └── cvr_model.pkl
│   │
│   ├── rtb_engine/
│   │   ├── predictor.py
│   │   ├── budget_manager.py
│   │   ├── strategy.py
│   │   ├── simulator.py
│   │   └── dataset_generator.py
│   │
│   ├── train_models.py
│   └── main.py
│
├── frontend/
└── README.md

---

📊 Dataset

Synthetic RTB dataset (50,000 rows) with realistic patterns:

• Evening hours → Higher CTR
• Campaign 2 & 3 → Better performance
• Mobile → Higher conversion rate
• Conversion depends on click probability

Columns:

• impression_id
• campaign_id
• hour
• device_type
• floor_price
• market_price
• click
• conversion

---

⚡ Performance Optimization

• Vectorized CTR & CVR prediction
• Reduced per-row ML inference
• Budget-aware early stopping
• Lightweight Logistic Regression models
• Optimized simulation loop

---

📈 Baseline vs Optimized Strategy

The system compares:

🔹 Baseline Strategy

Fixed bid for every impression.

🔹 Optimized Strategy

Expected Value–based dynamic bidding with budget scaling.

This demonstrates measurable improvement in score under budget constraints.

---

🚀 How To Run

1️⃣ Install Dependencies

pip install -r backend/requirements.txt

2️⃣ Generate Dataset

python backend/rtb_engine/dataset_generator.py

3️⃣ Train Models

cd backend
python train_models.py

4️⃣ Run Simulation

python main.py

---

🎯 Key Learning Outcomes

This project demonstrates:

• Real-world ad-tech optimization logic
• Budget-constrained decision systems
• Applied machine learning in auction environments
• Performance-aware algorithm design
• Sequential online optimization

---

🔮 Future Improvements

• Advanced budget pacing algorithm
• Multi-armed bandit strategy
• Reinforcement learning optimization
• FastAPI backend integration
• Modern analytics dashboard (Next.js)
• Deployment to cloud

---

👨‍💻 Author

Developed as part of a Hackathon RTB Optimization Challenge.