🚗 Lead Generation Prediction for Vehicle Listings

🎯 Project Overview

This machine learning project predicts the number of leads generated by vehicle advertisements using advanced feature engineering and gradient boosting algorithms. The solution helps automotive businesses optimise listings, pricing, and marketing strategies based on data-driven predictions.

Main Goal: Predict the number of leads (customer inquiries) that a vehicle listing will generate, based on vehicle characteristics, pricing, location, and advertisement features.

---

🏆 Model Performance & Results

Key Achievements

Metric	Value
📊 RMSE	6.922 leads
📈 R²	70.3% variance explained
⚡ Inference speed	< 1ms per listing
📋 Features used	16 (down from 48)
🔼 Baseline improvement	64.6%

Model Specifications

• Algorithm: LightGBM Regressor with Optuna hyperparameter optimisation (150 trials)
• Training data: 48,578 vehicle listings
• Validation: 5-fold cross-validation with overfitting detection
• Feature engineering: Target encoding, Jenks flag clustering, outlier removal

Business Impact Categories

Lead Range	Category	Recommended Action
0–5	🔴 Low Performance	Review listing quality, adjust pricing
6–15	🟡 Moderate Performance	Minor adjustments, standard monitoring
16–30	🟢 High Performance	Replicate success factors, scale strategy
31+	🌟 Exceptional Performance	Case study, premium placement

---

🔍 What Was Accomplished

1. Data Analysis & Cleaning

• Analysed 48,578 listings across multiple Brazilian states and cities
• Reduced 48 raw features to 16 optimised features (67% reduction, zero accuracy loss)
• Systematic outlier removal and missing value handling

2. Advanced Feature Engineering

• Geographic encoding: City/state target encoding with smoothing (prevents overfitting)
• Flag clustering: Jenks Natural Breaks to group vehicle feature combinations
• Price positioning: Market value vs. advertised price gap analysis
• Visual impact: Photo count optimisation (sweet spot: 8 photos)

3. Model Development

• Hyperparameter search via Optuna (150 trials)
• Learning curve analysis for regularisation guidance
• Sklearn-compatible production pipeline (.joblib)

4. Business Intelligence

• Top drivers: phone clicks, views, location, price positioning
• State-specific lead generation patterns identified
• Feature importance ranking for actionable ad improvement

---

🚀 Quick Start

git clone https://github.com/pelabdang/leads-analysis-prediction.git
cd leads-analysis-prediction
pip install -r requirements.txt

Batch Prediction

from src.models.model_trainer import ModelTrainer
import pandas as pd

listings_df = pd.read_csv('your_listings.csv')
trainer = ModelTrainer()
predictions = trainer.predict_batch(listings_df)

listings_df['predicted_leads'] = predictions
listings_df['performance_category'] = pd.cut(
    predictions,
    bins=[0, 5, 15, 30, float('inf')],
    labels=['Low', 'Moderate', 'High', 'Exceptional']
)

Real-time API

from src.models.model_trainer import ModelTrainer
from flask import Flask, request, jsonify

app = Flask(__name__)
model = ModelTrainer.load_model('complete_ml_pipeline')

@app.route('/predict', methods=['POST'])
def predict_leads():
    data = request.json
    prediction = model.predict_single(data)
    return jsonify({'predicted_leads': prediction})

---

📁 Project Structure

leads-analysis-prediction/
├── config/               # config.yaml — model & data settings
├── data/
│   ├── raw/              # Original immutable dataset
│   ├── processed/        # Cleaned, feature-engineered data
│   └── external/         # Supplementary sources
├── models/               # complete_ml_pipeline.joblib + artefacts
├── notebooks/
│   ├── exploratory/
│   ├── feature_engineering/
│   └── modeling/
├── reports/              # Modeling & feature engineering reports
├── src/
│   ├── data/
│   ├── features/
│   ├── models/
│   └── visualization/
├── requirements.txt
└── setup.py

---

📊 Key Lead Generation Drivers

1. 📞 Phone Engagement — strongest signal of buyer intent
2. 👁️ Visual Presentation — 8 photos = optimal lead generation
3. 🌍 Geographic Location — state/city-level variance is significant
4. 💰 Price Positioning — sweet spot relative to market value
5. 🚗 Vehicle Features — safety & comfort clusters drive engagement

---

📚 Reports

• 📊 Modeling Report
• ⚙️ Feature Engineering Report

---

📧 Contact: Angelo Pelisson · GitHub
📊 Performance: RMSE 6.922 · R² 70.3% · 64.6% over baseline