Used Car Price Prediction using Gradient Boosting

India's used-car market runs on asymmetry: sellers anchor high, buyers guess low, and neither side knows the fair price of a specific car with its specific mileage and history. Listing platforms show asking prices, not transaction prices, so the "market rate" is folklore. This project replaces folklore with a model: gradient-boosted trees trained on ~9,200 listings learn how age, mileage, fuel, transmission, brand and ownership depress or support prices, and return a fair-price estimate with a ±10% expected band. The estimate is explainable — the demo breaks down how much each factor contributed — which is what makes it usable in a real negotiation. The build also documents the honest limits: the model prices the car's attributes, not its accident history or mechanical condition.

1. ~9,200 listings are cleaned: duplicates removed, outliers trimmed, km/year/fuel/owner standardized, prices in ₹ lakh.
2. Features are engineered: vehicle age, log-mileage, brand, segment, fuel, transmission, owner count, city tier.
3. A gradient-boosting regressor is trained with cross-validation; R² and MAPE are reported on held-out listings.
4. Feature importances and partial-dependence plots reveal the depreciation structure for the report.
5. In the demo, entered specs flow through a depreciation model fitted to the listing statistics, returning price, band and breakdown.
6. Any listing's specs can be scored the same way to produce the fair/overpriced/suspicious verdict.

Tech stack:

• Python 3, LightGBM/XGBoost, scikit-learn
• pandas, NumPy
• Matplotlib (analysis plots)
• HTML/CSS/JavaScript estimator demo
• Indian used-car listings dataset (Kaggle)

Dataset & model details

• Dataset: Public Indian used-car listings dataset (Kaggle) — ~9,200 listings with brand, model, year, km driven, fuel, transmission, owner number, city, mileage, engine, seats and asking price; cleaned and deduplicated.
• Task: Regression; input = 12 listing attributes, output = fair price (₹ lakh).
• Model: Gradient boosting (LightGBM-style) with native categorical handling; tuned leaves, learning rate and regularization.
• Metrics: Test R² 0.93, MAPE ±8% (design targets for the built-to-order training run); residual analysis in the report. No measured figures are claimed before the run.

• [Fair-price estimator] Enter brand, segment, year, km, fuel, transmission and owner count; get a price in ₹ lakh with an expected range.
• [Estimate breakdown] Every estimate shows each factor's contribution (age depreciation, mileage penalty, fuel/transmission adjustments).
• [Depreciation curves] Segment-wise value-retention curves (indexed to new price) learned from the listings data.
• [Listing verdicts] Sample listings scored against the model: fair, overpriced or suspiciously cheap, with the buyer's playbook for each.
• [Price-vs-age scatter] The model's expected band plotted against real listings — outliers are visible instantly.
• [Gradient-boosted regressor] LightGBM-style training with categorical handling for brand/model/fuel.
• [Negotiation guide] The report translates model outputs into buyer/seller actions (when to negotiate, when to walk away, when to verify).

• Listings cleaning & feature-engineering scripts
• Gradient-boosting training & evaluation notebook
• Trained model with inference script
• Price-estimator web demo with breakdown
• Depreciation and market-trend analysis plots
• Project report PDF (background, market context, methodology, results, negotiation guide)
• PPT presentation for final review
• Viva Q&A preparation document (boosting, regression metrics, categorical handling, MAPE)

• R² and MAPE are design targets; the report records the actual metrics from the training run for the order.
• The model prices attributes, not condition — accident history, service records and mechanical state need physical inspection.
• Trained on asking prices (not final transaction prices), which skew slightly high; the report discusses this bias openly.
• Indian market only as shipped; other markets need local listings and retraining.
• The demo's depreciation model is illustrative; the shipped notebook trains the full gradient-boosted model on the real data.