elladridi/real-estate-price-prediction-tunisia

ML pipeline forecasting 3/5-year future real estate prices in Tunisia — CRISP-DM methodology, dual XGBoost models (villa/apartment), quantile-regression uncertainty intervals, and a Streamlit valuation demo. # Predicting Future Real Estate Prices in Tunisia ML pipeline forecasting 3/5-year future real estate prices in Tunisia — CRISP-DM methodology, dual XGBoost models (villa/apartment), quantile-regression uncertainty intervals, and a Streamlit valuation demo. Six-week, internship project. Built with the **CRISP-DM** methodology: 7 raw sources merged into a 42,161-row dataset, cleaned down to 12,971 usable listings, feature-engineered, and modeled with two specialized XGBoost regressors plus a quantile-regression uncertainty layer. ## Results | Segment | R² (test set) | Test set size | |------------|---------------|----------------| | Villa | **0.806** | 653 listings | | Apartment | **0.570** | 1,942 listings | - 95% prediction intervals cover the true price **93.9%** of the time on unseen data. - Median interval width ≈ 534,191 TND (wide, reflecting the underlying R²). ## Architecture ``` 7 raw sources ──► merge into shared schema (42,161 rows) │ ▼ clean, impute, deduplicate (12,971 rows) │ ▼ feature engineering: postal code, region, tourism-zone flag, controlled price-trend index (2018 base year) │ ▼ ┌────────────────────┴────────────────────┐ ▼ ▼ Villa XGBoost model Apartment XGBoost model (no postal code/region — (+ postal code, region — too little data to support it) enough data to benefit) │ │ └────────────────────┬────────────────────┘ ▼ quantile-regression models (2.5th/97.5th pct) │ ▼ forecast_price(property_features, years_ahead) │ ▼ Streamlit test interface ``` **Why two separate models instead of one?** Adding richer location features (postal code, region) to a single combined model helped apartments slightly (R² 0.536 → 0.550) but hurt villas badly (R² 0.734 → 0.641) — villa listings are too sparse for fine-grained location categories, so the model overfit. Splitting into two models, each with a feature set tai …