House Price Prediction using XGBoost

A home is the largest purchase most families ever make, yet its price is still usually set by slow, expensive and partly subjective appraisals — while the online estimators buyers actually use are black boxes that never explain a number. The underlying problem is a regression over messy, mixed data: dozens of attributes (living area, overall quality, garage capacity, basement finish, neighborhood, year built) interact non-linearly, some values are missing, and sale prices are heavily right-skewed. This project applies it to the Ames Housing dataset (2,930 homes, 79 explanatory variables, a standard regression benchmark): missing-value handling per feature type, ordinal encoding of quality ratings, a log transform of the skewed SalePrice target, XGBoost training with early stopping and cross-validation, holdout evaluation with RMSE, MAE and R-squared, SHAP-based feature analysis, and a demo that estimates any home's price from sliders and prices the effect of renovations.

1. The Ames Housing dataset (2,930 homes, 79 explanatory variables) is loaded; missing values are imputed per feature type, quality ratings ordinal-encoded, and SalePrice log-transformed to tame its right skew.
2. The data is split into train and holdout sets (e.g. 80/20); the target used for training is log(SalePrice), converted back to dollars at inference.
3. An XGBoost regressor is trained with early stopping on a validation fold; hyperparameters (max depth, learning rate, regularization) are tuned with cross-validation.
4. The holdout set is evaluated with RMSE, MAE and R-squared, plus predicted-vs-actual and residual plots that reveal where the model under- or over-predicts.
5. SHAP values and gain importance identify the strongest price drivers — overall quality, living area, garage capacity — for the report's analysis section.
6. At inference, the demo encodes a home's attributes with the same pipeline, predicts log-price, exponentiates it, and breaks the estimate down by feature contribution.

Tech stack:

• Python 3.10, XGBoost (gradient-boosted regression, early stopping)
• scikit-learn (pipelines, cross-validation, metrics)
• pandas, NumPy (data wrangling and feature tables)
• Matplotlib, Seaborn (predicted-vs-actual, residuals, importance plots)
• SHAP (feature contribution explanations)
• Jupyter notebook (buyer-run training and evaluation)
• Streamlit demo app (price estimator, renovation simulator, dashboards)

• [Ames Housing data pipeline] Loads the Ames dataset (2,930 homes, 79 features), imputes missing values per feature type, ordinal-encodes quality ratings (Excellent to Poor), and log-transforms the skewed SalePrice target.
• [XGBoost regressor with early stopping] Trains gradient-boosted trees with early stopping on a validation fold and cross-validation, with max depth, learning rate and regularization exposed as tunable hyperparameters.
• [SHAP and gain-based importance] Ranks price drivers — overall quality, living area and garage capacity typically top the list — with gain importance bars and SHAP summary plots the report can analyze.
• [Interactive price estimator] Move sliders for overall quality, living area, garage cars, basement area, year built and lot area and get an instant price estimate with a per-feature contribution breakdown.
• [Renovation simulator] Upgrade the kitchen or overall quality rating in the demo and see the predicted price delta — a concrete "what is the upgrade worth" analysis for the report.
• [Evaluation dashboard] Predicted-vs-actual scatter, residual plots, RMSE, MAE and R-squared on the held-out set, all computed by the notebook during the build.
• [Neighborhood price explorer] Median sale prices and price distributions across Ames neighborhoods, showing how location shifts the baseline.
• [Comparable-homes finder] Finds the nearest neighbors of an input home in feature space and shows their actual sale prices beside the prediction.

• Complete source code (data pipeline, training, evaluation, demo app)
• Jupyter training and evaluation notebook (buyer-run procedure on the Ames dataset)
• Project report PDF (background, methodology, SHAP analysis, evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (gradient boosting, log transform, RMSE vs MAE, SHAP, overfitting control)
• Setup guide (environment, dataset download, running the notebook and demo)

• The Ames data covers one Iowa city from 2006–2010; it does not represent current prices, other cities, or today's interest-rate environment — this is a teaching benchmark, not a valuation service.
• The model interpolates within the training distribution; unusual luxury homes or extreme feature values extrapolate poorly, as the residual plots show.
• Macroeconomic drivers (interest rates, local employment, school ratings beyond the dataset) are outside the feature set.
• The renovation simulator estimates price effects of upgrades, not their construction cost — return-on-investment needs cost data the project does not include.
• Currency and units are US dollars and square feet throughout, matching the dataset.