Business Context

HomeValue Analytics builds automated valuation models for regional real-estate brokers. The pricing team wants a regression model that generalizes well to new listings and can explain why regularization improves performance over an unregularized baseline.

Dataset

You are given a tabular dataset of 62,000 residential property sales from the last 24 months across 8 metro areas. The goal is to predict the final sale price from listing, property, and neighborhood features.

• Several numerical features are highly correlated (for example, square_feet and bedrooms; listing_price and price_per_sqft_listed).
• Categorical features include moderate-cardinality fields such as zip_code and seller_type.
• About 8% of rows have missing values in at least one feature, mostly in renovation and neighborhood fields.

Success Criteria

A good solution should outperform an unregularized linear regression baseline and demonstrate, with validation results, why regularization is useful for reducing overfitting. Target test RMSE below $42K and stable train/validation error gaps.

Constraints

• The pricing team prefers a model that remains reasonably interpretable.
• Batch scoring runs nightly on up to 200K listings, so inference must be fast.
• The solution should be easy to retrain weekly.

Deliverables

1. Train and compare Linear Regression, Ridge, and Lasso models.
2. Explain what regularization is and why it helps on this dataset.
3. Build a preprocessing pipeline for missing values, scaling, and categorical encoding.
4. Use cross-validation to tune regularization strength.
5. Report RMSE, MAE, and R² on validation/test data, and discuss coefficient behavior under regularization.