Summary
A real estate client wanted a model that predicts a home’s sale price from its listing details. I cleaned 20,000 King County (Seattle area) sales, engineered features, and trained gradient-boosted trees with XGBoost. The model was then scored on a hidden holdout set as a class competition.
| Model version | Test RMSE | Test R² |
|---|---|---|
| Engineered features + geographic clusters | $136,744 | 0.88 |
| Cleaned outliers + date features (final) | $113,938 | 0.90 |
Approach
1. Cleaning. I parsed the sale date into year and month, dropped ID columns, and removed implausible records: more than 15 bedrooms, bathroom counts outside 0.5–5.5, and extreme living and lot sizes.
2. Feature engineering. I used log transforms of square footage, living-to-lot and above-to-basement ratios, house age, years since renovation, and a renovation flag. K-means clusters on latitude and longitude capture neighborhood effects.
3. Model. I used an XGBoost regressor with an 80/20 train/test split, then ran the same cleaning function on the holdout set so the training and scoring data stay consistent.
def clean_housing(housing):
housing['datetime'] = housing['date'].str.split('T').str[0]
housing = housing.assign(
year=housing['datetime'].str[0:4].astype(int),
month=housing['datetime'].str[4:6].astype(int),
).drop(columns=['date', 'datetime', 'id'])
clean = housing[housing['bedrooms'] <= 15]
clean = clean[(clean['bathrooms'] <= 5.5) & (clean['bathrooms'] >= .5)]
clean = clean[clean['sqft_living'] <= 8670]
clean = clean[clean['sqft_lot'] <= 500000]
return clean
model = XGBRegressor()
model.fit(X_train, y_train)
predictions = model.predict(X_test)What I learned
Removing a small number of extreme listings improved the model more than adding extra engineered features did, cutting RMSE by about $23,000. Bundling the cleaning steps into a single function also made it simple to score new holdout files without mismatched columns.