House Price Predictor 🏠

An ensemble of XGBoost + LightGBM + sklearn GradientBoosting for predicting house prices in the King County (Seattle) area.

Model Details

• Dataset: inria-soda/tabular-benchmark (reg_num_house_sales config)
• Training samples: 15,137 | Validation: 3,234 | Test: 3,242
• Target: log(price) — exponentiate predictions with np.exp() for dollar amounts
• Based on: Grinsztajn et al. "Why do tree-based models still outperform DL on tabular data?" (NeurIPS 2022)

Results

🏆 Best Model: Ensemble (R² = 0.8915)

Feature Importance

Usage

import joblib
import numpy as np
from huggingface_hub import hf_hub_download

# Download and load model
model_path = hf_hub_download("richeechabhadiya/house-price-predictor", "xgboost_model.joblib")
model = joblib.load(model_path)

# Predict (input: 15 features as numpy array)
# Features: bedrooms, bathrooms, sqft_living, sqft_lot, grade, sqft_above,
# sqft_basement, yr_built, yr_renovated, lat, long,
# sqft_living15, sqft_lot15, date_month, date_day
sample = np.array([[3, 2.0, 1800, 7500, 7, 1800, 0, 1990, 0, 47.5, -122.2, 1700, 7500, 6, 15]])
log_price = model.predict(sample)
price_dollars = np.exp(log_price)
print(f"Predicted price: ${price_dollars[0]:,.0f}")

Ensemble Prediction (Best Accuracy)

import joblib
import numpy as np
from huggingface_hub import hf_hub_download

# Load all 3 models
xgb = joblib.load(hf_hub_download("richeechabhadiya/house-price-predictor", "xgboost_model.joblib"))
lgbm = joblib.load(hf_hub_download("richeechabhadiya/house-price-predictor", "lightgbm_model.joblib"))
skgb = joblib.load(hf_hub_download("richeechabhadiya/house-price-predictor", "sklearn_gb_model.joblib"))

# Ensemble prediction (average)
sample = np.array([[3, 2.0, 1800, 7500, 7, 1800, 0, 1990, 0, 47.5, -122.2, 1700, 7500, 6, 15]])
pred = (xgb.predict(sample) + lgbm.predict(sample) + skgb.predict(sample)) / 3
price = np.exp(pred)
print(f"Ensemble predicted price: ${price[0]:,.0f}")

Training

Trained with hyperparameters from NeurIPS 2022 benchmark research:

• XGBoost: 2000 estimators, lr=0.05, max_depth=6, early stopping (50 rounds) → stopped at 621 rounds
• LightGBM: 2000 estimators, lr=0.05, 63 leaves, early stopping (50 rounds) → stopped at 370 rounds
• sklearn GB: 500 estimators, lr=0.05, max_depth=6, early stopping (50 rounds)

Files

• xgboost_model.joblib — XGBoost model (2.4 MB)
• lightgbm_model.joblib — LightGBM model (2.1 MB)
• sklearn_gb_model.joblib — sklearn GradientBoosting model (1.9 MB)
• model_metadata.json — Full training metadata, results, and feature names
• feature_importance.json — Feature importance scores