Inspiration

"Location, location, location" — we've heard this golden rule of real estate countless times. But as we started exploring the BroadVail dataset, a
question emerged: Does this rule still hold after COVID-19?

What We Learned

  • Before COVID: Market name was #1 predictor (12.27)
  • After COVID: Property type shot to #1 (8.92)
  • type_main gained +8.35, mrkt_name dropped -7.05
  • Healthcare access appeared as top predictor for the first time

The new rule: Quality, scale, and access.

How We Built It

  1. Split-Era Modeling (Pre/Post COVID)
  2. Stacking Ensemble (XGBoost + LightGBM + ExtraTrees → Ridge)
  3. Feature Engineering (property_age, supply_pressure, etc.)
  4. 15-minute drivetime as optimal radius
  5. Strict leakage prevention

Result: RMSE 0.114

Challenges We Faced

  1. The Structural Break — solved with split-era modeling
  2. Data Leakage Risk — fold-specific calculations
  3. Interpreting the Shift — connected to real-world changes
  4. Balancing Accuracy and Insight — chose interpretable models

Built With

Share this project:

Updates