Editing AI for Real Estate (section)

== <span style="color: #FFFFFF;">Understanding</span> ==
Real estate ML has three core applications:

**Property valuation (AVMs)**: The foundational real estate AI problem. A property's value depends on thousands of features: physical (bedrooms, bathrooms, square footage, age, condition), locational (neighborhood, walkability, school districts, proximity to transit), and temporal (market conditions, interest rates, seasonality). Gradient boosting models (XGBoost, LightGBM) on structured features plus neural networks for photo features achieve median errors of 2–5%. The challenge: "location, location, location" — geo-spatial features are complex, hierarchical, and require careful encoding.

**Market forecasting**: Predicting where prices will go uses time-series ML on macro indicators (interest rates, employment, inventory), local market metrics (days on market, list-to-sale ratio), and leading indicators (building permits, mortgage applications). LSTM and Temporal Fusion Transformers capture complex temporal patterns across multiple spatial scales.

**Computer vision for properties**: Listing photos contain rich information about condition and desirability — not captured in structured data. CNNs classify room types, detect renovation quality, and score aesthetic appeal. Zillow's AI was trained on millions of agent-labelled photos to assess kitchen and bathroom quality. These vision scores improve AVM accuracy significantly.

**The iBuyer lesson**: Opendoor and Zillow Offers demonstrated both the power and risk of ML-based real estate. Zillow Offers famously lost $381M in Q3 2021 after its AVM failed to predict market turning points, causing massive overpaying for homes. This highlights that AVM errors are not independent — systematic biases across a portfolio are correlated, creating massive risk.
</div>

<div style="background-color: #8B0000; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">