Open Doorway
Open Doorway Model Card
← Back to Transparency Report

Open Doorway Price Appreciation Model

Version: appreciation_model_v3 | Last trained: 2026-08-25 | Status: Production

model_training Model Details

Algorithm

LightGBM Gradient Boosting

Task

12-month price appreciation regression

Training Data

20,390 properties across 24 markets

Features

Property, location, market, economic, image, satellite

Feature Store

Feast with Redis online store

Experiment Tracking

MLflow

speed Performance Metrics

R²

0.91

MAE

4.2%

RMSE

6.8%

MAPE

5.1%

ECE (calibration)

0.038

Brier Score

0.14

p95 Latency

48ms

Fairness ΔFPR

0.07

verified Confidence Tiers

HIGH Confidence ≥ 80%. Used for primary valuations and investment recommendations.
MEDIUM Confidence 60-79%. Suitable for market analysis and preliminary screening.
LOW Confidence < 60%. Supplement with manual due diligence. Limited feature availability.

bar_chart Feature Importance (Top 10)

market_velocity
0.92
price_to_income
0.87
historical_appr
0.83
infra_score
0.76
demographic_growth
0.71
mortgage_rate
0.68
climate_risk
0.63
job_growth
0.59
sqft_quality
0.52
school_rating
0.47

warning Known Limitations

  • • Model trained primarily on Latin American markets (60% Mexico). Accuracy may vary in underrepresented regions.
  • • Forecasts assume stable macroeconomic conditions. Extreme events (recession, policy changes) are not modeled.
  • • Limited to properties with sufficient comparable data. New-build or unique properties may have lower confidence.
  • • 12-month horizon is primary target. 36-60 month forecasts have wider confidence intervals.
  • • Satellite and image features are experimental and may not be available for all properties.

balance Ethical Considerations

  • • Fairness audits performed quarterly across regional subgroups. Worst-group ΔFPR: 0.07 (target < 0.1).
  • • No protected attributes (race, gender, religion) are used as model features.
  • • Confidence tiers ensure low-confidence predictions are flagged for human review.
  • • All forecasts are marked as "analytical estimates, not investment advice" in user-facing text.
  • • Data privacy: GDPR compliant. No personal data used in training without consent.

This model card follows the Model Card framework for transparent ML reporting. Updated: 2026-09-13.