Redefining 'Local' in the X-Minute City: Predicting apartment performance based on neighborhood amenities and understanding how renter preferences changed post-COVID.

Try the platform: https://revpar-analysis.streamlit.app/

This project predicts RevPAR (Revenue Per Available Unit) Growth for multifamily apartment properties using machine learning. We analyze how neighborhood amenities, location characteristics, and market dynamics influence apartment performance.

1. Time Period Effect: Pre vs Post COVID distinction accounts for ~67% of prediction importance
2. Location Matters: State and metro-level features are crucial predictors
3. Optimal Drivetime: 10-minute drivetime radius provides best predictive power
4. Model Performance: Ensemble model achieves 0.1092 RMSE (18.3% improvement over baseline)

# Navigate to project
cd "/Users/somildoshi/Somil Doshi/UH - Coding/Finance"

# Create virtual environment
python3 -m venv venv
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Install OpenMP for XGBoost (Mac only)
brew install libomp

# Create .env file
echo 'OPENAI_API_KEY=your_key_here' > .env
echo 'RANDOM_STATE=42' >> .env

# Train on 15-minute drivetime (default)
python train.py --drivetime drv15

# Or train on other drivetimes
python train.py --drivetime drv10
python train.py --drivetime drv30

python generate_full_submission.py

streamlit run app/app.py

Finance/
├── data/
│   ├── master_panel_drv10.csv    # 10-minute drivetime data
│   ├── master_panel_drv15.csv    # 15-minute drivetime data
│   ├── master_panel_drv30.csv    # 30-minute drivetime data
│   ├── scoring.csv               # Held-out test set
│   └── dictionary.csv            # Data dictionary
│
├── src/
│   ├── __init__.py               # Package init
│   ├── config.py                 # Configuration settings
│   ├── data_loader.py            # Data loading & wrangling
│   ├── features.py               # Feature engineering
│   └── model.py                  # Model training & evaluation
│
├── app/
│   └── app.py                    # Streamlit dashboard
│
├── models/                       # Saved model checkpoints
│   ├── baseline_ridge.pkl
│   ├── xgboost_model.pkl
│   ├── lightgbm_model.pkl
│   └── ensemble_model.pkl
│
├── outputs/                      # Predictions & analysis
│   ├── submission_final_*.csv
│   └── feature_importance_*.csv
│
├── train.py                      # Main training script
├── generate_full_submission.py   # Generate predictions
├── requirements.txt              # Dependencies
├── .env                          # Environment variables
└── README.md                     # This file

• Handled missing values (median imputation for numeric, 'Unknown' for categorical)
• Handled multi-phase buildings (same UBID with different yearbuilt)
• Outlier clipping at 1st/99th percentiles
• Created target variable based on time_window_tag

• Amenity Density: Count / trade area size
• Amenity Ratios: High-end grocery / total grocery, sit-down / total food
• Composite Scores: Proximity, transit, health, housing composites from AARP metrics
• Building Features: Age, renovation status, size indicators
• Supply Features: Growth rate, expansion indicators
• Affordability Features: Rent-to-mortgage ratio, rent advantage indicator

• GroupKFold with 5 folds
• Grouped by UBID to prevent data leakage
• Same property never appears in both train and validation

1. is_pre_covid (39.4%)
2. is_post_covid (28.0%)
3. time_window_tag (9.4%)
4. state (4.7%)
5. state_metro (4.3%)
6. zip_metro (1.4%)
7. aarp_met_house_access_step (1.1%)
8. aarp_met_trans_delay (0.7%)
9. aarp_met_health_hospital (0.6%)
10. mrkt_name (0.4%)

• Post-COVID, local amenities matter more than downtown proximity
• Properties with good grocery access outperform regardless of location type
• The 15-minute city concept should be adapted by neighborhood type

• For Renters: Look for "hidden gem" neighborhoods with high local amenity density
• For Urban Planners: Invest in local amenities in underserved suburban areas
• For Investors: Consider supply growth as a key risk factor

1. ✅ Code Repository: Complete with clear instructions
2. ✅ Methodology Summary: This README + notebook analysis
3. ✅ Predictions CSV: outputs/submission_final_*.csv
4. ✅ Dashboard: Interactive Streamlit app

• Abbiasov et al., "The 15-minute city quantified using human mobility data," Nature Human Behaviour 8, no. 3 (2024)
• BroadVail Capital Partners Finance Track Problem Statement

Rice Datathon 2026 Participant

This code is provided for the Rice Datathon 2026 competition only. Data must be deleted after the event.