Rice Datathon 2026 - BroadVail Finance Track

RevPAR Growth Prediction & Neighborhood Drivers

Predict which multifamily properties outperform their peers on RevPAR growth and discover which combinations of amenities, neighborhood context, and travel-time definitions best explain that outperformance across different locations and time periods.

Results Summary

Metric	Value
Best Model	CatBoost Ensemble (3 models)
CV RMSE	0.1368
Original CatBoost	0.1725
Baseline RMSE	0.2020 (Ridge)
Improvement vs Baseline	32.3%

Key Findings:

• the submission file is named as submission.csv in the root folder
• Property type importance increased 226% post-COVID
• Building age metrics decreased ~58% in importance post-COVID
• Only ~20% of properties maintained same performance quartile between periods

Quick Start

Prerequisites

• Docker and Docker Compose installed
• Dataset files (provided separately)

Setup

1. Place dataset files in Dataset/

   Dataset/
   ├── master_panel_drv10.csv
   ├── master_panel_drv15.csv
   ├── master_panel_drv30.csv
   ├── scoring.csv
   └── dictionary.csv
   

2. Build and start Docker environment

   docker-compose build
   docker-compose up

3. Access Jupyter Lab at http://localhost:8888

Run Notebooks

Execute notebooks in order (00-07) to reproduce the analysis.

Project Structure

├── submission.csv         # Final predictions on scoring.csv
├── data/                  # Processed data (gitignored)
├── Dataset/               # Raw CSV files (gitignored)
├── models/                # Trained models (gitignored)
├── notebooks/             # Analysis notebooks (00-07)
├── reports/
│   ├── figures/           # Visualizations
│   ├── tables/            # Analysis tables
│   ├── final_writeup.md   # Methodology summary
│   └── llm_outputs/       # LLM-generated summaries
├── src/                   # Python modules
├── Dockerfile
└── docker-compose.yml

Notebook Pipeline

#	Notebook	Description
00	00_data_audit.ipynb	Data schema validation
01	01_research_insights.ipynb	EDA and hypotheses
02	02_feature_engineering.ipynb	Feature pipeline
03	03_baselines.ipynb	Baseline models
04	04_advanced_ml.ipynb	Gradient boosting
05	05_interpretation_story.ipynb	SHAP analysis
06	06_make_submission.ipynb	Generate submission
07	07_writeup.ipynb	Final report

Methodology

Feature Engineering

• Density features: amenity counts / trade area size
• Share features: category proportions
• PCA: dimensionality reduction on amenity densities
• Clustering: neighborhood typology via KMeans

Model

• Algorithm: CatBoost gradient boosting
• Validation: 5-fold GroupKFold by market
• Interpretation: SHAP values for feature importance

LLM Integration

Generate executive summary using OpenAI API:

export OPENAI_API_KEY="your-key"
python scripts/run_llm_summary.py

Uses temperature=0 for deterministic outputs per competition rules.

Key Dependencies

• pandas, numpy, scikit-learn
• catboost, lightgbm, xgboost
• shap, matplotlib, seaborn
• openai (for LLM summaries)

Competition Rules

1. Models must be repeatable and deterministic
2. Data is confidential - delete after competition
3. Only use provided data - no external sources

License

Competition use only. See BroadVail data use notice.