-
-
comparison between drivetime definitions
-
-
-
-
-
-
-
Inspiration & Journey
We were drawn to this project by the opportunity to work with high-quality, structured real estate data on a problem with clear business impact. Predicting apartment property RevPAR growth directly informs investment decisions, and the dataset offered a chance to extract insights from complex, well defined features.
What it does
It predicts REVPAR growth based on a variety of features like property characteristics, trade area metrics, and AARP Livability Index scores
How we built it
We built our solution using a Gradient Boosting Regressor with Optuna hyperparameter optimization, training separate models for each drivetime definition (10, 15, and 30 minutes) to respect the requirement of not merging dataframes. Through 50 trials per drivetime, we optimized seven key hyperparameters, achieving an RMSE of 0.1246 (0.57× the target standard deviation) for our best model.
Challenges we ran into
We faced several challenges: persistent CatBoost installation issues on macOS forced us to pivot to scikit-learn's GradientBoostingRegressor; determining whether to include time_window_tag as a feature required careful consideration to avoid data leakage; and the hyperparameter tuning process required 6-9 hours of computation time. We also had to navigate the trade-off between feature engineering and overfitting, ultimately deciding against creating aggregated features. These challenges reinforced that successful machine learning projects require more than algorithms, and they demand careful data understanding, thoughtful methodology, and domain knowledge to extract genuine insights from complex real-world problems.
Accomplishments that we're proud of
Working with people we met for the first time. And getting a functional model despite all the challenges
What we learned
We learned that housing economics metrics, particularly the relationship between mortgage costs and rental rates, are the strongest predictors of RevPAR growth, more so than traditional property characteristics. Using SHAP values for interpretability revealed actionable insights, such as property size becoming more important post-COVID, while location became less critical.
What's next for Brodvail Prediction Challenge
Explore different models and feature engineering ideas
Log in or sign up for Devpost to join the conversation.