Hazard Count Prediction for Liberty Mutual’s Properties

Kaggle Ranking: Top 12% (248/2,362 competitors)

End Date: August 28, 2015

Description: Liberty Mutual needs to decide which homes are worth insuring, and this competition frames that underwriting decision as a supervised learning problem: predict the hazard count for a property from anonymized attributes such as its roof, foundation, windows, and siding. Getting these predictions right lets the insurer align its portfolio of home policies with its business goals.

Best Method: Weighted Ensemble (RF + XGBoost)

Reflections: This competition was particularly challenging because the anonymized features left little room for feature engineering. It was a fun exercise in ensembling and performing proper cross-validation to prevent overfitting on the public leaderboard, as tempting as that may be.