Winner of the $5,000 prize for internal Liberty Mutual Group employees Kaggle competition: “Liberty Mutual Group - Fire Peril Loss Cost”. This competition was held in parallel (same leaderboard) to the public Kaggle competition.
1st /21 LMG teams and 36th/634 teams overall position – top 10%.
The modeling problem: “In commercial insurance, fire losses account for a significant portion of the total property losses. Modeling is inherently difficult, as losses are high severity and low frequency and volatile. Your task is to predict the target using the provided information.”
The competition took place during the summer of 2014. At that moment, I was fascinated with Random Forests in R, so I started the competition using R as many others participants. After a while, I realized that the hierarchical structure of the data was really important. At that moment, I couldn’t find any R packages with Random Forests for mixed data so, I used SAS Generalized Mixed Models in order to take advantage of that critical data structure. That was my successful approach using only one model. In 2014 I wasn’t and expert using model stacking that are critical in this kind of competitions.
This is the presentation: