AAIS and Pinpoint submitted a Florida filing outlining a homeowners predictive modeling framework that uses machine learning, third party consumer data, and census block group level information to score insurance risk.
The filing describes models designed to predict claim frequency, severity, and loss costs using Gradient Boosting Machines built on CatBoost. According to the memorandum, the models draw from data sources covering roughly 260 million U.S. adults and use information matched through names, addresses, email addresses, and phone numbers.
Insurers would need to subscribe through a partnership with Pinpoint to use the models, which are intended for homeowners and auto insurance pricing. Participating carriers can either use an industrywide model or train the framework on their own loss experience.
The filing also addresses regulatory concerns around discrimination and proxy variables. Pinpoint said the models underwent third party review by ZwillGen, which tested for disparate impact across race, gender, and age categories. The certification stated the model’s adverse impact ratios remained within the “four fifths rule” threshold commonly referenced in discrimination testing.