Turning 14 billion miles of driving data into insight at Progressive

Progressive collects an enormous stream of behavioral data through its Snapshot usage-based insurance program, more than 14 billion miles and roughly 2 billion trips, closing in on 2 trillion records, with new sources arriving daily. The business runs on segmentation, and functions from pricing to claims depend on turning that data into predictions quickly. Basic models such as GLMs could not keep pace with the volume or the range of questions the analytics team wanted to ask.

Progressive built a predictive-modeling stack that pairs H2O with SAS, on top of DB2, SQL Server, and Hadoop, with Tableau and D3 for visualization. H2O let the team move beyond GLMs to more sophisticated machine learning and build many models in far less time. Snapshot devices capture speed, time of day, and hard-braking events, and more recently GPS, which the models turn into individualized risk scores that raise or lower a driver's rate.

The same fast-modeling approach extends across the organization. Customer churn and retention, fraud mitigation, claims management, and marketing all draw on the shared data foundation and H2O's algorithms.

Pairing H2O with the existing stack lets Progressive build models quickly enough to fold in new driving data as it arrives, rather than preparing models in advance. Snapshot analyzes more than 14 billion miles of driving data and has returned over $500 million in discounts to customers, who save 10 to 20 percent on average. Progressive can now apply advanced algorithms across telematics, churn, fraud, and claims from one data foundation.