How Horse Racing Predictions with H2O.ai Saved a Local Insurance Company $8M a Year
- Predictive AI
- H2O Driverless AI
- H2O Hydrogen Torch
Overview
Matthew Foster, chief data scientist at Simply AI, traces his path from a failed rules-based system to predicting the winner of the Melbourne Cup with an H2O AutoML model built on roughly 1,000 features per horse. He then shows how that same approach carried into enterprise work, describing an insurance company's Snowflake-integrated claims prediction model built with H2O, plus a Hydrogen Torch NLP project that labels and searches tens of thousands of unstructured policy documents through an H2O Wave app.
Key takeaways
- How the horse racing model used H2O AutoML (built years before Driverless AI existed) with around 1,000 features per horse, including trainer accuracy, jockey history, and even the racing form of a horse's parents.
- How, for an insurance company client, preprocessing and scoring happen inside Snowflake: a Driverless AI model (MOJO) is uploaded and scored there via a UDF, with results piped to a postprocessing step used for monitoring and dashboards.
- Why the presenter calls the claims-prediction model's savings, an estimate of around $2 million per year, conservative, noting the real number could be higher.
- The team's three operating principles: simplify (they removed about 40 percent of the preprocessing code with no loss in accuracy), automate (auto-generating and scheduling test cases across environments), and monitor (watching daily for model and data drift).
- How a separate NLP project used H2O Hydrogen Torch, trained on a small set of manually labeled policy document paragraphs, to then label and score around 50,000 unstructured policy documents for specific risk and trauma definitions.
- How the labeling and search tool was packaged as an H2O Wave app so business users could upload a PDF, get labeled paragraphs back in under a minute, and run full-text search across the document set.
Start owning your AI