321% lift in recovered fraud losses
A loss-ratio model that recovers more fraud losses than a top-five bank’s benchmark.
Context
Unrecovered fraud shows up directly in the loss ratio. The benchmark to beat was a model from a top-five bank.
What I did
I built and shipped a gradient-boosting plus logistic-regression model, with batch scoring in a Python and T-SQL pipeline.
Outcome
321% lift in recovered fraud losses versus the top-five bank benchmark.
Evidence
Client work, no public page.
Stack
Python, T-SQL, SQL Server, gradient boosting, logistic regression