Applied Scientist, Risk

About the role

Fraud models at Vantage make decisions in under thirty milliseconds on transactions we cannot get wrong in either direction. A false negative is money; a false positive is a customer who leaves. The threshold conversation here is a business conversation.

What you will do

Own fraud detection models from feature to deployment
Set and defend operating thresholds against real cost asymmetries
Work under class imbalance of roughly one in four thousand
Explain model decisions to compliance and to regulators

What we look for

Strong statistics — this is a genuine requirement, not a preference
Production experience with severe class imbalance
Gradient boosting depth; most of our production models are boosted trees
Able to explain a model to a non-technical audience

What the assistant changed, and why

Change

Summary rebuilt around imbalanced classification and threshold selection

Why

Their entire posting is about cost-sensitive decisions. Yours led with deep learning, which is not what this team runs.

Change

Foregrounded the XGBoost churn model over the neural ranking work

Why

They say most production models are boosted trees. Your strongest boosting project was third on the page.

Change

Added the cost matrix framing to that project

Why

You chose the threshold against a real cost asymmetry and described it as "tuned the threshold", which gives away the most relevant thing you did.

Change

Added your statistics coursework back in

Why

You had removed it as too junior. For a posting that calls statistics a genuine requirement, it belongs.

What matched

Imbalanced classification at one in two thousand, in production
Threshold selection against an explicit cost matrix
XGBoost and LightGBM depth
Presented model behavior to non-technical stakeholders

What it would not stretch

No financial services or regulatory experience. Stated plainly rather than stretched.

Awaiting your approval

The draft is written and the form is filled in on Vantage Financial’s own site. It sits here until you approve this one application. There is no approve-all, and nothing is submitted without you.