JD Case Study · Acima
Acima — Upbound Group’s lease-to-own engine — is hiring a data scientist to build the models that decide who gets funded. That’s the exact funnel I worked inside Acima: bank-linked, income-verified, lease-to-own underwriting where synthetic identities and first-payment fraud are the losses that actually matter. This isn’t a domain I’d ramp into. It’s the one I lived in.
One kind of number here. The analysis below runs on illustrative figures invented to show method. No Acima data is used or implied; the operating logic is the deliverable.
What they’re actually buying
Catch more fraud without turning away good customers. Stripped down, the job is one number, tuned continuously: every dollar of fraud loss against every good lease you decline — or every good customer you make jump through a hoop. A rules engine alone is too blunt for the gray zone; a black-box model alone won’t survive a fraud review or a chargeback dispute. What Acima is buying is the person who builds the hybrid — rules for the explainable hard blocks, ML for the gray zone — and owns where the operating point sits.
The first 90 days
The order matters. Nobody should change a system they haven’t reconciled, and nobody should present a verdict before the numbers under it can be trusted.
Days 1–30
Decompose fraud losses by typology — synthetic identity, account takeover, first-payment fraud — by channel and by vintage. Reproduce the fraud-review funnel and find where the incumbent rules (and any live model) actually sit on the precision/recall curve. Meet Fraud Prevention, Ops, and Analytics.
Days 31–60
Engineer the fraud-specific features — velocity checks, device/IP signals, behavioral profiles — and train an XGBoost challenger benchmarked against the rules at a fixed friction budget. Design the hybrid strategy with the Fraud Prevention Manager: rules hold the hard blocks, the model works the gray zone.
Days 61–90
Run champion/challenger in shadow with an explicit loss guardrail, wire the operating-point dashboard (fraud loss vs. approval/friction, by segment), and stand up monitoring + a retraining cadence so the model doesn’t decay as fraud adapts. Feed the same features into the chargeback-dispute process.
Signature analysis · illustrative data
Where the model earns its keep: the drop from the rules layer (64) to the funded book (56) is the gray zone — accounts a blunt rule can’t call. A tuned score blocks the fraud a rule waves through and recovers good customers a rule would decline. That’s the operating point I’d own: nudge it toward fewer losses while holding approvals flat. The whole game is buying loss reduction without friction.
Requirement → proof
Each line of the posting against something already built and running on this site. Where the fit is a ramp rather than a match, the row says so.
The same engine
The engine behind this whole portfolio — the approval/funding funnel, the first-payment gate, loss economics, and experimentation with a loss guardrail — is the same machine a fraud book runs on. I built it in public for lease-to-own, which is Acima’s own book; the fraud typologies change faster than the frontier does. I’d walk in fluent in the one trade-off this role exists to manage: loss against friction.
Acima’s fraud funnel isn’t a domain I’d ramp into — it’s the book I worked, and the one trade-off I’d own from day one. On-site in Draper is a feature, not a friction: I’m already in Utah. I’d rather show you the model than describe it — let’s talk about your loss curve. — Paul Brown