Plaid — Fraud (Plaid Protect) — Product Sense & Metrics

Role context: Data Scientist, Fraud Data team (Plaid Protect) · Est. study time: 50 min · 5 practice questions

How to prepare for this role

Day to day you measure how Plaid Protect performs for each kind of customer and each moment in the user lifecycle, build the dashboards the fraud org shares, run backtests that compare models and rules (and that help Sales show prospects what Protect would have caught), design the tables all of this sits on, and design experiments for new customer-facing features. Expect the interview to test the same things: precision and recall at an operating point, base rates, how a backtest can fool you, metric design for a B2B risk product, experiments when outcomes are rare and arrive weeks later, and SQL on score, decision and label tables.

Where to spend your prep time

• Precision, recall and false positive rate at a fixed flag rate, and how the base rate changes precision.
• Backtest design: out-of-time windows, matured labels, and the leakage traps.
• The friction side: what stepping up a good user costs the customer.
• Experiments with rare, delayed outcomes, and with fraud rings that span both arms.
• Explaining a model improvement to a customer in one plain sentence.

What Is Plaid Protect

Plaid is the network that connects apps to people's bank accounts. It links to about 12,000 financial institutions and thousands of apps, and more than half of banked people in the US have used it. Businesses pay Plaid for products like account linking, identity, balance checks, payments, and fraud tools.

Plaid Protect is the fraud product. Launched in 2025, it scores a user in real time at each moment that matters in a customer's app: signup, identity verification, linking a bank account, login, password reset, moving money. The model behind it is called the Trust Index. It returns a risk score plus the signals behind it, and the customer decides what to do: let the user through, add a step (a "step-up", like extra verification), send them to manual review, or block them. The model has gone through three generations in about a year, each adding data and features. The latest one follows a fraud graph of devices, accounts, identities and apps up to nine hops away.

Four facts make fraud analytics here different from most fraud jobs:

1. The edge is the network. A fraudster can look clean inside one app and obvious across the network: the same device behind five identities, a bank account linked and unlinked across a dozen apps in a week. No single customer can see that. Plaid can.
2. The customer sets the threshold. Plaid supplies the score; the customer chooses how many users to step up or block. The same model is a different product at a 2% step-up rate than at 10%. So performance only means something at an operating point.
3. The labels belong to customers. Whether a user was fraud comes from the customer, often weeks later (chargebacks, returns, charge-offs, a click in the dashboard). Labels arrive late, some never arrive, and every customer defines fraud a little differently.
4. The other side fights back. Fraudsters change tactics after they get blocked. Every model starts decaying the day it ships.

The product is early-scaling: a year old, landing customers through backtests and still building the shared way everyone measures it. The data scientist doesn't just report performance. They define how performance is measured.