Fintech Development

Fraud Signal Feature Extraction Prompt

Extract fraud-signal features from raw transaction data for a scoring model, covering velocity, device, and behavioral patterns.

Anselm Fowel / Public collection28 Sept 2026

Your prompt

The prompt

You are a fintech engineer who has built fraud detection feature pipelines feeding a real-time scoring model, and you know that the best fraud signals almost never live in the transaction record alone, they come from patterns across time and entities.

My raw transaction data includes [FIELDS AVAILABLE: AMOUNT, TIMESTAMP, CARD/ACCOUNT ID, MERCHANT, DEVICE FINGERPRINT, IP, GEOLOCATION, LIST WHAT YOU HAVE]. This feeds a [FRAUD MODEL TYPE: RULES ENGINE / ML CLASSIFIER] that scores transactions in [REAL-TIME / NEAR-REAL-TIME] for [FRAUD TYPE: CARD FRAUD / ACCOUNT TAKEOVER / SYNTHETIC IDENTITY].

First, list velocity-based features derivable from this data: transaction count and total amount per card/account over rolling windows (1 minute, 1 hour, 24 hours), and distinct merchant or geography count in those same windows. For each, explain what fraud pattern it's designed to catch.

Next, list device and network features: whether this device/IP has been seen on this account before, how many distinct accounts have used this same device or IP recently, and mismatch signals like IP geolocation versus billing address. Explain the false-positive risk each carries (e.g., shared NAT IPs, VPNs used legitimately) and how to dampen it.

Then list behavioral and contextual features: deviation from this account's typical transaction amount or merchant category, time-of-day anomalies relative to the account's history, and a first-time-large-transaction flag for new accounts specifically, since new-account fraud has different base rates than established-account fraud. Also address which of these features require a low-latency online store for real-time scoring versus which can be computed in a batch pipeline and cached, and the staleness tolerance for each.

Finally, output a feature table with columns: Feature Name, Description, Data Sources Required, Computation Window, and Latency Tier (real-time store vs. batch), covering at least 15 features, followed by pseudocode for computing the 3 highest-signal velocity features.

2,038 characters

Your prompt

Recent articles