Fraud and financial crime
Fraud detection features that work: a playbook for banks, telcos and credit unions
The best fraud models rely on a small set of feature families: velocity, deviation from the customer’s own baseline, newness, network links and context such as device and channel. Compute them once in streaming pipelines, store them in a feature store, use the same definitions for training and real-time scoring, and retrain on analyst outcomes. This playbook lists the features that work across banking, telecom, credit unions and fintech.
Published . Last updated
On this page
Which feature families matter most?
| Family | Example features | Why it works |
|---|---|---|
| Velocity | Transactions in the last 10 minutes, 1 hour, 24 hours | Fraud is fast; criminals move before they are stopped |
| Deviation | Amount compared with the customer’s 90-day average | Normal for one customer is fraud for another |
| Newness | New device, new payee, new country, days since phone change | Most fraud starts with something new |
| Network | Accounts sharing a device, unrelated senders to one account | Fraud runs in rings |
| Context | Channel, time of day, merchant category, cell tower | Separates real changes from suspicious ones |
Industry examples
- Banks: new beneficiary plus large instant transfer plus recent password reset.
- Telcos: outgoing-to-incoming call ratio, cell changes per day, calls to rare destinations.
- Credit unions: spend versus the member’s own baseline, contact details changed in the last 7 days.
- Wallets: wallets per device, minutes between money in and money out.
How do you keep training and scoring consistent?
Define each feature once and use that definition for both training and live scoring. If they drift apart, models look good in testing and fail in production. On DataNature, features are defined in the feature store, computed by StreamHouse, and served from Lakebase. See feature store and MLOps.
Where do labels come from?
- Confirmed fraud from chargebacks, customer claims and investigations
- Analyst decisions on alerts
- Blocks and confirmed SIM box SIMs for telcos
Labels arrive late. Train on data old enough for the outcome to be known, and track how long labels take to arrive.
How should you measure a fraud model?
Accuracy is misleading because fraud is rare. Use:
- Detection rate at a fixed alert volume your team can handle
- False positive ratio: false alerts for each confirmed fraud
- Value detected: money saved, not just cases found
- Time to decision for real-time use cases
How do you explain decisions?
Return the top reasons with every score, such as "new device" or "payee received money from 14 unrelated accounts today". Analysts work faster, customers get clearer answers, and model risk reviewers can check the model. Atlas keeps the score, the reasons and the feature values for audit.
Next steps
Start with velocity, deviation and newness features on one channel. Add network features once the basics work. The architecture is in how to build real-time fraud detection on a lakehouse.
Which features need StreamHouse, and which can be batch?
| Feature family | Example | Freshness needed | Built where |
|---|---|---|---|
| Velocity | Transactions per card in the last 10 minutes | Seconds | StreamHouse in-flight aggregation |
| Novelty | First transaction with this merchant or device | Seconds | StreamHouse, joined with history |
| Session | Login followed by payee change and transfer | Seconds | StreamHouse, keyed by customer |
| Behavioural baseline | Average spend by hour and channel over 90 days | Daily | dbt Gold, feature store |
| Network | Shared devices or beneficiaries across accounts | Hourly to daily | Spark job over Gold |
| Profile | Tenure, product holdings, KYC risk band | Daily | dbt Gold |
The model reads both at scoring time: short-window features from the hot path and slower features from the feature store. Every feature value used in a decision is written back with the score, which makes model risk review and retraining straightforward.
How does DataNature Atlas keep this secure and compliant?
Investigators see what they need; analysts see masked identifiers; every score and decision is auditable. DataNature Atlas is the platform's governance module:
- Knows your data: a catalog with automated PII detection and classification across Bronze, Silver and Gold.
- Controls access: role-based grants plus attribute-based policies, row filters and column masks (null, redact, partial, hash, year only).
- Governs AI: Voir agents are scoped principals; masks, row filters and PII guardrails apply to every tool call and answer.
- Proves it: lineage from source to metric to decision, and an audit log of who or what accessed which data.
- Stays local: Atlas runs inside your deployment, so policies, lineage and logs never leave your jurisdiction.
These controls support GDPR, PIPEDA, Quebec Law 25, CCPA, HIPAA and sector rules such as OSFI E-23. Read DataNature Atlas in depth or see security and compliance.
Platform components in this article
Learn more about each DataNature product and module mentioned here.
See it on your own data.
Book a demoDemos and quotes are handled by Dlytica, the company behind DataNature.