Fraud scoring turns the many signals around an order, a login or an account change into one number (or an accept, review or reject recommendation) so you can decide in milliseconds what to let through, what to check and what to stop. This guide explains what goes into a score, how to use it, how to judge a vendor, and where scoring stops being enough. The advice is vendor-neutral.
What a fraud score is
Every order gets a score, usually numeric and sometimes a simple approve, deny or review answer, produced by proprietary models from the data you send and the data the vendor already holds. You then act on the score: accept, reject, or send the order for manual review. Good services also return reason codes ("address unverifiable", "high velocity of use") so a reviewer knows which checks to run, rather than a bare number.
A score is a probability, not a verdict. It ranks orders by risk so you can put your effort where it pays off.
What goes into a score
Models weigh thousands of individual signals, some highly meaningful and some close to noise. The common families are:
- Identity and address data: address and phone verification, email age and reputation, reverse lookups.
- Device and network: device identification, IP geolocation and proxy detection.
- Velocity: how often the same device, card, email or address appears, and how often those details change (velocity of use, velocity of change).
- Lists: shared negative, warm and positive lists.
- Network data: activity the vendor has seen across all its other merchants. This is often the biggest advantage, because it can stop a fraudster you have never seen before.
The more complete the data you submit, the better the score. Quality matters as well as quantity: a score built on a device ID that keeps recognizing a returning fraudster is stronger than one built on many weak signals.
Types of fraud scoring
- Model-based order scoring: a vendor applies its own multi-merchant model to your card-not-present orders. Fast to deploy and bundles many tools, but everyone starts from the same base model.
- Login risk scoring: scores the login attempt, because a correct password alone does not prove the right person is logging in. Used to allow, restrict, or step up authentication, which is central to account takeover defense.
- Custom modeling: a model built on your own historical data. It takes longer and needs extensive data, but it is bespoke to your business. See is model-based scoring right for your business for how the two differ, and the best practices for custom modeling.
- Rules: hand-written internal rules and rule engines. Rules built only from past fraud need constant patching, and models usually cope better with changing fraud. Many teams run both. See should legacy rules-based systems be ruled out?
How to use the score well
- Send every data element you have. Accuracy scales with completeness.
- Run authorization checks before fraud screening, not after, so you do not pay to score orders that will be declined anyway.
- Set your own thresholds from your own chargeback history, not the vendor's generic default. Typically you auto-accept below a floor, send the middle band to manual review, and reject the top band.
- Use the reason codes to route reviews to the right tool or technique.
- Bring in an analyst for the first implementation if you do not already have one.
The tradeoff that never goes away
Better fraud screening catches more fraud but also flags more good customers. In our experience, even good services catch roughly 40 to 70 percent of fraud attempts, and a higher catch rate almost always brings a higher false-positive rate. Declined good orders are a real cost, not a rounding error: see measuring false positives is not an exact science. Choose the point on the curve that fits your margins, and ask any vendor what insult rate to expect at each score threshold.
How to evaluate a scoring vendor
Our longer guide is three core considerations for evaluating model-based fraud scoring providers. In short:
- Which tools and signals are included (address verification, geolocation, device ID, reverse lookups), and how strong the device recognition and data sharing are. For data sharing, ask how many merchants are in the pool and how many model features use shared velocity, not just shared negative lists.
- Expertise: overall data science depth and your industry specifically. Ask what modeling approach is used (for example random forests, statistical pattern recognition or neural networks) and talk to someone who can explain it.
- Machine learning: is it used, is it supervised or unsupervised, and how often is the model refreshed? Read maximizing model performance and shelf life.
- Model currency and data freshness: monthly or real-time updates beat a model trained on year-old fraud patterns.
- Output: a numeric score with a defined range plus descriptors, not just pass or fail.
- Risk sharing: whether a chargeback guarantee comes with the score.
- Coverage and fit: web, mobile browser and native app channels, manual review tooling, customization, volume pricing and how it bills (per transaction, flat, or basis points).
- Proof: a backtest on your own historical orders, measured on your own chargebacks and false positives, beats any brochure.
How to know it is working
Track four numbers monthly: fraud catch rate, false-positive rate, the share of orders going to manual review, and your chargeback rate. A model that looked good at launch drifts as fraud changes, so recheck thresholds on a schedule. To see how scoring and decision confidence fit together, read about the confidence indicator and behavioral monitoring.
Learn the method. Our Modeling & Analytics for CNP Fraud Management course explains what custom modeling is, the terms and concepts behind it, and how it fits into a fraud and risk strategy. For a vendor-neutral read on which approach fits your business, talk to a senior consultant: see our consulting services.
Guides in this section
- Model-Based Fraud Scoring: vendor checklist, how it works and sample vendors for order scoring.
- Login Risk Scoring: scoring the login event to stop account takeover.
- The rise of machine learning and AI in fraud prevention