Choosing an approach

Transaction monitoring: rules, graphs or machine learning?

· 8 min read

Every payment company ends up building some form of transaction monitoring: a system that looks at payments and accounts and decides what to allow, what to review and what to block. There are three main ways to do it. Each catches different fraud, and each fails in its own way.

1. Rules

Rules are explicit conditions: “block payments over this amount from accounts younger than seven days”, “review any card used in two countries within an hour”.

Strengths

  • Easy to understand, explain to auditors and change quickly.
  • Good for hard policy: limits, sanctions lists, blocked countries or BINs.
  • Work from day one, with no historical data.

Weaknesses

  • Fraudsters learn the thresholds and stay just below them.
  • Rule sets grow until nobody knows why half of them exist.
  • High false positives: blunt rules block good customers.
  • They judge one payment at a time and cannot see coordinated activity.

2. Machine learning scoring

A model learns from past payments and outcomes (chargebacks, confirmed fraud, analyst decisions) and gives each new payment a risk score.

Strengths

  • Weighs hundreds of signals at once, so it is much harder to game than a single threshold.
  • Fewer false positives than rules at the same catch rate, once trained.
  • Improves as more decisions are fed back.

Weaknesses

  • Needs labelled history, which young companies often lack.
  • A generic model trained on someone else’s data may not fit your customers or your market.
  • A score with no reason is hard for analysts to act on and for regulators to accept.
  • Still mostly payment-by-payment, so organised rings can slip through.

3. Graph analysis

Graph analysis links accounts, cards, devices, IPs, phones and counterparties, then looks for clusters and flows that do not look like normal customers. It is how you find a fraud ring.

Strengths

  • Sees organised fraud: mule networks, card testing farms, multi-accounting, collusive merchants.
  • Turns many small alerts into one case, which saves analyst time.
  • Finds fraud that never produced a chargeback, so it does not depend on labels.

Weaknesses

  • Depends on good identity resolution: phones, emails and devices have to be normalised first.
  • Harder to run in real time unless the graph is kept up to date as payments arrive.
  • Clusters still need to be explained before someone can act on them.

Comparison at a glance

QuestionRulesMachine learningGraph
Works without history?YesNoYes
Catches organised rings?RarelyPartlyYes
Explainable?YesOnly with extra workYes, if cases are written up
Easy to game?YesHarderHard

The answer is usually all three

The strongest setups layer the approaches rather than picking one:

  1. Rules for policy: limits, sanctions and lists that must always apply.
  2. Graph analysis to find rings in your history and as new links appear.
  3. A model trained on your own data that scores every new payment in real time, using ring membership as one of its strongest signals.
  4. Plain-language reasons for every alert and every ring, so analysts can act and auditors can follow.

Argos is built that way. It connects read-only to your database or a file, works out what each column means, masks personal data, finds the rings already in your data, and then scores every new payment with a model trained only on your company’s own data and decisions, alongside the rules and lists you define.

More from the blog