Fraud Systems Need Trust Before Intelligence

It is tempting to start a fraud platform with machine learning. The promise is attractive: feed the system enough data and it will discover abuse patterns automatically.

But early fraud systems usually do not fail because they lack advanced models. They fail because they cannot explain their decisions.

Before Bayesian inference, graph intelligence, or game theory, the platform needs a deterministic foundation.

Start With Events

Fraud reasoning begins with clean facts:

• who acted;
• what event happened;
• when it happened;
• which tenant owns the event;
• which identifiers were involved;
• whether this event is a duplicate;
• what evidence was preserved.

If event ingestion is unreliable, every later model inherits that uncertainty.

Rules Are Not Primitive

Rules are sometimes treated as unsophisticated compared to AI. That is the wrong comparison.

A good rule engine gives the system important properties:

• versioned logic;
• reproducible evaluation;
• human-readable explanations;
• clear score impact;
• auditability;
• testable boundaries.

For example:

Rule: New account requesting high-value refund
Evidence: account age 2 hours, refund value $420
Impact: +35 risk points
Explanation: refund requested before trust history exists

That explanation is immediately useful to an analyst.

Then Add Probability

Bayesian reasoning becomes useful after the deterministic layer exists.

The system can then ask:

Given this behavior and prior history,
how should the fraud probability update?

But the probability should still point back to evidence. A fraud probability that cannot explain its inputs is hard to trust and hard to debug.

Game Theory Comes Later

Fraud is adversarial. Users adapt when platforms change their rules.

Game-theoretic thinking is useful when the system starts distinguishing behavior classes:

• honest;
• opportunistic;
• adversarial.

But that layer should come after the platform can already ingest events, evaluate rules, score risk, store decisions, and collect analyst feedback.

The Build Order Matters

The right sequence is:

events
  -> deterministic rules
  -> risk scoring
  -> explanations
  -> analyst feedback
  -> probability
  -> graph and anomaly signals
  -> strategic behavior models

This is less glamorous than starting with AI. It is also more likely to produce a fraud system that product teams, analysts, and customers can trust.