Research Question

How can a fraud system explain not only that an event is risky, but why the actor may be behaving strategically?

This research direction is connected to the local StrategicFraudIntelligenceSystem project. The current codebase is a Spring Boot scaffold with CI, security baseline, package boundaries, and project documentation. The Bayesian and game-theoretic layers are documented roadmap phases, not finished implementation.

Planned Research Stack

The project roadmap separates the intelligence layers deliberately:

| Layer | Purpose | | --- | --- | | Deterministic rules | Explain simple fraud indicators first | | Static risk scoring | Convert rule signals into a normalized score | | Bayesian inference | Update P(fraud | behavior) as evidence arrives | | Behavioral sequences | Detect suspicious action transitions | | Graph intelligence | Detect linked accounts and fraud rings | | Game theory | Classify honest, opportunistic, and adversarial behavior | | Decision optimization | Balance fraud loss against user friction |

That sequence matters. Bayesian and game-theoretic models need clean event history, labels, explanations, and review outcomes before they are credible.

Why Bayesian Reasoning Fits

Fraud rarely has one perfect signal. A Bayesian layer can combine weak evidence:

prior fraud probability
  + behavior likelihoods
  + evidence quality
  + historical review outcomes
  -> updated fraud probability

The research challenge is to keep the probability explainable enough for analysts and product teams to trust.

Why Game Theory Fits

Fraud is adversarial. Once a platform changes its rules, attackers adapt.

The game-theory layer will explore strategy classes such as:

• honest;
• opportunistic;
• adversarial.

The goal is not only to flag risk, but to explain whether behavior looks accidental, incentive-driven, or strategically abusive.

What This Will Demonstrate

This research will show how to design an intelligence system in stages:

• deterministic rules before probabilistic scoring;
• probability updates before strategic inference;
• graph and anomaly signals before autonomous decisions;
• explanations and audit trails at every layer;
• decision policies that account for business cost and user friction.

The honest portfolio framing is important: this is a planned research track with a scaffolded system, not a completed fraud engine yet.