Leaders Insights
Leaders Insights

Rester au meilleur niveau, un peu chaque jour.

DomainesMarketingDataFinanceIA
RessourcesApprendreTestOutilsBlogGlossaire
© 2026 Leaders Insights — Tous droits réservés.
Formations/Data in insurance/Data in insurance/Building fraud detection models on claims data
3/4+150 XP

Data in insurance

1Reading the actuarial data stack that prices every policy+1502Turning telematics and new risk signals into pricing power+1503Building fraud detection models on claims data+1504Governing fairness and compliance in pricing models+150

Building fraud detection models on claims data

# Building Fraud Detection Models on Claims Data

A bodily-injury claim lands on an adjuster's desk: a rear-end collision, three occupants, all reporting soft-tissue neck injuries, all treated at the same clinic, all represented by the same attorney within 48 hours of the accident. Individually, nothing is illegal. Together, it is a textbook signature of an organized fraud ring.

Your job as the data person is to teach a model to see that pattern, and to do it without flagging the thousands of legitimate whiplash claims that look superficially similar.

Why fraud detection is a data problem, not a rules problem

Insurers have used rules for decades. A rule says: "flag any claim over $10,000 with treatment starting more than 30 days after the accident." Rules are transparent and easy to audit, but fraudsters learn them and stay just under the threshold.

Soft fraud (exaggerating a real claim) and hard fraud (a fully staged or fabricated loss) both evolve. The interesting, expensive fraud is organized: rings of providers, attorneys, and "patients" who recycle the same playbook across many claims.

That is a pattern problem. And patterns across thousands of claims are exactly what data models are good at.

The Coalition Against Insurance Fraud publishes solid, free background on how these schemes work. Skim their fraud resources before designing features; understanding the crime makes you better at engineering signals for it.

Start with the unit of analysis

Beginners model "the claim." Fraud rings live *between* claims. So you actually need three levels:

  • Claim level: one loss event.
  • Entity level: a person, clinic, vehicle, or attorney appearing across many claims.
  • Network level: how those entities connect.

A single suspicious claim tells you little. The same chiropractor appearing on 200 low-speed rear-end claims with near-identical treatment plans tells you a lot.

Feature engineering: turning the scene into signals

Let us convert our opening claim into features a model can read.

Claim-level features

  • Days from accident to first treatment (very short or oddly delayed both matter).
  • Treatment intensity: number of visits relative to injury severity.
  • Ratio of soft-tissue (hard to disprove) to objectively verifiable injuries.
  • Was the accident reported to police? Were there independent witnesses?
  • Time of day and location (some staged accidents cluster in specific spots).

Entity-level features

This is where organized fraud shows up.

  • How many claims does this clinic, attorney, or repair shop appear on this year?
  • What share of a provider's claims involve the same injury type?
  • Do the same three occupants appear together on prior accidents?

Velocity and coincidence features

  • Number of occupants who all filed injury claims (staged crashes maximize claimants per crash).
  • Attorney retained within X hours (fast legal representation before pain would realistically set in).
  • Multiple claimants sharing an address or phone number.

None of these is proof. Each is a weak signal. The model's job is to combine many weak signals into a calibrated risk score.

The graph is where rings hide

Standard tabular models miss the connective tissue. Fraud rings are naturally a graph: nodes are people, clinics, vehicles, phone numbers, and bank accounts; edges are shared claims.

Build the graph, then compute features from it:

  • Degree: how many claims a node touches.
  • Connected components: tightly linked clusters of claimants, providers, and attorneys that recur together.
  • Shared-attribute links: two "unrelated" claimants using the same phone number is a strong edge.
python
import networkx as nx

# Nodes: claimants, providers, attorneys, phones, VINs
# Edges: appears-on-same-claim or shares-attribute
G = nx.Graph()
G.add_edge("claimant_A", "clinic_9")
G.add_edge("claimant_B", "clinic_9")
G.add_edge("claimant_A", "phone_555")
G.add_edge("claimant_B", "phone_555")  # suspicious shared attribute

# Find tight clusters that recur across claims
components = list(nx.connected_components(G))
suspicious = [c for c in components if len(c) > 15]

A cluster of fifteen "strangers" all routing through one clinic and one phone number is the ring. That structural fact is a feature you feed into the downstream model.

Choosing the model: supervised, unsupervised, or both

You rarely have clean fraud labels. Confirmed fraud is a tiny, biased slice: only the cases someone already caught. So use a layered approach.

Supervised models

Train on confirmed fraud (investigated and proven) plus confirmed legitimate claims. Gradient-boosted trees (XGBoost, LightGBM) handle mixed features well and give feature importances adjusters can inspect. Good for catching *known* patterns.

The weakness: they only learn fraud that looks like past fraud.

Unsupervised anomaly detection

For novel schemes, use anomaly detection: isolation forests, autoencoders, or clustering. These flag claims that deviate from normal without needing labels. The goal is not to declare "fraud"; it is to say "this is statistically unusual, look closer."

The practical stack

Most mature insurers run both: a supervised score for known typologies, an anomaly score for the unknown, and graph features feeding both. A claim that is high on all three goes straight to the Special Investigations Unit (SIU), the internal team that investigates suspected fraud.

The false-positive problem is the whole game

Here is the trap. Suppose fraud is roughly 10 percent of claims (a commonly cited estimate; real rates vary by line and are hard to measure precisely because undetected fraud is, by definition, uncounted).

If your model flags 30 percent of claims to "be safe," you have buried your SIU. Adjusters lose trust, ignore the alerts, and the whole system dies. This is alert fatigue, and it kills more fraud programs than bad math does.

Two disciplines fix this:

1. Optimize for precision at low volume. You do not need to catch every fraud. You need the top 2 percent of flagged claims to be genuinely worth an investigator's time. Measure precision at k: of the top 100 claims you flag, how many are real? That is the number your SIU cares about.

2. Make scores explainable. An adjuster will not act on "the model said 0.87." Give reasons: "same clinic on 180 claims, three occupants co-appeared on a prior loss, attorney retained in 6 hours." Tools like SHAP values turn a score into a short list of drivers. Explainability is also increasingly a regulatory expectation, since flagging a policyholder as suspicious has real consequences.

Governance: this is regulated territory

Fraud models touch protected decisions. A few guardrails:

  • Do not use prohibited or proxy variables. Features correlated with race, religion, or national origin can create illegal discrimination even if unintentional. Test for disparate impact.
  • Keep a human in the loop. The model routes claims for review; it should not auto-deny. Denial is a human, documented decision.
  • Document everything. Regulators and courts may ask why a claim was investigated. "The model flagged it" is not enough.

The NAIC (National Association of Insurance Commissioners) has ongoing guidance on the use of AI and data in insurance; it is worth tracking as expectations tighten through 2026.

Vérification des acquis

1. Why does the lesson argue that fraud detection is fundamentally a data problem rather than a rules problem?

2. A single low-speed rear-end claim with soft-tissue injuries is, on its own, hard to distinguish from legitimate whiplash. What does this illustrate about the appropriate unit of analysis?

3. The lesson warns against flagging the 'thousands of legitimate whiplash claims that look superficially similar' to fraud. What core modeling concern does this highlight?

CHOIX MULTIPLES

4. Select ALL correct answers. According to the lesson, which levels of analysis are needed to detect organized fraud rings?

Sélectionnez toutes les réponses correctes.

CHOIX MULTIPLES

5. Select ALL correct answers. What is true about the distinction between soft fraud and hard fraud as described in the lesson?

Sélectionnez toutes les réponses correctes.

Closing the loop: feedback makes it work

A fraud model is not "trained once." Every SIU investigation produces a new label: fraud confirmed, or cleared. Feed those outcomes back.

Two cautions:

  • Confirmation bias. If you only investigate what the model flags, you only ever learn about claims the model already suspected. Periodically investigate a small random sample of *low-score* claims to check for fraud the model is missing. Otherwise the model gets confidently blind.
  • Concept drift. Rings adapt. When your top signal (say, one clinic) gets shut down, the pattern moves. Monitor whether your feature importances shift over time, and retrain on a rolling window.

Bringing it back to our claim

Our opening claim: three occupants, one clinic, one fast attorney, all soft-tissue. Run it through the stack:

  • Claim features: high occupant count, fast attorney, low-verifiability injuries. Elevated.
  • Entity features: that clinic appears on an unusual share of low-speed rear-end claims. Elevated.
  • Graph: two of the three occupants co-appeared on a prior loss together. Strong edge.

Combined score: high, with three explainable drivers. It routes to SIU with a clear rationale, and the ten legitimate whiplash claims processed that same afternoon (single occupant, normal treatment cadence, no shared entities) sail through untouched.

That is the goal: catch the ring, spare the honest claimant, respect the adjuster's time.

Key Takeaways

  • Model the network, not just the claim. Organized fraud lives in connections between clinics, attorneys, and repeat claimants; graph features expose it.
  • Layer supervised and unsupervised models. Supervised scores catch known typologies; anomaly detection catches novel ones.
  • Precision at low volume beats raw recall. Protect your SIU from alert fatigue by flagging only the highest-value claims, with explainable reasons.
  • Close the feedback loop, and audit for blind spots. Sample low-score claims and monitor for drift so the model keeps learning as rings adapt.
  • Treat it as regulated decision-making. Test for disparate impact, keep humans deciding denials, and document why every claim was flagged.

Précédent

Turning telematics and new risk signals into pricing power

Suivant

Governing fairness and compliance in pricing models