Return Fraud Risk Scoring

Return fraud risk scoring ranks each return request by how closely it matches patterns associated with abuse, wardrobing, receipt fraud, and organized retail crime. The model does not approve or deny the return. It produces a risk score and short rationale that a loss-prevention (LP) analyst uses to decide whether to investigate, request more evidence, or process the return under normal policy.

Inputs typically include the customer's transaction and return history, item category and value, time between purchase and return request, channel (in-store, ship-from-store, ecommerce), payment method, and known loss patterns from prior cases. When those signals are incomplete, especially when transaction history is missing, the system should return empty output rather than invent a score. A missing history is not "low risk"; it is "not enough to score."

Refund volume and speed pressure create a gap between policy and practice. Agents and self-serve flows are optimized to clear the queue. Fraud and abuse hide in that queue as requests that look ordinary until you stack them against history and category norms.

A score gives LP a triage layer. High-risk requests get human review before money leaves. Medium-risk requests may get a lighter check (proof of purchase, serial verification, condition photos). Low-risk requests stay on the standard path. The score does not replace policy, disposition rules, or the refund decision. It ranks attention.

Related workflows that sit next to scoring include Automated Return Disposition for routing accepted returns, Refund Policy Compliance Checker for policy fit, and Return Reason Classification for structuring the customer's stated reason. Scoring answers a different question: how likely is this request to be abusive or fraudulent given what we know?

Useful models combine several signal families rather than a single red-flag rule.

Transaction and return history. Frequency of returns relative to purchases, concentration in high-resale or high-wardrobing categories, refund method preference, and prior exceptions or investigations all matter. A first-time buyer with one return is not the same case as a customer whose refund rate and category mix match known loss cohorts.

Item and category context. Apparel and accessories often show different abuse patterns than hard goods. Serial-tracked, high-ASP, or easily resold items may warrant tighter scrutiny when timing or channel also look off. Category alone never proves fraud; it adjusts base rates.

Timing and channel. Same-day or near-window returns, returns after heavy promotional use, ship-to-home then return-to-store mismatches, and returns without a matching order ID are classic investigation triggers. Timing is evidence for the analyst, not an automatic deny.

Known loss patterns. LP case libraries encode MO patterns: empty-box returns, switched merchandise, counterfeit receipts, and multi-account clusters. Models trained or calibrated against those patterns can surface similarity without claiming certainty.

When any required history field is absent (no linked order, no prior customer graph, incomplete POS join), emit no score. Downstream systems should treat empty output as "queue for manual identity or order match," not as clearance.

The model scores risk; LP still investigates. That separation protects customers and the brand.

1. Ingest the return request with order ID, customer ID, items, reason code, and channel.
2. Run scoring only when required history is present. If not, return empty output and flag for data repair or manual lookup.
3. Present score, top drivers, and comparable cases to an LP analyst or authorized supervisor, not as a customer-facing message.
4. Analyst decides next action: proceed under policy, request evidence, open an investigation, or escalate for organized retail crime review.
5. Log outcome (false positive, confirmed abuse, inconclusive) so calibration and thresholds can improve without turning the score into an auto-deny rule.

Do not auto-deny a return from the model score alone. Denial or refund refusal is a policy and investigation outcome. Auto-deny from a probabilistic score creates false refusals, chargeback risk, and trust damage that LP cannot easily reverse.

Risk scoring is most effective as a gate before refund execution and before automated disposition for high-value or high-risk SKUs. Policy compliance can run in parallel: a request can be policy-compliant and still high risk, or out of policy and low fraud risk. Reason classification improves analytics and coaching; it should not be confused with fraud likelihood.

A practical operating model:

• Low score + policy OK → standard refund or exchange path.
• Elevated score → hold for LP review; do not promise instant refund in self-serve UI until review completes (communicate delay honestly).
• Empty score (missing history) → resolve identity/order linkage first; do not invent a default low score to keep automation running.

Disposition automation (Automated Return Disposition) should consume the LP decision after review, not the raw model score as a silent deny signal.

Thresholds and calibration. Start with conservative high-risk thresholds so the first wave of reviews is high precision. Expand coverage as false-positive rates and analyst capacity allow. Recalibrate when assortment, promotions, or return windows change.

Explainability. Analysts need drivers they can act on (return rate vs peers, category mix, timing anomalies), not a black-box number. Opaque scores get ignored or over-trusted.

Privacy and access. Scores and case comparisons are LP tools. Restrict who can see them, retain investigation notes under existing privacy and retention policies, and avoid surfacing internal risk labels to store associates in ways that invite profiling without process.

Failure modes to design for. Missing history must yield empty output. Conflicting IDs (guest checkout, shared cards, gift returns) need explicit handling. Promotional spikes will shift base rates; treat seasonality as a calibration input, not as proof of fraud.

Success measures. Track investigation yield (confirmed abuse per reviewed return), false-positive burden on good customers, refund leakage avoided on confirmed cases, and time-to-decision. Optimize for better triage, not for higher deny rates.

Return fraud risk scoring earns its place when it makes LP attention scarce and accurate: score when the data supports it, stay silent when it does not, and leave deny-or-refund decisions to humans with policy and evidence in hand.