The Outlier Lab

Basket and SKU Anomalies

Plausibility of what was bought (basket composition, offer-qualifying SKU concentration, quantities, prices and return behavior), used to judge receipts, offer claims, refunds and orders.

What it measures

Image forensics asks "is this document real?"; basket analysis asks "is this purchase believable?" The second question survives even when the first is answered by a perfect fake.

Feature familyExamples
Offer-qualifying concentrationShare of basket value or line items that qualify for active offers; baskets containing only qualifying SKUs; qualifying items vs. the platform-wide average for that merchant
Quantity and limit behaviorQuantities exactly at or just under offer limits; multiples of the minimum required; the same SKU repeated across many receipts and accounts
Price plausibilityItem prices vs. known shelf, promotional and regional price ranges for that retailer and date; discounts or coupon lines inconsistent with retailer practice; tax computed incorrectly for the store's jurisdiction
Co-occurrenceWhether items in the basket normally appear together; "staple" items (produce, milk, bags) absent from very large grocery baskets; product mixes that look assembled rather than shopped
Store and time contextSKU sold at that banner or region; product launched yet; store hours; purchase time consistent with store location
Diversity across accountsBaskets from many accounts that are nearly identical in composition; sudden convergence of many users on the same item set
Post-purchase behaviorQualifying items later returned or refunded (buy–claim–return); partial returns that leave the payout intact

Why it matters

In Cashback, Coupons and Rewards, offers are funded by brands to drive purchase of specific products, so the qualifying SKU is the money trigger. Abusers optimize for it: they buy (or claim to buy) precisely the items that pay, in precisely the quantities that maximize payout, and often return them afterwards. Legitimate shoppers' baskets are noisy (mixed categories, staples, non-qualifying items, prices that reflect local promotions), and abusers' baskets usually aren't.

Basket analysis is also the counterweight to generative image tools. As AI-generated receipts became the majority of flagged fake expense receipts by spring 2026 (per AppZen data), pixel-level detection lost reliability. A generator still has to choose a basket, though, and producing baskets that match real price, tax, co-occurrence and regional assortment data across thousands of submissions is much harder than producing a convincing image.

Elsewhere on the money path:

How to collect it

  • Extract line items from receipts (OCR) and order systems, normalized to a product catalog: SKU or UPC/GTIN where available, otherwise normalized descriptions.
  • Assemble reference data: the offer catalog (which SKUs qualify, limits, value, dates); retailer assortment and price ranges by region and date (from partner feeds, e-receipt integrations, or learned from your own high-confidence submissions); and tax rules by jurisdiction.
  • Build population baselines: co-occurrence statistics and typical basket size and mix per merchant, from trusted accounts and verified transactions.
  • Link return and refund events back to the original purchase and payout.
  • On build vs. buy, basket plausibility depends on your offers and your merchant mix, so it's primarily built. Product catalog and retail price data can be licensed; ecommerce fraud vendors score cart contents on generic resale-risk features.

How to use it

  • Score at submission: qualifying concentration × account trust × image risk. A brand-new account submitting a receipt that is 100% offer-qualifying, at limit quantities, from a merchant it has never shopped, deserves review even if the image is clean.
  • Offer-level monitoring: track qualifying-SKU claim volume vs. retail sales expectations and brand forecasts. Claims exceeding plausible sell-through for a region are a sentinel for counterfeit receipts or a leaked exploit (Playbook - Business Logic Exploit Response Members).
  • Cross-account similarity: cluster baskets by composition; tight clusters across unrelated-looking accounts indicate shared-receipt rings or templates (Entity Link Analysis, Fraud Rings and Collusion).
  • Tie to returns: match qualifying purchases against retailer return feeds or card refund events where data partnerships allow, and claw back where terms permit (Reward Vesting and Clawback).
  • Design feedback: if one offer attracts disproportionate anomalies, change the offer: add per-household limits, require a minimum non-qualifying basket, or vest payouts past the retailer return window (Promotion Design Guardrails).

Strengths and limitations

Strengths: holds up as image fabrication improves; directly tied to offer economics; interpretable for reviewers and brand partners; useful across receipts, orders, returns and coupons.

Limitations: depends on reference data quality, and thin price and assortment data for small merchants weakens it; adaptive adversaries can degrade concentration features over time, so pair them with external transaction confirmation and cross-account similarity; legitimate customer behavior shifts quickly when a deal goes viral.

False positives and fairness

  • Extreme couponers and deal communities deliberately shop only promoted items at maximum quantities. They follow the rules on purpose and are often a segment the business values.
  • Small business owners and resellers using personal accounts to buy in bulk.
  • Shoppers who split trips (one quick trip just for the deal item).
  • Households stocking up during sales or shortages.
  • Regional and cultural differences in basket composition that baselines built on majority populations may misread.

Distinguish "optimized but real" from "fabricated or reversed" using external confirmation and return linkage. Handle optimization through offer design and terms, and keep it out of fraud labels.

Privacy and compliance considerations

Line-item data can reveal sensitive information (pharmacy, baby, alcohol, health products). Minimize retention of non-qualifying items, restrict reviewer access, and disclose use in privacy notices. Sharing basket-level data with brand partners should be aggregated and consistent with contracts and privacy commitments. See Privacy and Data Use in Fraud Prevention.

Mechanisms it helps detect

Receipt Fraud (Cashback) · Card-Linked Offer Abuse · Coupon Counterfeiting and Misredemption · Refund and Return Fraud · Promo Code Abuse · Business Logic Exploitation · Card-Not-Present Fraud · Triangulation Fraud · Fraud Rings and Collusion

Receipt and Document Image Forensics · Promo and Reward Redemption Patterns · Dispute and Return History · Payment Instrument Signals · Receipt and Purchase Verification · Promotion Design Guardrails · Claims and Returns Policy Controls · Playbook - Receipt and Offer Fraud Members · Playbook - Refund and Claims Abuse Members · Case Study - Talens Counterfeit Coupon Scheme (2020)

Sources and further reading