Putting a euro figure on the fraud the system didn't catch
Remaining Fraud Damage, a method that benchmarks each customer segment against trusted customers to size the logistics refund leakage existing controls let through, in euros, per market, every week.
Senior Product Analyst, Risk & Abuse, Zalando
Produced the first euro-denominated estimate of logistics refund leakage that survived existing controls, broken down by customer segment and by market across six countries, refreshed weekly and monthly for business reviews.
Remaining Fraud Damage = Return Damage + Delivery Damage, as a share of GMV.
Everyone tracks the fraud they catch. The damage that matters for the business is the part that slips past every control: manual refund leakage on missing-delivery, item-not-received, and parcel-missing claims. It's invisible precisely because nothing flagged it, so there was no euro figure to act on.
I built the Remaining Fraud Damage measure. The idea is simple. Trusted high-value customers (A/VIP) set the honest baseline refund/damage rate, and every other value segment's excess over that benchmark, applied to its GMV, is the leakage estimate. Remaining Fraud Damage = return-damage + delivery-damage across the analysed segments, expressed as a share of GMV, tracked weekly and monthly across the top six markets. It runs on manual-refund and Salesforce case data joined to customer value segments and AbPP risk signals, built in PySpark/Databricks over the Zalando data lake.
Risk and leadership got a defensible, recurring euro estimate of the damage that survived existing controls, broken down by segment and market with weekly and monthly trends. That turned 'fraud we can't see' into a number that could be argued about, owned, and reduced.
The leakage stays invisible because nothing flagged it: no euro figure to budget against, no segment or market to target, and no way to tell whether refund damage is getting better or worse.
The fraud you catch is the easy half to talk about. The number leadership actually needed was the other half: how much logistics refund damage (missing delivery, item-not-received, parcel-missing) was leaking past every control? By definition nothing had flagged it, so there was no figure at all.
I built one. The trick is choosing an honest baseline. Trusted high-value customers (A/VIP) refund and claim at some natural rate that isn’t abuse, so treat that as the floor. Then every other value segment’s refund/damage rate above that floor, applied to its GMV, is an estimate of the excess that shouldn’t be there. Sum the return-damage and delivery-damage across the analysed segments and you have Remaining Fraud Damage as a share of GMV.
The plumbing is real e-commerce data work: manual-refund and Salesforce case tables joined to customer value segments, GMV denominators (before and after returns), manual-refund reasons, and AbPP risk signals. All of it built in PySpark on Databricks over the data lake, across the top six markets, refreshed weekly and monthly.
What it changed is the conversation. “Fraud we can’t see” became a defensible euro estimate, split by segment and market, with a trend line. You can’t reduce a number you’ve never put on the table. This put it on the table.