Six Questions a Payment-Fraud Dashboard Should Answer
Abstract
Fraud dashboards usually report activity — alerts raised, cases closed, blocks applied. Activity is easy to count and tells an operator almost nothing about whether losses are being prevented. This note sets out six questions a payment-fraud dashboard should answer, with the metric and the trap for each, and argues that the last two questions are the ones that separate a monitoring screen from a control.
1. The problem
A dashboard full of activity metrics creates a false sense of control. Alert volume rises, cases close, the screen is green — while the loss rate climbs, because nobody is measuring the loss rate on that screen. The metrics that matter are usually absent because they require joins across systems: the payment record, the dispute record, the case record and the customer record.
2. Why current practice falls short
Three habits get in the way.
Counting inputs. Alerts, blocks and reviews are inputs. They say how busy the team is.
Aggregating away the detail. A monthly loss figure hides the fact that losses doubled in one payment method, one merchant category and one time window. Aggregates conceal the pattern you need.
No counterfactual. Nobody records what would have happened without the control, so nobody can say whether the control earned its cost.
3. The proposed approach
Six questions, in the order an operator should read them.
| Question | Metric | The trap |
|---|---|---|
| What is leaking? | Loss rate per ₹1,000 processed, by method | Reporting absolute losses; volume growth masks a falling rate |
| Where? | Loss rate by channel, merchant category, geography, ticket band | Averages hide a single hot cell — always break down to the cell |
| From whom? | Share of loss from repeat actors and linked accounts | First-party fraud looks like a customer problem until you count repeat devices |
| Why now? | Change versus the prior period, with the trigger identified | Attributing drift to "seasonality" without checking |
| What did we miss? | False-negative rate from post-hoc review of approved-and-disputed transactions | Only measuring the alerts you raised |
| What did we get wrong? | False-positive rate and customer friction cost per block | Treating declines as free |
The two that matter most are the last two. A false-negative review — sampling transactions that were approved and later disputed — is the only way to see the fraud your rules did not catch. And a false-positive count is the only way to see what the control costs in lost customers, which is a real loss line that never appears in a fraud budget.
4. How it works in practice
A worked week, synthetic figures:
- Loss rate 0.41 per ₹1,000, up from 0.28 the prior week — question one flags it.
- Breakdown puts almost all of the increase in one ticket band and one payment method — question two localises it.
- Repeat-actor share in that cell is 63%, with a cluster of accounts sharing two device fingerprints — question three reframes it from "fraud spike" to "one ring".
- Change coincides with a checkout release — question four identifies the trigger.
- Post-hoc review of 200 approved-and-disputed transactions finds 9 with the same device cluster — question five proves the rule gap.
- Blocked customers in that cell number 41, of whom 6 complained — question six prices the control.
The output is not a screen. It is three sentences and one decision: tighten the rule for one cell, revert part of the checkout change, and accept the friction cost in the other.
5. Trade-offs and failure modes
- False-negative review is expensive. Sampling needs judgement and time; a 200-transaction sample per week is a real cost and worth it.
- Friction is hard to price honestly. Complaint counts undercount silent abandonment, so treat the friction figure as a floor.
- Segment definitions drift. If "merchant category" is free text, every breakdown is noise.
- Dashboard latency invites gaming. A daily figure can be managed; a weekly trend cannot.
6. What to measure
- Loss rate per ₹1,000 processed, weekly, by cell.
- False-negative rate from the weekly post-hoc sample.
- False-positive rate and complaint rate per 1,000 blocks.
- Repeat-actor share of total loss.
- Time from first signal to case decision, median and 90th percentile.
7. Conclusion
If a dashboard cannot answer "what did we miss?" and "what did we get wrong?", it is a status screen, not a control. Both questions are answerable with data most payment teams already hold — they simply require sampling and a join that nobody has been asked to build.
Disclaimer. The views here are the author's own. All figures are synthetic and used to illustrate a method.