Projects / Which card transactions to approve, review or decline

Which card transactions to approve, review or decline

A three-tier policy on a fraud score cuts the monthly cost of fraud, false declines and reviews by at least $64k.

Self-built case study. IEEE-CIS public transaction data. Not client or employer work.

PythonSQL
Board deckFor leadershipTechnical documentMethod, results, notebook and files

A three-tier policy on a fraud score cuts the monthly cost of fraud by at least $64k

Total cost each month

–$64k

at least; $127k against the base rules

Fraud dollars stopped

71.2%

vs. 38.4% with the current rules

Good transactions declined

2.46%

inside the 3% limit; 0.28% today

Decision: approve a test on 10% of transactions.

Source: IEEE-CIS card transactions; analysis. Public data; illustrative results.

Fixed rules stop only 38.4% of fraud dollars, and stricter rules decline good customers

Total cost each month with the current rules: $299,432

Fraud loss that gets through

38.4%

of fraud dollars stopped today

Cost of false declines

3% + $10

lost margin and goodwill for each good customer declined

Cost of reviews

120

alerts a day: the capacity of 3 analysts

The current rules and the costs are assumptions. Source: IEEE-CIS card transactions; analysis. Public data; illustrative results.

Q: Which policy gives the lowest total cost with the same review team?

Score each transaction, then approve, review or decline by its expected cost

Tier 1

Approve

The lowest expected cost. No friction for the customer.

Tier 2

Review

The middle tier, inside the capacity of 120 alerts a day.

Tier 3

Decline

The highest expected cost. Limit: 3% of good transactions.

Expected cost in dollars sets the tier, not the fraud score alone.

Source: IEEE-CIS card transactions; analysis. Public data; illustrative results.

On the test months the new policy stops 71.2% of fraud dollars and costs 42% less

New policy against the current rules

Later months that the score never saw; same review team

Share of fraud dollars stopped

New policy
71.2%
Current rules
38.4%

Total cost each month

New policy
$172,612
Current rules
$299,432

–42%

total cost, against the base current rules

The price of the saving

  • Good transactions declined: 2.46% vs. 0.28% today
  • The limit is 3%

Source: IEEE-CIS card transactions; analysis. Public data; illustrative results.

The saving stays above $53k a month under every assumption we tested

Decrease in total cost each month

$, new policy against the current rules

$126,820

Base rules

$64,274

Strongest rules

$53,544

Lowest case

$64k

each month: the figure we report

So what: the current rules are an assumption, so the result against the strongest rules is the one we can defend.

Lowest case = the lowest result across the cost assumptions tested. Source: IEEE-CIS card transactions; analysis. Public data; illustrative results.

Test the new policy on 10% of transactions, with the 3% limit as a stop rule

Actions

  1. Test the new policyon a random 10% of transactions, about 16 days
  2. Retrain the score each monthmean PR-AUC 0.576 vs. 0.487
  3. Keep the team at 3 analystsa fourth saves only $1,534 a month

Risks and controls

More good customers are declined: 2.46% vs. 0.28% todayThe 3% limit is a stop rule of the test
The cost of a false decline is an assumptionFinance and Marketing confirm it before the test; it sets the size of the decline tier
The score loses 15% of its PR-AUC after month 4Monthly retrain, and a check of the PR-AUC each month

Source: analysis. Public data; illustrative results.

1 / 6

Total cost splits into three parts, and the policy must lower the sum

  • Total cost each monthCurrent rules: $299,432
    • Fraud loss that gets through
      • Fraud dollars attemptedFixed: the test months
      • Share of fraud dollars stoppedThe lever: the score and the three tiers

        Does the new policy stop more fraud dollars?

        Yes. 71.2% vs. 38.4%.

    • Cost of false declines
      • Good transactions declinedThe limit: 3% of good transactions

        Does the decline rate of good customers stay inside the limit?

        Yes. 2.46%, up from 0.28% today.

      • Cost of each false decline3% margin on the amount, plus $10 of friction

        Does the saving hold if this cost changes?

        Yes. At least $53,544 each month under every cost assumption.

    • Cost of reviews
      • Alerts reviewed each dayCapacity: 3 analysts, 120 alerts

        Does a larger team pay?

        No. 4 analysts is the lowest-cost team, but only $1,534 each month below today. No change before the test.

      • Cost of an analyst dayFixed: $300

Lever what the recommendation changes Limit what must not get worse Fixed outside the decision