The Dashboard That Passes Every Check
A retail bank deploys a machine-learning model to flag suspicious transactions for its AML team. Validation shows 96% accuracy on the holdout set. The model flags 1.8% of transactions, and the AML team can review 2,000 alerts per week. Total weekly volume is 400,000 transactions. The fraud rate is 0.3%. The vendor's dashboard reports '96% accurate' and the compliance officer cites it in the board pack. A data analyst notices the holdout set was drawn from the same quarter as training and that the model was tuned on the holdout set during development. Separately, the alert queue shows that 70% of reviewed alerts are closed as false positives, and the AML team has begun deprioritizing alerts from a segment that historically generates most of them.
1. Evaluate the 96% accuracy claim and identify the metrics the board should actually see.
Read the worked answer
96% accuracy is nearly meaningless at a 0.3% base rate. With 400,000 weekly transactions, true positives are about 1,200. A model that flagged nothing would be 99.7% 'accurate.' The 96% figure is dominated by the majority class.
Also, the validation is compromised: a holdout drawn from the same quarter as training leaks temporal information, and tuning on the holdout inflates the estimate. The 96% is an in-sample-ish number, not a generalization estimate.
Metrics the board should see:
- Precision (of flagged alerts, how many are true): at 1.8% flags = 7,200 alerts; if 70% are false positives, precision ≈ 30%.
- Recall (share of true fraud caught) - unknown and critical.
- Alert volume vs. review capacity: 7,200 alerts vs. 2,000 reviews/week means 5,200 alerts go unreviewed. The team is structurally overloaded.
- False-negative cost vs. false-positive cost, and a proper out-of-time validation on a later period.
The board is being shown a vanity metric that hides a capacity and validation problem.
2. The team's deprioritization of one segment - what risk does this create and how should it be governed?
Read the worked answer
Deprioritizing a segment because it generates many false positives creates a feedback loop: fewer reviews in that segment means fewer confirmed frauds there, which makes the model look more accurate there and justifies further deprioritization. The segment may also be a protected class or a proxy for one, raising fair-lending and fair-treatment risk.
Governance responses:
- Measure recall and precision by segment, not just overall.
- Set a minimum review floor per segment and monitor alert-closure rates by segment over time.
- Require documented, approved criteria for any deprioritization, with model risk and compliance sign-off.
- Add drift monitoring and periodic out-of-time revalidation; retrain only on data not used for tuning.
- Report model limitations and the unreviewed-alert backlog to the board, not just the headline accuracy.
The exam point: a model that looks good on a dashboard can conceal a capacity failure and a fairness risk. Governance must measure what the dashboard omits.