Set the fraud threshold too conservatively and legitimate transactions are declined. Cardholders contact the bank, then may stop using the card. Revenue falls. Chargeback rates may rise when customers dispute alternative manual payments. False positives create measurable harm that is often undervalued beside false negatives.
Lowering false positives without allowing more fraud through requires more than changing one threshold. Legitimate and fraudulent activity overlap in the score distribution, while calibration affects customer segments differently. Here is how enza's fraud scoring layer approaches threshold tuning to reduce false positives without materially increasing fraud exposure.
What Score Distributions Show in Practice
A fraud score estimates probability, usually from 0 to 1. Higher values represent greater assessed fraud likelihood. On a real transaction set, a calibrated model places most activity near 0 because most transactions are legitimate, with a long tail toward 1 representing a small high-risk minority.
The threshold is the score at which a transaction is declined or flagged. At 0.7, scores of 0.7 or higher are declined. At 0.5, more transactions are declined, including more legitimate ones, while more fraud is caught. At 0.9, only the strongest fraud signals are declined, so more fraud passes but fewer legitimate transactions are rejected.
Most calibration work sits roughly between 0.4 and 0.8. Activity in this band is truly uncertain: concern is elevated, but confidence is limited. Treatment of this middle range drives much of the false positive rate and marginal fraud exposure.
Configure Thresholds by Segment
One threshold seldom suits every transaction. Customer segments show different transaction behaviour, and their score distributions may vary substantially.
A small-business owner processing several daily business transactions has a very different velocity profile from a retail consumer making a few weekly purchases. Without segment-specific training, the model may consistently over-score the owner's legitimate activity. One shared cutoff then either rejects too many small-business transactions or lets too much retail-consumer fraud through, depending on its level.
Review false positives across useful segments: amount bands, merchant category codes, time-of-day windows, card age, and geography. A segment where false positives are unusually high compared with fraud is a candidate for its own threshold adjustment.
Avoid excessive segmentation. Every added threshold dimension increases maintenance work and introduces cases where one transaction matches conflicting segments. Begin with the two or three segments showing the clearest calibration problem, measure results, then expand only if warranted.
Soft Declines and Step-Up Authentication
A score crossing the threshold does not always require a hard decline. For the middle band, step-up authentication can require the cardholder to complete another check, such as an OTP sent to the registered phone number, before approval.
Step-up changes the question from "decline or approve" to whether the user can prove access to the account owner's registered channel. Legitimate cardholders can complete it, while fraudsters with stolen card data generally cannot. In the middle band, false positives become a step-up success rate, which is higher than the existing legitimate approval rate there.
The cost is added friction for legitimate cardholders. OTP delivery depends on a matching phone number on file. It is broadly practical for card-not-present e-commerce, but more disruptive at an in-person point of sale. Use it selectively where friction is acceptable and false positives are most severe.
Threshold Calibration Feedback Loops
Threshold calibration depends on closing the feedback loop. For declined transactions, determine how many were fraudulent and how many were legitimate. Without that evidence, you cannot tell an overly conservative threshold from a properly calibrated one.
Two sources provide this evidence: cardholder disputes for declines and confirmed fraud on approvals from chargeback data. Decline disputes show whether cardholders contacted the bank about an incorrect rejection, a strong false positive signal. Chargebacks on approved transactions reveal the fraud that passed the model.
Link each outcome to its original score range. Chargebacks concentrated between 0.2 and 0.4 indicate underweighted fraud patterns and may call for a lower approval threshold or retraining. Erroneous decline disputes concentrated between 0.5 and 0.7 identify the range where adjustment should cut false positives most effectively.
Manual review sampling of approved transactions can also create labeled training data. Randomly approve a small share of mid-range scores, watch for chargebacks, and use the outcomes for a better-labeled retraining set. Keep the sample small enough that elevated-score fraud does not materially change the overall chargeback rate.
The Tradeoff Cannot Be Removed
No threshold can deliver zero false positives and zero false negatives at once. Legitimate and fraudulent score distributions overlap, and calibration cannot remove that overlap. Fewer false positives in the shared region therefore costs some fraud detection there, and the reverse is also true.
The goal is a better operating point on the precision-recall curve: a lower false positive rate at the same fraud detection rate, achieved by separating the distributions more effectively. Better features, representative training data, a model suited to African transaction patterns, and retraining as behaviour changes improve the curve. Threshold tuning selects a point on it, while model improvement shifts it.
Banks and fintechs using enza's fraud layer can change thresholds in the dashboard without an API update or redeployment. Review false positive metrics on a set cadence, not only after incidents. Monthly or quarterly reviews using segment feedback provide steadier calibration than ad-hoc changes prompted by cardholder complaints.
African Transaction Patterns in Model Calibration
Models trained mainly on North American or European data may perform poorly on African transactions. The issue is not wholly different fraud, but a different legitimate-activity baseline that changes score distributions.
Mobile money top-ups and wallet transfers, common across West and East African markets, may look like unusual velocity to a model trained on Western card data. Informal merchants, cash-heavy economies with high card-present amounts, and uneven cardholder geography also create legitimate signatures unlike those in global fraud-model training data.
Training and calibrating on African transaction patterns lowers false positives for African cardholders at the same fraud detection threshold. An identical nominal threshold can produce different actual false positive rates on African-calibrated and North American-calibrated models. Correct training data and correct thresholding are inseparable.