Case Study: Global Marketplace

More trusted users and fewer losses at scale

This global marketplace was facing a difficult tradeoff: raise defenses and risk turning away real customers, or loosen controls and invite costly fraud. Elephant helped them break the tradeoff, improving approval rates, cutting chargebacks by more than two-thirds, and reducing operational burden in the process.

The challenge

This marketplace processes over a million transactions monthly, with a fraud team that was deeply familiar with behavioral patterns, rules-based logic, and risk thresholds. But despite their maturity, outcomes weren’t improving fast enough.

Chargeback rates remained stubbornly high The review queue consumed too much analyst time. And the path to higher approval rates felt closed off, blocked by the fear of letting more fraud through. They needed a new way to recognize when a user should be trusted, not just when one shouldn’t.

Their goals were clear:

Improve approval rates without increasing chargebacks

Reduce manual review volume and cost

Drive down chargeback rates

Unlock new approvals from trustworthy users previously flagged as risky

Where static systems broke down

Too many safe transactions stalled

Even real customers were getting routed to review or declined, damaging conversion and trust

Manual reviews were costly and slow

Almost 10% of their transactions required manual review, creating delays from fraud analyst bandwidth

Chargebacks were high despite rules

Even their strictest rules couldn't reduce their chargebacks rates to below 12%

Signals weren't informing strategy

Identity indicators like email and phone strength were available, but weren't integrated into scoring logic

How the system changed with Elephant

Elephant introduced trust scores calibrated to the marketplace's own data, using a blend of labeled and unlabeled transactions, enriched identity signals, and outcome-driven scoring thresholds. Rather than focusing solely on what looked risky, the new system prioritized what looked credible.

What we implemented:

Historical approval and fraud outcomes were mapped to identity-based signals
Email, phone, IP, and address signals were scored for both risk and consistency
Connection strength between user attributes was used to validate identity
Output thresholds aligned to business targets for fraud, approval, and review rates

Proof of signal intelligence:

Emails found in Elephant's identity graph were 59% less likely to be fraudulent
Phone and address pairs matched to a known identity showed 42% lower fraud rates
Email + phone matched to the same identity resulted in 3.2x less fraud
Transactions with 3+ strong identity signals had 80% lower fraud rates

The real win? Confidence throughout the funnel

Once Elephant's Trust Score was deployed, the impact was felt across the entire system. Approvals moved faster, review queues shrank, and fraud dropped, all while preserving the control teams needed to stay confident.

Chargeback rate dropped from 12% to 4%

Adaptive scoring identified credible users earlier, reducing loss from risky transactions

Manual review rate reduced from 9% to 3.5%

Signal-based separation made borderline cases cleaner, easing analyst load and cost

19.44pp increase in approval rates at same fraud threshold

Calibrated trust scores allowed the team to approve more good users without increasing exposure

ROC-AUC lifted from 0.75 to 0.83, PR-AUC doubled to 0.18

Model improvements delivered sharper signal separation and better fraud protection, even in low-prevalence environments