Elephant

A model built for one problem, trained at the scale that problem demands

Payment fraud is a domain problem. It requires a model that understands not just whether a signal looks unusual, but whether it looks unusual in the context of how fraud actually behaves in payment environments. Elephant was designed around that distinction. Most attempts to improve generic fraud scoring involve adding rules, tightening thresholds, or layering supplemental signals on top of an existing system. Elephant takes a different approach. The differences are structural and begin with how the model was trained.

Precision that reduces the approval and fraud tradeoff

The authorization rate and fraud exposure tradeoff isn't a fixed law. It's a product of scoring imprecision. When scoring more accurately separates legitimate transactions from fraudulent ones, that constraint loosens. Elephant's architecture is designed to reduce the imprecision that makes the tradeoff feel inevitable.

Adaptive retraining as fraud patterns evolve

Elephant is retrained continuously to reflect the evolving fraud patterns of each deployment environment, staying aligned with how fraud actually behaves as conditions change. That alignment reduces the drift that forces compensating rules, broader thresholds, and expanded manual review.

Signal relationships evaluated, not just signal presence

Elephant resolves identity, behavioral, and device signals together, evaluating whether their combination is consistent with legitimate transaction patterns. A signal unremarkable in isolation can carry significant fraud risk in the wrong combination. Elephant's architecture is built to recognize that distinction.

Domain-specific training on payment fraud signals

Elephant was trained on payment fraud signals, not adapted from a broader risk or fraud model. Payment fraud has specific behavioral signatures and signal patterns that general-purpose training data doesn't capture. Domain specificity is what allows Elephant to score with precision where generic models approximate.

Built on the data foundation a model like this requires

A large risk model is only as strong as what it was trained on. Elephant is built on Pipl's two decades of experience making fragmented global data usable in high-stakes decisions. It carries the accumulated signal depth of billions of identity connections, trillions of payment events, and twenty years of infrastructure built to make that data reliable at the point of decision.

That foundation is what allows Elephant to operate with the specificity and confidence that payment fraud decisioning demands.

Trained on over one trillion payment fraud signals

Built across connections between five billion digital identities

Twenty years of global data infrastructure underlying the model

Designed for the signal complexity of real payment environments, not training sets

What makes Elephant architecturally different

Elephant wasn't adapted from a general-purpose model. It was built from the ground up for a single domain. These are the architectural decisions that define it.

Relational signal evaluation

Elephant evaluates how identity, behavioral, and device signals relate to each other, not just whether they're present. Fraud risk is expressed in signal relationships. Elephant's architecture is built around that reality.

Payment fraud pattern specificity

Elephant scores against the patterns that define legitimate and fraudulent transactions in payment contexts specifically. That specificity is what separates it from models optimized for broad risk classification.

Identity graph depth

Elephant resolves identity across five billion digital identities and one trillion signals. That depth of coverage is what allows it to evaluate identity consistency with the precision that payment fraud decisioning requires.

Adaptive calibration

Elephant is retrained continuously to reflect the specific fraud patterns of each deployment environment. Its understanding of fraud stays current rather than degrading against a fixed historical baseline.

How Elephant compares to other models

Our Model

Our Model

Training domain

Payment Fraud

Signal evaluation

Relational — identity, behavioral, and device signals evaluated together

Calibration

Deployment-specific, retrained to each environment

Update frequency

Continuous, no manual intervention required

Domain knowledge

Identity infrastructure built over 20 years, applied exclusively to payment fraud

Objective

Real-time trust scoring for payment fraud decisioning

Other Models

General Purpose

Other Models

Training domain

Broad, multi-domain risk

Signal evaluation

Isolated inputs weighted independently

Calibration

Not deployment-specific

Update frequency

Infrequent, not environment-specific

Domain knowledge

Shallow in payment fraud specifically

Objective

Wide applicability across risk contexts

Other Models

Fraud Classifier

Other Models

Training domain

Fraud broadly, not payment-specific

Signal evaluation

Rule-based or threshold-driven

Calibration

Fixed at point of training

Update frequency

Manual, requires intervention

Domain knowledge

Degrades as fraud patterns evolve

Objective

Binary fraud classification

How Elephant compares to other models

	Our Model Elephant	Other Models General Purpose	Other Models Fraud Classifier
Training domain	Payment Fraud	Broad, multi-domain risk	Fraud broadly, not payment-specific
Signal evaluation	Relational — identity, behavioral, and device signals evaluated together	Isolated inputs weighted independently	Rule-based or threshold-driven
Calibration	Deployment-specific, retrained to each environment	Not deployment-specific	Fixed at point of training
Update frequency	Continuous, no manual intervention required	Infrequent, not environment-specific	Manual, requires intervention
Domain knowledge	Identity infrastructure built over twenty years, applied exclusively to payment fraud	Shallow in payment fraud specifically	Degrades as fraud patterns evolve
Objective	Real-time trust scoring for payment fraud decisioning	Wide applicability across risk contexts	Binary fraud classification

A large risk model for online transactions

A model built for one problem, trained at the scale that problem demands

Precision that reduces the approval and fraud tradeoff

Adaptive retraining as fraud patterns evolve

Signal relationships evaluated, not just signal presence

Domain-specific training on payment fraud signals

Built on the data foundation a model like this requires

What makes Elephant architecturally different

Relational signal evaluation

Payment fraud pattern specificity

Identity graph depth

Adaptive calibration

How Elephant compares to other models

Our Model

Elephant

Training domain

Signal evaluation

Calibration

Update frequency

Domain knowledge

Objective

Other Models

General Purpose

Training domain

Signal evaluation

Calibration

Update frequency

Domain knowledge

Objective

Other Models

Fraud Classifier

Training domain

Signal evaluation

Calibration

Update frequency

Domain knowledge

Objective

How Elephant compares to other models

The large risk model built to perform where others don't