Fraud Scoring

What Is Fraud Scoring?

Fraud scoring is the process of assigning a numerical risk rating to a payment transaction to estimate the likelihood that it is fraudulent. A fraud scoring model ingests data points about the transaction, the payer, and the surrounding context, then outputs a single score that payment systems use to approve, flag for review, or decline the transaction. The entire evaluation typically completes within milliseconds.

Early fraud detection relied on static rule sets: if the transaction amount exceeded a threshold or originated from a flagged country, the system would block it. Modern fraud scoring replaces these rigid rules with machine learning models trained on millions of historical transactions. These models identify complex patterns that no human-written rule set could capture, such as the combination of a new device, an unusual purchase time, and a shipping address that deviates from established behavior.

Fraud scoring sits at the core of transaction monitoring infrastructure. Every major payment processor, issuing bank, and acquiring institution runs some form of fraud scoring on every transaction it handles.

How It Works

A fraud scoring system processes each transaction through three stages: signal collection, model evaluation, and decision routing.

Signal Collection

The model ingests hundreds of data points per transaction. Common input features include:

• Transaction attributes: amount, currency, merchant category code, payment method, and whether the card is present or not
• Velocity signals: number of transactions in the last hour, day, or week from the same card, device, or IP address (see velocity checks)
• Device data: browser fingerprint, operating system, screen resolution, installed fonts, and whether the device has been seen before
• Geolocation: IP address location compared to billing address, shipping address, and the cardholder's historical transaction geography
• Behavioral biometrics: keystroke dynamics (how fast and in what rhythm the user types), mouse movement patterns, mobile touch pressure, and swipe velocity
• Account history: age of the account, previous chargeback rate, lifetime transaction volume, and whether the email domain is disposable

Model Evaluation

Machine learning models compare these signals against patterns learned from historical fraud and legitimate transactions. Common model architectures include gradient-boosted decision trees, neural networks, and ensemble methods that combine multiple models for higher accuracy.

The output is a single score, typically normalized to a 0-100 scale (where 100 indicates the highest fraud probability) or a 0-to-1 probability. Some proprietary systems use different ranges: FICO's Falcon platform scores on a 1-999 scale across more than 10,000 financial institutions worldwide.

A simplified representation of how input features map to a fraud score:

// Simplified fraud scoring feature vector
{
  "transaction": {
    "amount": 2499.99,
    "currency": "USD",
    "merchant_category": "electronics",
    "card_present": false
  },
  "velocity": {
    "txn_count_1h": 4,
    "txn_count_24h": 12,
    "distinct_merchants_1h": 4
  },
  "device": {
    "fingerprint_seen_before": false,
    "browser": "headless_chrome",
    "vpn_detected": true
  },
  "geo": {
    "ip_country": "NG",
    "billing_country": "US",
    "distance_miles": 5842
  }
}

// Model output
{
  "fraud_score": 94,
  "risk_factors": [
    "new_device",
    "geo_mismatch",
    "high_velocity",
    "vpn_detected"
  ],
  "recommendation": "decline"
}

Decision Routing

The fraud score maps to an action through configurable thresholds. Most payment systems use a three-tier model:

These thresholds are not universal. Each merchant and processor tunes them based on their risk tolerance, industry, and fraud rate. A high-margin digital goods merchant might set an aggressive decline threshold at 60, while a low-margin retailer with thin chargeback tolerance might decline at 40.

False Positives vs. False Negatives

The central tension in fraud scoring is the tradeoff between two types of errors:

• False positives (legitimate transactions declined): lost revenue, customer frustration, and potential permanent customer loss. Studies estimate that false declines cost merchants significantly more than actual fraud losses.
• False negatives (fraudulent transactions approved): chargeback fees, merchandise loss, and potential penalties from card networks if fraud rates exceed acceptable thresholds

Lowering the decline threshold catches more fraud but increases false positives. Raising it reduces friction for legitimate customers but lets more fraud through. AI-powered fraud scoring systems have improved this tradeoff substantially: modern models achieve 90-98% accuracy at major institutions with detection speeds up to 50 times faster than legacy rule-based systems.

US merchants incur approximately $4.61 in total cost for every $1 lost to fraud when accounting for chargeback fees, operational overhead, and lost merchandise. This multiplier effect makes accurate scoring critical to payment economics.

Fraud Scoring in Crypto and Blockchain Payments

Cryptocurrency transactions present fundamentally different challenges for fraud scoring compared to traditional card payments. These differences reshape what fraud scoring means in the blockchain context.

Why Crypto Fraud Scoring Is Different

• Irreversibility: credit card transactions can be reversed through chargebacks, providing a safety net for fraud victims. Confirmed blockchain transactions are immutable, making pre-transaction fraud detection essential rather than optional.
• Pseudonymity: traditional payments are tied to verified identities through KYC/AML processes. Crypto wallets are pseudonymous, and a single user can create unlimited addresses. Privacy-enhancing tools like CoinJoin and mixing services further obscure transaction trails.
• No intermediary dispute resolution: traditional payment networks have built-in dispute resolution through issuer and acquirer banks. Peer-to-peer crypto transfers have no equivalent fallback.
• Cross-chain complexity: funds can move across multiple blockchains through bridges, making it harder to trace the full transaction path compared to a single payment network

Blockchain-Based Risk Scoring

Chain analysis platforms have emerged to fill the fraud scoring gap in crypto. Instead of scoring individual card transactions, these tools score wallet addresses and transaction patterns by analyzing on-chain data:

• Chainalysis: the largest provider with over 1,500 clients, having mapped more than $24 trillion in value across over 1 billion addresses. Its KYT (Know Your Transaction) product provides real-time risk scoring for exchanges and financial institutions.
• Elliptic: covers 50+ blockchains with a focus on DeFi and NFT forensics, providing continuous risk scoring and compliance monitoring
• TRM Labs: emphasizes sanctions screening workflows with explicit confidence levels and reasoning for every attribution

These platforms analyze transaction graphs, identify connections to known illicit addresses (darknet markets, ransomware wallets, sanctioned entities), and assign risk scores that exchanges and on/off-ramp providers use for compliance decisions. For a deeper analysis of how stablecoins address payment fraud challenges, see the research on payment fraud and the stablecoin advantage.

Use Cases

Card-Not-Present Transactions

E-commerce transactions carry the highest fraud rates because the card is not physically present. Fraud scoring is the primary defense layer, evaluating device fingerprints, shipping addresses, and behavioral signals before the transaction reaches the acquirer. Global e-commerce fraud losses reached an estimated $48 billion in 2025, growing 16% year over year.

Exchange and On-Ramp Screening

Cryptocurrency exchanges apply fraud scoring at the fiat-to-crypto boundary. When a user initiates a purchase with a credit card or bank transfer, the exchange scores the transaction for stolen card indicators, account takeover signals, and money laundering patterns. This is where traditional payment fraud scoring intersects with blockchain-specific chain analysis.

Real-Time Payment Networks

Instant payment systems like FedNow, Pix, and Faster Payments require fraud scoring that executes in under 500 milliseconds. Because these payments settle in real time with no chargeback mechanism, pre-authorization fraud scoring is the only line of defense. For more on how fraud prevention works across digital payment types, see the research on fraud prevention in digital payments.

Stablecoin Payments

As stablecoin payment rails mature, fraud scoring adapts to evaluate both the on-chain risk profile of receiving addresses and the off-chain context of the payment initiation. Platforms like Spark that facilitate stablecoin transfers can integrate chain analysis scoring alongside traditional payment fraud models to provide comprehensive risk assessment across payment types.

Risks and Considerations

Model Bias and Discrimination

Fraud scoring models trained on historical data can inherit and amplify existing biases. If past fraud patterns correlate with geographic regions or demographic groups, the model may systematically score legitimate transactions from those groups higher, creating discriminatory outcomes. Regulatory scrutiny of algorithmic fairness in financial services is increasing globally.

Adversarial Adaptation

Fraudsters continuously adapt their techniques to evade scoring models. As models learn to detect specific patterns, attackers shift tactics: using residential proxies instead of data center IPs, spoofing device fingerprints, or distributing transactions across multiple stolen cards to stay below velocity thresholds. This creates an ongoing arms race that requires continuous model retraining.

Privacy Concerns

Effective fraud scoring requires extensive data collection: device fingerprints, behavioral biometrics, location tracking, and transaction history. This creates tension with privacy regulations like GDPR and user expectations around data minimization. The EU's PSD3 directive (expected to take effect between 2026 and 2028) formally recognizes behavioral biometrics as a valid "inherence" factor for strong customer authentication, but also expands requirements around how this data must be handled.

Over-Reliance on Scores

Fraud scores are probabilistic estimates, not deterministic verdicts. Treating a score as an absolute truth leads to poor outcomes. Organizations that combine automated scoring with human review for borderline cases, feedback loops from confirmed fraud and false positives, and layered defenses beyond a single score achieve the best balance between fraud prevention and customer experience.

This glossary entry is for informational purposes only and does not constitute financial or investment advice. Always do your own research before using any protocol or technology.