Скоринг фрода
Fraud Scoring
Fraud scoring assigns a transaction, payment or player a numerical fraud-risk score based on rules and models. Based on score thresholds, the system approves, sends for manual review, or declines the operation.
What is Скоринг фрода (Fraud Scoring)?
## How it works
Each operation (deposit, withdrawal, bet, registration, card payment) receives a risk score, usually on a 0–100 or 0–1 scale. The score combines rules (for example, "if the card BIN is from country A and the IP is from country B, add weight") with the probability produced by an ML model trained on chargeback history and confirmed fraud. Thresholds define three zones: approve, send to manual review, decline.
## Signals used
In payments: card BIN and issuer country, mismatch between IP and billing country, device and browser data, speed of repeated attempts, age of email and phone, card history within your own network and in consortiums. In gambling and betting: multiple accounts linked to one device or payment instrument, withdrawals right after a bonus, atypical betting patterns (arbitrage, surebets, suspected manipulation), logins via proxy or VPN.
## Impact
Strict thresholds reduce chargebacks and fraud losses, but false declines rise: legitimate customers leave and approval rate falls. Weak thresholds have the opposite effect. The balance is tuned using metrics: decline rate, chargeback rate by channel, and the cost of manual review. The score also affects the ability to work with a processor and the terms of the reserve.
## Common mistakes
Using one general score for all geos and payment methods without segmentation. Not updating the model after a change in traffic or a new payment method. Assuming fraud scoring replaces KYC and AML: these are separate processes, and scoring does not cover them. Looking only at the decline rate without checking it against actual fraud and friendly fraud.
FAQ
How does fraud scoring differ from KYC?
KYC verifies a customer's identity using documents, while fraud scoring assesses the risk of a specific operation based on behavior and signals. Scoring runs in real time, whereas KYC is usually one-off or periodic.
Can scoring decisions be fully automated?
Yes, for simple cases with very low or very high scores. Borderline scores are usually sent to manual review, otherwise false declines and chargebacks increase.
Why does approval rate drop after enabling fraud scoring?
Thresholds may be too strict, or the rules may not account for legitimate traffic characteristics. You need to analyze false positives and adjust weights by segment.