> ## Documentation Index
> Fetch the complete documentation index at: https://docs.fingerprint.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Fraud Score Overview

Fraud Score is a machine learning platform that trains a custom fraud detection model for your business, using your own confirmed fraud outcomes combined with Fingerprint device intelligence.

Unlike a generic risk score, Fraud Score learns what fraud looks like in your traffic specifically. A model trained on your chargebacks is not the same as a model trained on your competitor's chargebacks.

## How it works

Fraud Score follows a continuous loop:

1. **You submit labels** - confirmed fraud and non-fraud outcomes attached to specific events
2. **We build a training dataset** - validated, balanced, and versioned
3. **A model is trained** - one model per fraud type, per client type
4. **The model is reviewed** - by Fingerprint, then deployed manually or automatically depending on your settings
5. **Predictions appear in your API response**
6. **The model retrains automatically when enough new labels arrive**

## Supported fraud types

| Category | Examples |
| :- | :- |
| `payment_fraud` | Chargebacks, stolen card use |
| `account_takeover` | Credential stuffing, session hijacking |
| `fake_account` | Synthetic identity, duplicate registration |
| `promotion_abuse` | Bonus farming, referral fraud |
| `multi_accounting` | Ban evasion, duplicate accounts |
| `platform_abuse` | API abuse, toll fraud |
| `content_abuse` | Spam, fake reviews |
| `other_fraud` | Custom use cases |

## API response

Predictions appear in the `labels` array inside your existing identification response.

```json theme={"theme":"github-dark-dimmed"}
{
  "labels": [
    {
      "label": "chargebacks",
      "category": "payment_fraud",
      "prediction": true,
      "ml_score": 0.872,
      "threshold": 0.650,
      "version": "4"
    }
  ]
}
```

One entry appears per active model. A customer with payment fraud and account takeover models on Web and iOS will see up to four entries.

| Field | Description |
| :- | :- |
| `label` | Your label name as submitted |
| `category` | Fraud type from Fingerprint taxonomy |
| `prediction` | `true` if the score exceeds the threshold |
| `ml_score` | Model confidence score (0-1) |
| `threshold` | Configured cutoff for this model |
| `version` | Active model version |

## Prerequisites

* Fingerprint Smart Signals must be active on your subscription.
* You need at least 1,000 confirmed fraud outcomes per fraud type per client type to train your first model. For best results, aim for 10 confirmed non-fraud events for every fraud event.
* Fraud Score is a paid add-on. Contact your account team to enable it.