Catch new attack patterns before the chargebacks.

The attack started three weeks before your chargeback rate moved.

Rules and supervised models only know the fraud you already labelled. A new pattern matches no rule and appears in no training set, so you find out when the disputes settle. Fraud Drift watches the shape of ordinary behaviour per segment and flags the day it changed.

No spam. One email when it is ready to try.

DetectedDevice reuse rate · channel=affiliate-bchangepoint 2026-08-13 · 4 signals corroborate in channel=affiliate-b
1.61%10.0%18.5%changepointJun 29Aug 27
QuietChargeback rate · all channelssame window · no changepoint · disputes have not settled
0.49%0.62%0.76%Jun 29Aug 27

Fifteen behavioural series, one injected bot ring, detection computed live in your browser, including the chargeback rate that has not moved.

The problem

Every detector you own is trained on fraud that already cost you money.

A rules engine encodes attacks somebody wrote down after seeing them. A supervised model learns from transactions somebody labelled after the chargeback landed. Both are backward-looking by construction, and neither is wrong about that. It is just that the gap between a novel pattern starting and either one being able to see it is measured in weeks. The attacker knows this and sizes the run to fit inside it.

120 days
How long a cardholder has to dispute most transactions. The label your model needs to learn a new pattern can arrive four months after the transaction it describes.
Visa and Mastercard chargeback time limits
Zero labels
What changepoint detection on behavioural distributions requires. "This population stopped looking like itself" is computable on day one of the attack.
4 of 4
Independent behavioural signals in the demo run that shift inside the same channel on the same day, while the same signals on every other channel stay flat. One signal moving is a campaign. Four is a population change.

The insight

You cannot recognise a new attack. You can notice a population stopped looking like itself.

Recognition needs a label. Noticing does not. When a scripted ring enters an acquisition channel, it does not just add fraudulent transactions. It changes the distribution of dull, unlabelled signals nobody thinks of as fraud features. Session duration compresses because a script does not read the page. Device reuse rises because emulators are expensive. Time to first transaction collapses because the account exists to transact. Geo entropy shifts because the proxy pool is finite. Those moves are all present on day one, weeks before any dispute settles. This does not replace the supervised model, which is still the thing that adjudicates a transaction. It buys the model the weeks it needs to get labelled.

Method

CUSUM on day-over-day percentage changes plus Bayesian Online Changepoint Detection (Adams & MacKay 2007) per behavioural signal per segment, then Benjamini-Hochberg FDR control across the whole run, then a corroboration rule: no alert unless several independent signals in the same segment changed regime within the same window.

How it works

Four steps, no data science team

01
Point it at events you already emit

Signup, session, and transaction events from your existing pipeline. No labels, no chargeback join, no model retraining. It reads the boring fields you never scored on.

02
It builds behavioural distributions per segment

Session duration, device and fingerprint reuse, time to first transaction, velocity, geo and ASN entropy, funnel timing, sliced by acquisition channel, product, geography, and payment method.

03
Two detectors, a correction, and a corroboration rule

Every signal-by-segment series runs through CUSUM and BOCD. Benjamini-Hochberg then filters across the entire run, because thousands of simultaneous tests at α=0.05 produce alarms by arithmetic alone. What survives still has to be joined by other signals in the same segment before anyone hears about it.

04
One alert that names the segment and the day

Not "risk is elevated". The channel, the date the regime changed, which signals corroborate, and the cohort of accounts to label so your supervised model has training data before the disputes arrive.

Who it is for

The risk lead who explains last quarter to the board

Fintechs and marketplaces past the stage where one analyst can eyeball it: enough volume for distributions to be stable, enough channels for an attacker to pick a soft one, and a supervised stack already in place that keeps finding out late.

Pricing

Pilot
$0
One segment dimension, 60-day history
  • Full detection engine
  • Six behavioural signals
  • Slack alerts
  • Backtest against your last incident
Most common
Team
$1,200/mo
Up to 20M events/month
  • Unlimited signals and segments
  • Corroboration rules
  • Cohort export for labelling
  • Case-management webhooks
  • 12-month history
Scale
$4,500/mo
Unlimited events
  • VPC or on-prem deployment
  • Custom signals
  • SSO and audit log
  • Model-team support on tuning

Competition

What exists, and what it does not do

WhoWhat they doThe gap
SiftSupervised fraud scoring trained on a network-wide labelled dataset.Strong on patterns the network has already seen and charged back. A pattern with no labels anywhere is not in the training data by definition.
SardineDevice intelligence and behavioural biometrics scored per session.Scores individuals against a threshold. A ring that keeps every session inside normal bounds is invisible per-session and obvious in aggregate. Different question, different math.
Unit21No-code rules and case management for risk operations.Writing the rule is the step after noticing. Noticing is the part that takes weeks, and it is the part nothing in the category automates.
Arize / FiddlerML monitoring with feature-drift detection on deployed models.Built to tell a data scientist when to retrain, not to alert an on-call risk team. Threshold-based, no multiplicity control across thousands of segment tests, no corroboration rule. It would flood the channel.
How this fails

The honest weakness: unsupervised drift tells you a distribution changed, not that fraud caused it. A marketing push into a channel changes session duration and time to first transaction too, and so does a checkout redesign. If early alerts are mostly campaigns, the risk team learns to ignore them inside a month and the product is dead. A fraud tool that cries wolf is worse than no tool. The mitigation is that precision is the roadmap rather than a feature: FDR control across every test in the run, and a hard corroboration requirement that several independent signals in the same segment moved together, since a campaign usually moves one or two and a scripted ring moves the whole behavioural profile. That will still miss slow, human-operated fraud that never changes a distribution. This is a detector for scaled attacks, and we should say so.

Market

Sold against the cost of one undetected ring, not as a line item

Mid-market fintechs and marketplaces already spend $50k to $400k a year on fraud vendors. This is an early-warning layer beside that spend, not a replacement for it. Two thousand accounts at the Team tier is $29M ARR, and a single ring caught three weeks earlier usually pays the year.

Request a backtest

No spam. One email when it is ready to try.

Or just go look at the demo first →