Finance ops · Production

Monzo uses TensorFlow machine learning to cut card fraud losses to under 0.01% of top-up volume

The problem

Criminals were purchasing stolen card details online and topping up Monzo prepaid cards to convert stolen card numbers into physical spending power; Monzo would then absorb chargeback losses from the genuine cardholders. Prepaid card schemes were especially attractive targets because the conversion from stolen details to a usable card was straightforward.

First attempt

In the early months of the fraud system (May–June 2016), the false positive rate was extremely high — Monzo was banning six genuine users for every three fraudsters caught — and financial losses from fraud reached a high of 0.84% of top-up volume.

Workflow diagram · grounded in source

Top-up request received

trigger

“When a Monzo card is topped up with money from another debit card, Monzo is the acquiring merchant for that transaction”

ML fraud risk scoring

ai_action

“Our fraud prediction model is built using Google's Tensorflow library and analyses a large number of metrics including links between users and behavioural patterns”

3D Secure routing decision

routing

“our fraud engine makes a decision based on how risky it thinks a particular top up is and only puts a small percentage of top ups through 3D Secure”

3D Secure authentication

validation

“the customer will be redirected to the bank that issued the card for further authentication. This typically consists of entering several characters from a password”

False positive monitoring

feedback_loop

“One metric we keep a very close eye on is the number of false positives, where we mistakenly ban a genuine customer”

Reported outcome

Monthly fraud losses fell to under 0.01% of total top-up volume (from a high of 0.84%), the false positive ratio improved from 6 genuine users banned per 3 fraudsters to 1, and the overall fraud rate is now an order of magnitude below the financial services industry average.

Reported metrics

Monthly fraud losses as share of top-up volume (current)less than 0.01%

Monthly fraud losses as share of top-up volume (previous high)0.84%

Fraud rate vs industry averageorder of magnitude lower than the financial services industry average

False positive ratio (before improvement)6 genuine users for every 3 fraudsters

Show all 6 reported metrics

monthly fraud losses as share of top-up volume (current)less than 0.01%

monthly fraud losses as share of top-up volume (previous high)0.84%

fraud rate vs industry averageorder of magnitude lower than the financial services industry average

false positive ratio (before improvement)6 genuine users for every 3 fraudsters

false positive ratio (after improvement)1 genuine user for every 3 fraudsters

fraudulent top-up rate reductionsignificantly reduced our rate of fraudulent top ups

Reported stack

Tensorflow3D Secure

Source

https://monzo.com/blog/2017/02/03/fighting-fraud-with-machine-learning

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Monthly fraud losses fell to under 0.01% of total top-up volume (from a high of 0.84%), the false positive ratio improved from 6 genuine users banned per 3 fraudsters to 1, and the overall fraud rate is now an order o…

What tools did this team use?

Tensorflow, 3D Secure.

What results were reported?

Monthly fraud losses as share of top-up volume (current): less than 0.01%; Monthly fraud losses as share of top-up volume (previous high): 0.84%; Fraud rate vs industry average: order of magnitude lower than the financial services industry average; False positive ratio (before improvement): 6 genuine users for every 3 fraudsters (source-reported, not independently verified).

What failed first in this deployment?

How is this finance ops AI workflow structured?

Top-up request received → ML fraud risk scoring → 3D Secure routing decision → 3D Secure authentication → False positive monitoring.