FRAUD & RISK
Fraud detection: find what deserves a closer look
Fraud detection with TIMi estimates, for each transaction or case, the probability that it is fraudulent, and shows the factors behind the alert.
Suspicion score
Each transaction or case gets a probability, not a simple yes or no.
Rare patterns
Modeler works on rare events, as long as the history is representative.
Explained per alert
Each alert comes with the factors that triggered it.
HOW IT WORKS
How fraud detection works.
Three steps, in one engine, on a standard server with no cluster.
1. Prepare the data
Anatella gathers the transaction history, the context (channel, amount, frequency, location) and the cases already confirmed as fraud, then structures them into variables.
2. Build the model
Modeler compares the profile of known fraudulent transactions with current ones and estimates a probability of fraud, including on rare or infrequent patterns.
3. Score and explain
Each transaction or case gets a suspicion score with the factors behind it, such as a gap with usual behavior or an unusual mix of amount and channel.
Anatella
ModelerGETTING IT RIGHT
Choices that shape a fraud model.
A fraud model works alongside your existing controls, not instead of them.
Combine rules and model
Business rules capture known, documented cases. The model adds the ability to spot subtler or evolving ones. The two work together.
Be realistic about new schemes
A model trained on known cases first detects patterns close to those already seen. Completely new schemes are harder, and are best covered by adding general anomaly detection.
Decide the pace
Scores can be produced in near real time or in batches, depending on the deployment architecture and your operational needs.
EXPLAINABILITY
An alert says look here. The factors say where.
A suspicion score alone says a transaction should be examined, not why. A white-box model shows the factors behind each alert, so investigators check a precise lead instead of starting from a signal with no context.
For fraud teams
Investigation time per alert goes down, because the triggering factor is already identified.
For transparency requirements
A model that justifies each result variable by variable gives you the documentation to answer transparency expectations such as GDPR and the AI Act. The compliance assessment itself stays with your legal team.
For model owners
Fraud schemes evolve. Modeler’s built-in Drift Report shows when a model loses accuracy, so you update it when needed instead of on a fixed calendar.
DATA
What data a fraud model uses.
Anatella connects to more than 80 sources, so this data can be gathered without prior custom development.
Transaction history
Amount, frequency, channel, timestamp.
Context
Location, device, account history.
Confirmed fraud cases
The learning base.
Existing business rules
Integrated alongside the predictive model.
PROOF
Fraud models in production.
We do not yet have a published customer story dedicated to generic transaction fraud. These two deployments are the closest references.
Recovery costs cut by half
A predictive model flagged suspicious behavior from the first month after a subscription was activated. Net gain: €1.7 million.
150 billion transactions a month on one server
Anatella processes the monthly transactions of a national telecom network, including fraud control, on a standard server in 5 to 7 hours.
FAQ
Common questions.
Can the model detect fraud schemes it has never seen?
A model trained on known cases first detects patterns close to those already observed. Detecting completely new schemes is more uncertain, and benefits from general anomaly detection used alongside.
How many confirmed fraud cases are needed?
There is no universal threshold, but very few cases make learning more fragile. Modeler can work with rare events, as long as the available history is representative.
Does the model replace our business rules?
No, it complements them. Rules capture known and documented cases, the model adds the ability to detect subtler or evolving ones.
How fast is a score available after a transaction?
Scores can be produced in near real time or in batches, depending on the deployment architecture and operational needs.
RELATED USE CASES
Keep exploring.
Claims Fraud Detection
Flag insurance claims likely to be fraudulent or exaggerated, with the reasons.
Anti-Money Laundering Pattern Detection
Spot unusual transaction structures between accounts, with the factors behind each alert.
Tax Fraud Detection
Prioritize the returns most likely to contain irregularities, with a documented reason.
SEE FOR YOURSELF
Test it on your own data.
Run TIMi on your own data with a free trial, or talk to us about your case.
