Service 04 — Predictive Modeling & Fraud Detection
Predict what's next. Stop what's fraud.
Keap Technologies builds predictive models that forecast demand, revenue and risk, and machine learning fraud detection that catches more fraud with fewer false alarms. For a US fintech, our graph-based models cut false positives by 75% while catching 95% of fraudulent activity. We serve payments, lending, brokerage and e-commerce teams in Cyprus, EMEA and the US.
Updated
- fewer false positives for a US fintech
- 75%
- Keap Technologies client
- of fraudulent activity caught
- 95%
- Keap Technologies client
- payment fraud in the EEA in 2024
- €4.2bn
- EBA & ECB, Dec 2025
- of credit-transfer fraud value came from scams
- 74%
- EBA & ECB, 2024 data
What we build
01
Transaction fraud detection
Real-time risk scores for card, transfer and wallet transactions, tuned to your risk appetite and analyst capacity.
02
Graph & network analytics
Link accounts, devices, beneficiaries and merchants to expose mule networks and organized fraud rings.
03
AML transaction monitoring
Machine learning that prioritizes alerts and finds suspicious patterns that static thresholds miss.
04
Credit risk & scoring
Explainable probability-of-default and affordability models, documented for model-risk and AI Act review.
05
Demand & revenue forecasting
Forecasts with honest uncertainty ranges for planning, inventory, liquidity and capacity.
06
Churn & lifetime value
Identify at-risk customers early and focus retention spend where it pays back.
What is predictive modeling?
Predictive modeling uses statistical and machine learning techniques to estimate an unknown or future outcome — a demand figure, a default, a cancellation, a fraudulent transaction — from patterns in historical data. A good predictive model comes with a measure of its own uncertainty and is judged by how much better it decides than the current process.
How does machine learning fraud detection work?
Machine learning fraud detection scores each transaction or event by how closely it resembles known fraud and how far it departs from normal behavior. Production systems usually combine three layers:
- Supervised models trained on confirmed fraud and legitimate activity, using features such as transaction velocity, amount deviation, device and location changes, and beneficiary novelty.
- Anomaly detection that flags behavior unlike anything seen before — essential because fraud patterns change faster than labels arrive.
- Graph analytics that score the relationships between accounts, devices and counterparties, exposing coordinated fraud that looks innocent one transaction at a time.
Scores feed a decision layer — approve, step-up authentication, hold or review — and analyst outcomes flow back as new labels, so the system improves every week.
Why graph analytics catches what rules miss
Rules look at one transaction at a time. Organized fraud works across many accounts at once: mule networks, synthetic identities and collusive merchants each look ordinary in isolation. Graph databases make those connections queryable, and graph-based features make them learnable.
“Very professional work, they used advanced data science techniques and graph databases to transform our fraud prevention. Their sophisticated models identified complex fraud patterns across networks, reducing false positives by 75% while catching 95% of fraudulent activities.”
Payment fraud in Europe, in numbers
The latest joint report from the European Banking Authority (EBA) and the European Central Bank (ECB), published in December 2025, shows fraud shifting from stolen credentials towards scams that manipulate customers into paying:
| Measure | Value |
|---|---|
| Total value of payment fraud in the EEA | €4.2bn in 2024 (€3.5bn in 2023) |
| Credit-transfer fraud | €2.5bn |
| Card fraud | €1.3bn |
| Share of credit-transfer fraud caused by scams (payer manipulation) | 74% by value (65% in 2023) |
| Share of credit-transfer fraud losses borne by customers | About 85% |
| Card fraud that is cross-border | About 70% |
Scams are authorized by the victim, so they pass strong customer authentication. Detecting them depends on behavioral and network signals — exactly where machine learning outperforms static rules. Read our fraud detection playbook for payment and e-money institutions.
Regulation-ready risk models
Under Annex III of the EU AI Act, AI used to evaluate the creditworthiness of individuals or set their credit score is high-risk — "with the exception of AI systems used for the purpose of detecting financial fraud." Following the Digital Omnibus (Regulation (EU) 2026/1744), those high-risk obligations apply from 2 December 2027.
Keap Technologies builds credit models with reason codes, documentation and human-oversight steps that support AI Act and model-risk review, and builds fraud models that remain explainable to analysts and auditors even where the AI Act's high-risk regime does not apply.
Frequently asked questions
What's the difference between rules-based and machine learning fraud detection?
Rules flag transactions that cross fixed thresholds, which is transparent but generates many false positives and misses new patterns. Machine learning weighs hundreds of signals together and adapts as fraud changes. The strongest systems combine both: rules for hard policy limits, models for risk scoring and prioritization.
Is fraud detection AI high-risk under the EU AI Act?
No, not under the credit-scoring category. Annex III point 5(b) of the EU AI Act makes creditworthiness assessment and credit scoring of individuals high-risk, but explicitly excludes AI systems used to detect financial fraud. GDPR and AML obligations still apply.
How do you reduce false positives without missing fraud?
By adding behavioral and network features that separate genuine unusual activity from fraud, calibrating scores to your analysts' capacity, and measuring precision and recall on recent, held-out data. For a US fintech this approach cut false positives by 75% while catching 95% of fraudulent activity.
Can you work alongside our existing transaction monitoring system?
Yes. Most projects add a machine learning scoring layer next to existing rules and case-management tools, so analysts keep their workflow while alerts become fewer and better prioritized.
Keep exploring
Let's build what's next.
Tell us about your data and the decision you want to improve. We'll come back with a clear plan.