Fraud Detection & Risk Management: Connecting Real-Time Decisions with the Bigger Picture
Explore how ClickHouse and Snowflake can support faster fraud decisions, historical risk analysis, and measurable improvements in financial services.

€4.2 billion. That was the total value of fraudulent payment transactions reported across the European Economic Area in 2024, according to the joint European Banking Authority and European Central Bank report. The report also highlights an evolving threat: fraudsters are increasingly manipulating people into making payments themselves, requiring institutions to adapt their defenses. EBA–ECB payment fraud report.
A customer might log in from a new device, add a beneficiary, and make a transfer. On their own, these actions seem normal. What matters is the order, timing, and how they relate to the account’s history, which could signal a problem.
This creates two main challenges for financial institutions: detecting suspicious activity early enough to respond, and gathering enough context to make the right call. If the system is too lenient, losses can increase. If it’s too strict, real payments might be blocked and customer trust could suffer.
A strong fraud detection system depends on good data quality, up-to-date information, solid risk assessment, and being able to trace each decision. This is where platforms like ClickHouse and Snowflake work well together.
One Transaction, Two Analytical Perspectives
Fraud detection immediately asks: should we act on this event now? Important signals include the current session, recent payment attempts, device changes, and sudden increases in activity.
Managing fraud risk means looking at the bigger picture. Which situations cause the most loss? Where do rules create too many alerts? How does the model perform over time? To answer these, you need transaction history, investigation results, and confirmed fraud cases.
These perspectives should support each other. Historical analysis guides operational decisions, and the outcomes of those decisions help improve the process further.
The same data can also help with AML processes. However, fraud detection and AML have different goals and investigation methods. Sharing a data platform doesn’t mean one model can solve every financial crime issue.
ClickHouse as an Operational Analytics Layer
ClickHouse can handle the transaction and behavior analytics needed for a risk decisioning service. It can track things like how many payment attempts happen in a short time, the total value of transfers to new beneficiaries, or how many accounts are linked to the same device.
Kafka integration lets you stream data as it comes in. Incremental materialized views shift some of the processing to when data is ingested, so later queries can use pre-aggregated data. ClickHouse Kafka integration, incremental materialized views documentation.
Moniepoint shows how this works in real life. Their engineers precompute customer state in ClickHouse and write it to a key-value store for the transaction evaluator. With the same traffic and rules, they cut p90 aggregate lookup latency by 93.1%. This result is from one specific deployment stage. Moniepoint presentation.
Keeping these roles separate matters. ClickHouse gives the analytical signals. A decisioning service uses those signals, adds rules and model results, and then works with the payment system to stop a transaction or send it for more checks.
Snowflake for Historical Analysis and Model Development
It’s much more helpful to assess a payment in the broader context of the customer. This includes their behavior, connections between accounts, and past alert results.
Snowflake’s fraud detection reference architecture includes data integration, feature engineering, transaction scoring, and keeping a record of scores. Both analysts and fraud model teams can use this setup. Snowflake fraud detection reference architecture
Snowflake ML provides capabilities for feature preparation, model training, model version management, and monitoring. Its documentation also covers online feature retrieval and model deployment for inference. Treating Snowflake solely as a historical archive would therefore oversimplify its capabilities. Snowflake ML documentation.
You can also combine both platforms. Whether this is the right choice depends on your environment, required latency, scale, and the cost of running extra components.
How a Suspicious Transfer Could Be Evaluated
For example, a customer might log in from a new device and then make several transfers to unknown beneficiaries.
- Login and payment events enter a shared event stream, with identifiers linking the account, session, and transaction.
- An analytics layer prepares current signals, such as the number of new beneficiaries and the total value of recent transfers.
- A decisioning service combines those signals with the historical profile, business rules, and a model score.
- The system approves the transaction, requests additional authentication, or routes the case for review according to the institution’s policy.
- The decision and eventual investigation outcome become part of the history used to assess system performance.
In this setup, ClickHouse handles current aggregates, while Snowflake supports historical analysis and model development.
The design should also explain how to handle a transaction before it appears in the aggregates. Late events, duplicates, and concurrent payments all matter. Responding quickly with outdated information can still lead to poor decisions.
Measure Beyond Model Performance
Business results depend on balancing fraud prevention with smooth processing of real transactions. Useful metrics include:
- Fraud losses relative to payment volume.
- The proportion of fraud detected and the rate of legitimate transactions incorrectly flagged.
- The number of cases requiring manual investigation.
- End-to-end decision latency, particularly p95 and p99.
- Data freshness and processing cost per transaction.
Evaluation should also consider how things change over time. Testing a model on transactions that happen after the training data is much closer to real-world use. You must leave out any information that wasn’t available when the decision was made, like later fraud confirmations.
From Pilot to Operational Process
A good place to start is with a clear scenario, such as account takeover or a series of suspicious card payments. This helps you decide what data you need, how fast decisions should be made, and what counts as success.
You can start by running the system in shadow mode, where it calculates risk assessments but doesn’t affect payments. You can then compare its results to the current process to see how many investigations it triggers and how many false alarms it might cause.
Before enabling automated actions, you need clear rules for what to do if data is missing or the model isn’t available. Each decision should keep track of the rule and model versions, input signals, and timestamp. This lets teams review what happened if there’s an error and see exactly how the case was handled.
Platform selection should be driven by the requirements. If the existing environment is already meeting the requirements around freshness, performance and cost then there's no real need to add another platform. Where the operational analytics and the historical analysis have very different requirements then the combination of ClickHouse and Snowflake is something that's definitely worth a go.
Planning your next move in Fraud Detection & Risk Management? Get in touch with Cadvel about a data architecture that brings the risk team's requirements right up to the product.
Sources & Further Reading
- ClickHouse: Integrating Kafka with ClickHouse
Documentation covering options for streaming data between Kafka and ClickHouse. - ClickHouse: Incremental Materialized Views
How incremental materialized views shift computation to ingestion time to accelerate analytical queries. - Moniepoint: Precomputing Fraud Signals with ClickHouse Materialized Views
An engineering presentation explaining how precomputed customer state reduced aggregate lookup latency in a fraud detection workflow. - Snowflake: Fraud Detection Reference Architecture
A reference architecture covering data integration, feature engineering, transaction scoring, and historical analysis. - Snowflake ML: Overview
Documentation on feature management, model training, deployment, monitoring, and ML lineage. - EBA & ECB: 2025 Report on Payment Fraud
Joint report examining payment fraud across the European Economic Area from 2022 to 2024, including transaction values, payment methods, and emerging fraud patterns.



