Scaling Network Analysis for Fraud Prevention with BigQuery Graph

Based in the UK, Curve are building a financial super-app, a smart wallet that consolidates all your debit and credit cards into a single app and card, simplifying how millions of users spend, send and save money.

However, operating at this scale means confronting a high-volume, ever-evolving landscape of financial crime. While traditional fraud detection models are excellent at flagging suspicious individual transactions, they often miss the "bigger picture"—the complex networks and hidden relationships that characterize organized fraud rings.

To uncover these connections, we realized we needed to move beyond traditional relational data modeling. By partnering with Google Cloud to implement BigQuery Graph, we’ve been able to conduct deep network analysis at scale, helping us identify hidden fraud networks and achieve significant transaction savings.

The Challenge: The Multi-Hop Problem

Fraudsters rarely operate in isolation. They often share a subset of attributes across multiple accounts—such as a common device, a specific funding card, or shared contact information. In a standard relational database, identifying these links requires complex "multi-hop" analysis.

Attempting to scale this using standard SQL presented two significant hurdles:

• Computational complexity: Uncovering a chain of connections (e.g., User A connects to User B, who connects to User C) requires multiple, massive self-joins. At our volume of millions of users and tens of millions of connections, these queries quickly became computationally expensive and difficult to maintain.

• Data scale: Our most granular signals involve billions of potential connections. Standard relational approaches struggle to process these relationships without hitting performance bottlenecks or exhausting system resources.

The Solution: Native Graph Analytics in the data platform

We transitioned our network analysis to BigQuery Graph to take advantage of its native Graph Query Language (GQL) support. The primary advantage was the ability to stop moving our data and start connecting it directly within our existing environment. We had previously explored other popular graph databases - however, being able to keep our data within our BigQuery existing data warehouse gave us significant time and cost savings compared to having to migrate to a new graph database.

By modeling our payment ecosystem as a property graph—where users are nodes and their shared identifiers are edges—we simplified our architecture significantly. Instead of writing dozens of lines of complex JOIN logic, we can now use intuitive GQL syntax to "match" patterns of suspicious behavior across our entire dataset. This approach allows us to:

• Traverse billions of connections: We can now analyze massive datasets, including user-level, device-level, and card-level connections, with high performance.

• Unify our data experience: Because BigQuery Graph is built into the data platform, we can combine graph traversals with standard SQL analysis, search, and machine learning workflows in a single query. We could therefore leverage our existing SQL pipelines to build the nodes and edges tables, switch to GQL for traversing the graph, and then perform final aggregations with standard SQL. This flexibility makes it accessible to more analysts, without having to upskill in a new language.

Impact and Results

Since integrating BigQuery Graph into our fraud mitigation strategy, the impact on our operational efficiency and bottom line has been profound.

• Financial impact: We estimate that the automated blocks triggered by these graph-based insights have saved Curve ~$12M in transaction losses in 2025 alone.
• Precision and accuracy: Our graph-powered queries have achieved an accuracy of approximately 72% in identifying fraudulent users. This high precision allows our fraud mitigation agents to focus their manual reviews on high-certainty cases rather than chasing false positives.
• Operational speed: Moving to GQL allowed us to streamline our graph queries and refresh our fraud rules more frequently. Previously we were limited to one-hop queries in our hourly rules, but GQL allowed us to optimize these slow-running scripts to stay one step ahead of organized crime.
• From rules to ML: The faster we can traverse the network, the faster we can serve graph-based features to our machine learning models. While rebuilding and traversing the graph on a daily basis is sufficient for training models, it is simply too slow at inference-time when transactions can be authorised in less than a second. GQL is allowing us to move towards micro-batch or streaming traversals to serve fresh data to our fraud monitoring models.

Looking Ahead

Our success with BigQuery Graph has opened new doors for our data science and security teams. We are currently working on fully incorporating our highest-volume signals—including billions of IP address connections—into our real-time detection loops. We are also exploring native graph visualization to give our analysts a more intuitive way to explore and "see" fraud webs as they form.

By treating our data as a living network of relationships rather than just rows in a table, Curve is ensuring that our security remains as efficient and robust as our customer experience.