Real-Time Fraud Detection for a Global Payments Provider


Global payments provider

Tech stack

Google Cloud


Fraud detection


AI + Machine Learning

Our client is a global company that provides payments processing technology and services to over 1 billion merchants worldwide and oversees over 100 million transactions per day. Using Machine Learning and MLOps best practices, Datatonic built a next-generation Fraud Detection model, which leveraged the best of Google Cloud technologies and detected fraud at over 90% accuracy, whilst maintaining high-quality user experience.

Our impact

  • Built a model that detects fraud with over 90% accuracy (with <6% false-positive rate), representing ~$100M in fraudulent transactions per year
  • Achieved a serving latency of ~20ms, ensuring that fraud decisions are made quickly, without impacting the customer experience.
  • Enabled the client to work with agility, and expand the scope of how they measure fraud and build specialist models, e.g. using different sets of features depending on geography.


The challenge

For over a decade, our client had been using machine learning models to power their fraud detection solutions. At the time, they were looking to address gaps and opportunities to accelerate their model development and improve their models’ performance in terms of latency and load. The client was seeking to modernise their existing fraud detection solution with the following goals in mind:

  1. Leverage best-in-class Machine Learning capabilities
  2. Enable the team to work agilely and flexibly
  3. Reduce time taken to iterate models
  4. Own the IP


Our solution

To answer the client’s needs, Datatonic developed a proof-of-concept real-time fraud detection solution on Google Cloud Platform.

In order to build the Next-Generation Fraud Detection model that took into account the business’ size, complexity and velocity of transactions, we identified three key areas to tackle:

1. Card presence

The cases of fraud when the card is not present, (e.g. eCommerce transactions) are ever-increasing in the digital world, our team developed an ML solution taking this key factor into account.

2. Stage of the payment request

There are many stages across the life of a payment request (see below), with benefits and disadvantages to embedding a fraud detection engine at each point. In order to pick the right stage, our experts considered the criticality in detecting fraud and the amount of data available at that stage, choosing to focus in the middle section, building an engine between merchant and acquirer.

3. Selecting the right metrics

Fraud detection measures a mix of factors and is a highly imbalanced problem, which can make it a challenging ML problem. Our team wanted to ensure that our model captured fraud (especially with high-value transactions), without compromising the customer experience. The aim was to:

+ Maximise the number of true positives (fraudulent transactions classified as fraudulent), and
+ Keep the number of false positives (non-fraudulent transactions classified as fraudulent) as low as possible.

Using these concepts, our team built a proof-of-concept real-time fraud detection model, which leveraged best practices in MLOps to accelerate model development and improve model performance in terms of latency and load. Our solution improved the performance of the existing fraud model, enabling them to predict fraud with over 90% accuracy while ensuring that fraud decisions are made quickly and without impacting customer experience, with a serving latency of ~20ms.

For more details on our solution and Fraud Detection, watch our on-demand webinar with Google Cloud.