By constantly serving our algorithms with information about their performance, we are able to hone them so they remain effective in the ever-evolving financial crime landscape, writes Jérôme Bovay

NetGuardians is better at preventing financial crime than other fraud fighters because our software constantly learns from what it does.

Its outputs – alerts grading the degree of risk in a bank transaction – are reviewed for accuracy. These reviews are fed back to the algorithms and this feedback loop teaches our software to spot new alerts fast, giving the best possible protection to banks in the ever-evolving fraud landscape.

A mixed diet of data is key

Machine-learning algorithms are intelligent in that they learn from the data they are fed. If the algorithm is fed only once, it will learn from that data and make predictions based on that alone. If, on the other hand, the algorithm is continuously fed information about how accurate it has been, it will be able to learn from its experience, becoming more refined and more accurate over time and therefore able to make better predictions.

At NetGuardians, we constantly feed our algorithms information from banks about their performance. Feedback is twofold: is the alert a real fraud case or, if not, was it sufficiently suspicious to generate further investigation by the analyst, such as a callback to the customer? The more feedback we receive, the more we can train our algorithms, meaning fewer false alerts and better protection for the bank and its customers.

This constant learnt improvement via performance review is vital because while fraud is a global problem and can happen to anyone, it is in fact relatively rare. A bank may see only one fraud in 100,000 transactions. At this frequency, it is hard for an algorithm to learn to spot different fraud types. However, it can learn about fraud by learning what is suspicious yet still genuine. In this way, a perfect diet for a fraud algorithm consists of data about transactions that are highly suspicious, a bit suspicious, and genuine, all mixed up together and delivered frequently.

The exact recipe is NetGuardians’ unique selling point. We use active learning techniques, mixed with an efficient set of classification algorithms. As a chef, we use qualitative ingredients, but their combination makes the difference!

Such a diet allows the algorithm to learn fast and become more effective. The feedback on its output creates a virtuous circle. Without it, the algorithm can quickly become out of date.

Minimal input from banking staff

From the banks’ perspective, giving feedback couldn’t be simpler. The process is automated – bank staff need just review a dashboard and press a button to let us know the accuracy of any alert.

The feedback loop using Active Learning has always been a central part of NetGuardians’ modus operandi. Together with our anomaly-detection software and our library of known frauds, it allows us to offer banks highly accurate fraud and financial crime protection from day one. Indeed, performance tests show that our software finds more fraud while generating fewer false alerts. This helps eliminate up to 99 percent of all payment risk. It also generates up to 85 percent fewer false alerts, reducing operational costs by up to 75 percent.

As you would expect, our algorithms are improving and adapting all the time – thanks to the feedback loop. It’s a win-win for everyone – apart from the fraudsters.

Jérôme Bovay is Lead Data Scientist at NetGuardians.