By Jimmy Hennessy, Director of Data Science at ACI Worldwide
Online fraud is a growing problem, causing significant damage to a merchant’s bottom line and often costing the business far more than just the lost goods. In fact, research shows that every pound of eCommerce fraud hits merchants with around three times as much in associated costs (i).
Detecting and preventing fraud is difficult. Merchants are attacked by fraudsters using a number of different methods and guises, and working hard to stay under the radar and evade detection. Spotting attempted fraud in among the millions of transactions being processed by merchants every day, across multiple channels, is a challenging task. Manually reviewing every transaction is impossible and hard coded fraud rules are not flexible enough to keep up with constantly evolving fraud trends.
The data battle
The ability to commit fraud is reliant on data. Whether it is stolen credentials, faked accounts, account takeover – in every case a fraudster uses certain data points or identity markers to successfully make a fraudulent purchase. This data leaves a trail, and merchants have the chance to block fraud by identifying those data trails.
This is where machine learning can turn the tide in the battle against fraudsters. Modern machine learning models can quickly and efficiently analyse vast amounts of data, to identify patterns and trends that are too complex to spot through other means.
Models learn quickly from millions of historical transactions, remember behaviours and learn the difference between genuine and fraudulent transactions. They can then map these learned patterns and use certain attributes to make predictions about new transactions in an automated manner. Used in this way, machine learning can help turn data into intelligence that can augment customer profiles, spot fraud signals and combat emerging fraud threats.
When properly trained and tailored for an individual merchant, machine learning models can help increase fraud detection accuracy by as much as 40-50% (ii) and act rapidly at the point of sale, without the customer noticing any intervention. This not only powers better fraud prevention success, it also helps to minimise false positives and friction at the checkout, both of which can undermine your customer relationships and impact sales.
Maximising machine learning effectiveness
Machine learning is a sophisticated tool and it’s important to understand that merchants can’t just ‘plug and play’ and expect optimal results. You must make sure that the machine learning models you use are geared not only to improve fraud prevention, but also to underpin a friction-free experience for genuine customers.
Machine learning models perform best in addressing these goals when they follow these three key principles:
1. Feed in the right data
For a model to learn and then assess transactions accurately, the data used to train it must be relevant, complete, correct, timely and in mass volumes, based on historical transaction data from your own customer base.
Supplementing this with external data can significantly increase the effectiveness of your machine learning models, by adding the context of fraud trends within and across market segments and geographies. Consortium data, along with a continuous feed of confirmed fraud intelligence, can help increase accuracy and provide early warning of emerging fraud trends.
2. Use the support of experts
Machine learning models need to be built, trained and optimised by data science experts, with input from skilled fraud analysts. Experienced specialists are best equipped to understand the nuances of each merchant’s customer base, as well as having knowledge of other businesses and sectors, to help define the right data requirements and build, train and optimise each model.
3. Continuously monitor performance
Customer and fraud trends constantly evolve, and machine learning models need to be trained, tested and re-trained in line with changing behaviours. You can’t just build them and forget them – models need to be monitored on an ongoing basis to ensure they are performing as expected. It’s important to understand that performance levels will degrade over time and traditional machine learning models will need to be re-trained as new behaviours emerge.
Next generation machine learning
Machine learning is advancing all the time, but a significant development has recently become available with new ‘incremental learning’ models. These models are able to “think for themselves” and adapt to the smallest changes in fraud and spending patterns, as they happen. Because they can adapt to new behaviours without needing to re-learn everything they already know, incremental learning models perform optimally for longer, reducing the need for re-training.
To ensure high performance, new data needs to be fed to the algorithm every 24 hours, so that the model is using the most up-to-date fraud intelligence and customer behaviour data. Because they are continually adjusting in response to that data, incremental learning models can protect merchants more effectively, even as customer preferences and fraudster behaviours evolve.
Fraudsters will continue to adapt their tactics in an effort to evade detection – and incremental learning is an important advance that will help merchants keep up with fraudsters and stop them in their tracks.
The bigger fraud prevention picture
Machine learning is an extremely valuable tool in the fight against fraud, but it isn’t a silver bullet.
Because machine learning models learn from experience, there are sometimes nuances or exceptional circumstances that the model can’t take into account. Machine learning models, standing alone, don’t always offer the flexibility needed to adapt quickly to unusual periods of trading, such as sales peaks, or new product launches. In cases where customer and fraudulent behaviour is changing rapidly, rules can often be flexed more easily to ensure that genuine customers are not mistakenly blocked or fraudsters inadvertently let through.
It’s not just about combining rules and machine learning either. Fraudsters are often smart, sophisticated and organised. They work to anticipate the industry’s next move and to circumvent the controls or predictive measures we use to combat them – always looking for an opening, or the weakest link in the chain. A single tool or layer of fraud prevention is not enough to stop fraud – good fraud prevention requires a solution with multiple dimensions.
Consortium data and shared intelligence, profiling, automated decisioning, human expertise and alerts should all form part of your overall fraud prevention solution, in addition to machine learning models, if you want to stay one step ahead of the fraudster.
Interested to know more? Watch our on-demand webinar: Find a New Ally in the Fight Against Fraud with Machine Learning
By Jimmy Hennessy, Director of Data Science at ACI Worldwide
(i) 2019 True Cost of Fraud, LexisNexis
(ii) ACI use cases 2018