It is fair to say that there has been a lot of hype, buzz – and probably also confusion – with regards to machine learning. When considering machine learning in the context of fraud prevention, it should not imply a sort of existential threat to humanity, which some claim artificial intelligence (AI) could pose.
Machine learning, to be fair, is a subset of AI – but it is concerned with killing fraud, not killer robots.Machine learning is a discipline within computer science concerned with the discovery of patterns in data via algorithms that can learn from and make predictions based on data. It is also referred to as predictive analytics or predictive modelling – and it can deliver tangible results in fraud detection and prevention.
>See also: Is business data AI compatible?
Algorithms are used to build predictive models, which assess the probability of a given transaction being fraudulent. This is done by combining historical transaction data from fraudulent activity with information from genuine customer transactions. It is the ability of machine learning algorithms to extract meaning from complicated data that means they can be used to identify patterns and highlight trends that are simply too complex (or require too much data to be crunched) to be noticed by human fraud analysts or through other traditional fraud detection techniques.
By running specific algorithms, and using them to make automated decisions or generate alerts for suspicious activity, it is possible to save manual review time, reduce the number of false positives, and quickly stop fraud attempts.
Predictive models learn by example, so it is imperative to ‘feed’ them the most valuable, complete, and relevant data possible, according to ACI Worldwide’s paper ‘Driving Up Conversion with Effective Fraud Management.’ Access to merchant, payment provider and FI data across sector and geographies helps to enhance and train machine learning models, resulting in improved accuracy.
Within the retail fraud sector, sector models (based on data from multiple merchants) and tactical models (developed for a specific segment, data strata, or merchant) are combined with sophisticated rules engines that allow fraud strategies to be tailored to individual needs. A rules engine should be flexible, not only to adapt to emerging fraud trends, but also to respond to business changes so that machine learning models do not constantly need to be ‘retrained.’
According to ACI Worldwide‘s paper, the successful deployment of machine learning models within the payments and fraud space enhances the work that fraud analysts are already doing, and is not a replacement. It is becoming imperative for fraud departments to be built on cross-functional teams, so that the relevance and performance of machine learning models are continually refined to remain effective.
The justification for this investment – both time and resources – is that fraud prevention solutions are no longer seen as purely a roadblock to fraud; they are a way of improving checkout conversion rates and thus boosting the bottom line. But the right balance – and being able to build an effective fraud filter – requires merchants to first see fraud to understand what their fraud looks like, and to recognise how fraud differs from genuine transactions.
Data, as well as feeding into machine learning models, can help merchants to build a picture of both fraudster and customer behaviour, which helps to establish the context for fraud management. Payment providers and merchants need systems capable of capturing and collating data, so that it can be analysed for trends, even as those trends are still emerging or evolving. The best data solutions dig deep into every transaction, gathering intelligence from every conceivable data point.
The data that one merchant gathers can be used to sketch out patterns within their customer base, but is also critical to building rich intelligence and a good understanding of emerging fraud trends within – and across – market segments and geographies.
The richer and more intelligent data becomes, the more accurate, effective and efficient fraud detection strategies become. This in turn reduces the impact on genuine customers and increases checkout conversion rates.
By more accurately pinpointing fraud, machine learning models also help to support better conversion rates, by reducing false positives and ensuring genuine customers do not get unnecessarily declined or delayed by manual review processes.
But the retail sector needs to bear in mind that machine learning is not a magic bullet; it is not a replacement for looking strategically at how to deploy the right fraud prevention tools and processes. Merchants need to select the technology and techniques that are best suited to their market segment, business, geography and customer base.
>See also: Why 2017 will be the year of automation
And critically, because things – and people – change, it is important that fraud management is adaptive. It needs to be adjusted for temporary periods of acceleration, such as peak trading periods, shopping holidays, and promotions, as well as for longer term shifts in customer behaviour and lane changes such as new market entries.
This requires a level of flexibility for which a uniform “one size fits all” solution is not applicable, and for which any machine learning capabilities need to be complemented by human fraud consultants and traditional fraud techniques.