It’s an event that plays out thousands of times across the UK every day. A consumer tries to pay for their weekly grocery shop using a credit card but with the bags packed at the till is unexpectedly told that it has been ‘declined’. The card is well within its credit limit, the PIN number is correct, the consumer has made numerous other purchases in the preceding weeks and yet there is no way around the reality of having no plastic money to spend.
For the financial services industry these ‘false positives’ have become a growing issue. As well as annoying customers and merchants they cost the industry in terms of the manual intervention necessary to authenticate customers and unblock cards.
Every financial services firm will claim it has specialised fraud-detection systems that are good enough to differentiate fraud from legitimate use but there can’t be a card user in the country who hasn’t had their card incorrectly refused on several occasions. No matter how sophisticated these systems say they are, the fact that false positives have become such an embedded issue suggests that something is amiss.
It is into this world of creaking fraud platforms that a Cambridge-based newcomer is pitching a technology called Adaptive Behavioural Analytics that it believes can far more accurately tell which transactions are good and which aren’t. The company is Featurespace and the platform it sells uses the ARIC engine, a machine learning system designed to monitor complex events for anomalies in minute detail.
It’s a field given recent worldwide attention by another Cambridge-founded enterprise, Google’s DeepMind, which also uses machine learning concepts and Bayesian statistical voodoo in its technology. That system advertised itself by playing Go against a human champion but its prowess is rally a window into the way that machines are set to infiltrate a huge array of decision-making systems in the next decade.
The firm’s journey from its foundation as a Cambridge University concept project a decade ago by David Excell (now CTO) and Professor Bill Fitzgerald (who died in April 2014) has been a slow and steady one, as much about proving the ideas as gung-ho business. However, since the firm appointed a new CEO in 2012, Martina King, things appear to have stepped up noticeably, culminating in a £3 million ($4.5 million) finding round in 2014 from a consortium of investors including Cambridge’s famous son Mike Lynch of Autonomy fame.
In conversation, the obviously very busy King (who previously headed Yahoo UK and Capital Radio in a lengthy career) is a likable mixture of jargon-free enthusiasm and thoughtfulness. Engineers can sometimes become wrapped up in technical details but King’s vision for Featurespace has a clarity and simplicity about it.
“When you work on things that are new and different it takes time to bring people along with you and show them that there is a better way of doing things,” says King, who steadfastly believes that what was once an esoteric lab technology - Adaptive Behavioural Analytics – is now being received very differently.
“It’s word of mouth.”
King first came across Featurespace after being introduced by Mike Lynch to co-founder Fitzgerald, of whom she speaks with affection. Her sense of loss – and that of Featurespace – to losing him in 2014 is still palpable.
King is less complimentary about the failures of the fraud detection systems used by banks. “They [financial services] have been losing so much money,” she bemoans. “The attacks are so much more sophisticated that you can’t predict what they will be.”
The problem of false positives, false negatives (i.e. missing real fraud) and customer dissatisfaction have gone from being an inconvenience to a major issue in a handful of years. But why do conventional systems struggle? According to King it is simply that they are based on fraud-detection rules that model known patterns. As soon as something novel comes along, they either break or become expensive to manage. Detecting fraud ends up being a complicated, manual process.
The Adaptive Behavioural Analytics of ARIC models the real world of fraud in far greater detail, understanding anomalies by putting events into context. It’s a theory of understanding the real world that has gained traction in other sectors of computer security which isn’t to say that it’s trivial to pull off.
“It is hard to do and you need real depth of knowledge in statistical profiling and to deliver this on an enterprise scale,” says King. “We have created the market in many respects. [Until recently] there weren’t people that were talking about machine learning.”
Its strength is that all events, including fraud, are executed by or on behalf of human beings. That gives the Bayesian maths underpinning ARIC something to grasp on to in its search not only for understanding the significance of an event but predicting what might happen next.
The challenge for Featurespace is simply to make fraud detection better, cheaper and faster.
“When I wake up in the morning I think how can I shortcut the time?”
She estimates the global cost of coping with false alarms alone could be $6.4 billion (£4 billion), a figure that is a significant percentage of real fraud. It’s as if the solution to fraud is becoming as burdensome as the problem it was meant to be solving.
“We know we can reduce alerts by 70 percent and so you don’t need as many people dealing with so many customers caught up in rules,” says King.
The company has collected a growing list of customers although King is especially proud of the five-year deal it struck in 2015 with mobile payments processor, Zapp, which works in the UK with Barclays. Processing thousands of transactions per second, that proves that Featurespace’s technology scales, an important QED.
ARIC is also being used by a major US bank although King is not able to name the company, as well as a range of other financial institutions for which product secrecy is necessary for competitive reasons.
In the end what will decide the company’s continued growth isn’t simply its customer list but the status of the whole field of anomaly detection itself. With DeepMind now world famous, following a path built by Autonomy a decade ago, Featurespace is building its house on the right plot at the right moment.
In the nick of time, Cambridge has outgrown its reputation as a place for interesting computer science and mathematical cleverness and has established itself as one of the world’s foremost centres of machine learning. Computers will be put to good use, not simply to make the world work in a different way but a better one too. Without that, King could argue, commerce will slowly sink under the hell of its own inability to transact reliably.
“The market has moved in our direction.”