The traditional approach to risk analysis relies on predetermined rules. An online gambling company might flag up any bets over a certain monetary value, for example, or a bank might block any financial transactions outside the customer’s home country.
This technique is as fallible as the rules themselves. Having one’s debit card blocked on the first day of a holiday is now a common inconvenience of modern life.
An alternative approach is to focus on the probability of certain outcomes, using a method of statistical analysis known as Bayesian inference. Named after 18th century mathematician Thomas Bayes, the technique produces a probability distribution for a given outcome that starts with an assumption of randomness and is gradually refined as relevant data becomes available.
According to William Fitzgerald, professor of applied statistics and signal processing at the University of Cambridge, Bayesian inference lends itself to applications in business because it works with data that has not been collected using the scientific method.
“The Bayesian approach tells you the probability that a hypothesis is correct when you can’t do random trials or reproducible experiments,” he says.
To put this argument into practice, in 2005 Fitzgerald and PhD student David Excell created Featurespace, a software company whose technology applies Bayesian inference to business risk analysis.
The software allows businesses to make educated estimations of risk based on statistically valid analysis, rather than exact predictions based on flawed models, Fitzgerald claims: “Rather than solving the wrong model exactly, we solve the real model approximately.”
Featurespace’s first customer was a large website operator that needed help with its fraud detection analysis. The company had a team of risk analysts using rules, but due to technical constraints they could only look at a subset of the transactions that went through their website.
Using Featurespace’s software, the company now compiles a model of every customer’s likely behaviour based on their online transactions. If a customer interacts with the website in a way that does not match their profile, the Featurespace system alerts the fraud prevention team. Instead of relying on predefined markers for fraud, the company is now using each customer’s historical behaviour as its guide.
An important characteristic of the technology is that once data has been incorporated into the probability distribution, it is no longer needed for future risk analysis. “This allows you to iteratively update your probabilistic beliefs as new data becomes available, but you don’t need to store the data,” says Fitzgerald.
This may prove especially useful in light of the European Commission’s proposed data protection reforms, which would oblige companies to delete customer data on request unless it is essential to their business operations.
Featurespace has won customers including short-term lender Wonga, bookmaker William Hill and spread- betting company IG Index, and in 2010 it raised “seven figures” in a Series A round of funding.
The company has plans for triple- digit growth from 2013, and is looking to adapt its technology for marketing applications such as predicting customer churn. “Say a customer doesn’t come onto your website for a period of time – that might be an indication that you are about to lose them to a competitor,” Fitzgerald says.
An example of the UK’s academic excellence translated into practical technology, Featurespace is a candidate to succeed Autonomy as the Cambridge cluster’s next great software success story. But as Autonomy’s acquisition by Hewlett-Packard shows, even the sturdiest of software companies are prone to assimilation. How long Featurespace can last as a stand-alone company is open to debate.