It is no misnomer to say that every business is becoming data-driven – but of course it is what you do with it that counts. To get a truly effective data science team firing on all aces, your organisation needs to be able to offer various comparisons, gauging the ‘insights within the insights.’
The insurance sector is therefore one which is tailor-made for this mission. Last April, Raconteur summed up this rising trend. “What is new is that data has grown in volume, quality and accessibility, and there is now the ability to combine and analyse multiple data sources – which is giving insurers plenty to think about,” wrote Sooraj Shah. “How quickly companies can adapt and ensure data science is a part of their organisation will determine how competitive they’re likely to be in the years to come.”
Pardeep Bassi (left) has been head of data science at LV= since June 2017, having spent all of his career in insurance in some capacity, firstly at AXA and then at Domestic and General. It makes for an interesting analysis to compare where the space has come in just a handful of years.
“Back then, it was called the innovation team – data science wasn’t a term widely used,” says Bassi of starting at AXA in 2012. “We were a statistically focused team looking at new technology and better ways to predict outcomes, predicting models using open source technology. The change which has happened now… [it] has become much more mainstream.”
“We’re at a stage now where we are scaling it up significantly and are encountering a lot more difficulties in terms of things we hadn’t known about last year”
LV= utilises this expertise across its general insurance arms, focusing on car, home, pet and travel insurance. Each of these use cases has differences, but the underlying template to build each of these models is reusable. One particular example was where LV= General Insurance worked with Microsoft to create a scalable machine learning solution to more easily solve the 20% of car insurance claims where liability is a grey area and claims can take up to 12 months to resolve.
Bassi notes that following the rollout, the company’s Net Promotor Score (NPS) had gone up while the average time to settle a claim had dropped ‘significantly’. “Not only are we making our processes more efficient from an operational perspective, we’re helping our customer experience,” he says.
This is not to say that everything is plain sailing, however. As with many companies, Bassi explains, the challenge is one of overhauling legacy systems. “From a technical point of view, we can build pretty much any algorithm we want to at the moment using open source technology – we have the right compute power, [and] because we’re in the cloud we’re able to spin up however much compute we need,” he says. “Where [we’ve] got legacy systems, [it’s] not only how do you build these models to start off with, but how do you actually integrate them back into existing systems to make sure they’re actually used and having an influence on the business?
“We’ve overcome this by having not only data scientists in the team but data engineers,” adds Bassi. “It’s their role to productionise and implement models – so they’ve got to understand existing systems and where it plugs in. A lot of the models we build require a real-time link to systems… so it’s understanding if the system is fit for purpose, can it accept API calls, does the latency have to be within a certain time window?”
Overall however, Bassi says that in the journey from first efforts to everything becoming machine learning-led, his team is ‘somewhere in the middle.’ Bassi spoke at last year’s AI & Big Data Expo where he said LV= was ‘implementing a couple of solutions’ – but he is back this year and is interested to not only explain his journey, but explore that of the audience too.
“A language is only a tool – what’s more important for a data scientist is having the core theoretical understanding of how a model works”
“Considering our own journey, there are various different stages of applying machine learning,” says Bassi. “The first stage is – can you get something live? The second stage is can you scale up, and how many different decisions can you affect? And the final stage is [where] everything is machine learning-led.
“We’re at a stage now where we are scaling it up significantly and are encountering a lot more difficulties in terms of things which we hadn’t known about last year,” he adds. “I think the rest of the industry will follow that same suit – so there’s going to be a number of people who were thinking about it last year who have now started doing it, and they can hopefully learn from the experiences that we’ve had.”
Much like Ben Dias, head of data science at Royal Mail, Bassi – who also holds a strong mathematical background – prioritises principles over the practical per se when looking for talent. “A language is a tool… I think what’s more important is having the core theoretical understanding of how a model works,” says Bassi. “Do you have that intuitive understanding right from first principles what a machine learning model is doing? Without that, you will really struggle to be a true data scientist and really know the limitations of certain algorithms.”