Machines that learn from data have the potential to transform insurance processes and counter fraud by humans
A formal branch of artificial intelligence, machine-learning builds systems that learn directly from the data they are fed and effectively program themselves to analyse that data and make accurate predictions. Having already helped multiple business sectors create new models and drive competitive advantage, now it’s the turn of the insurance industry.
So just how is machine-learning changing the way insurers do business? “It gives insurers three distinct advantages,” explains Max Richter, managing director in Accenture’s UK insurance analytics group. “The first is to mine greater volumes of data, the second to scale analytics across the organisation by working smarter and faster, and lastly by answering more complex questions from ‘will this customer leave me at renewal?’ to ‘what can I do about it?’”
As such it is quickly becoming an essential tool for the insurance sector, specifically enabling companies to yield higher predictive accuracy as it can fit more flexible and complex models. As opposed to traditional statistical methods, machine-learning takes advantage of the power of data analytics and is capable of computing seemingly unrelated datasets whether structured, semi-structured or unstructured.
“Insurers have introduced machine-learning algorithms primarily to handle risk similarity analytics, risk appetite and premium leakage,” says Hortonworks general manager of insurance, Cindy Maike. “It is also widely used to aid the frequency and severity of claims, manage expenses, subrogation (general insurance), litigation and fraud.”
Learn from claims
One of the most impactful machine-learning use-cases is the ability to learn from audits of closed claims. “Claim audits are traditionally a manual process by nature,” says Ms Maike. “Machine-learning techniques provide an uplift in the ability to learn from those by applying enhanced scoring and process methods throughout the claims life cycle.”
As machines learn more about us, they are able to predict our risk profiles at much greater accuracy and granularity than was possible before
Beyond traditional data sources, machine-learning is also opening the analytic doors to new information. “With the advent of telematics, wearables, and other sensor and remote monitoring technologies, more information will be collected about our behaviours and habits, such as how you drive your car or how much exercise you do,” says Heidi O’Leary, principal consultant with Market Gravity.
“As machines learn more about us, they are able to predict our risk profiles at much greater accuracy and granularity than was possible before. This will lead to more personalised insurance products and allow insurers to offer other value-adding products, such as getting the car serviced before it breaks down.”
Another area where wearables and machine-learning meet is on the building site. Ashley Hirst, chief operating officer and chief underwriting officer at AIG, says: “AIG has invested in Human Condition Safety (HCS), an early-stage company matching wearable technology with artificial intelligence.” This is a platform that enables workers to reduce injuries and employers to improve operational efficiency. “HCS’s technology can detect when a worker carries too much weight, makes a bad bend or enters an area that puts them at risk of injury,” says Mr Hirst. All of which helps create new efficiencies in the underwriting process.
So why isn’t everyone using it? Are all insurers aware of, and open to, the impact of machine-learning as a new way of risk-pricing and loss estimation? “Like any new technology this will require significant innovation both from a practical development through to changing organisational culture,” says Rod Bryson, principal, wealth, long-term savings and insurance, with Capgemini.
“For a sector built on underwriters and actuaries, and overall prudent decision-making, such an approach will not happen overnight, but the transformation is almost inevitable in the longer term.” Those who can experiment, at least in the short term, may have the greatest chance to succeed in this arena. “It’s not about building the capability or technology, it’s about how and where in the business do you start to apply it,” says Mr Bryson.
And if it is applied, will the application of machine-learning be able to deliver a measurable decline in human fraud? George Robbins, Europe, Middle East and Africa vice president for commercial solutions, at BAE Systems Applied Intelligence, is in no doubt. “Our customers often detect twice as much fraud using a combination of complex analytical and machine-learning techniques,” he says.
Scott Horwitz, senior director in insurance solutions, at global analytics company FICO, says: “One of our clients doubled the amount of fraud dollars detected and identified 33 per cent more claims in a fraud ring in the first 200 claims they reviewed.” Further, in the first nine days after deploying machine-learning and predictive analytics to try and combat insurance fraud, a large auto insurer found more than £250,000 in suspected fraud, says Mr Horwitz. “The combination of FICO predictive analytics, link analysis and business rules have found up to 50 per cent more fraud over a traditional, rules-based system,” he adds.
APPLYING MACHINE-LEARNING TO FRAUD
Cindy Maike, general manager of insurance at Hortonworks, tells how machine-learning can be applied in the real world of insurance fraud detection:
“Using machine-learning, insurers can load claims data of all types into a huge repository, often called a data lake. This method differs from traditional predictive models, which only leverage structured data. Claims notes, diaries and documents are key in discovering fraud and developing fraud models. In case of fraud detection, the procedure would consist of a learning phase, prediction phase and continuous learning phase.
“The learning phase is when you are learning from training data or claims which are fraudulent and those which are valid. It consists of pre-processing (normalisation, dimension reduction, image processing if you are using photos and so on), learning (supervised and unsupervised) and error analysis (precision, recall, overfitting, test/cross validation).
“The prediction phase uses the model from the learning phase and applies it to new data which is deployed for detecting and flagging fraud.
“The continuous learning phase is when your models are continuously recalibrated with new data and behaviours.
“In addition, using graph analytics with Apache Spark/GraphX is a newer method being leveraged as it enables the usage of neural network and social networks which is key in claims fraud analysis. This method is becoming quite popular compared with traditional claims scoring or business rules as these methods may result in too many false positives and it allows insurance companies to visualise fraud patterns more quickly compared with those traditional scoring models.”