Data Analytics in life insurance: lessons from predictive underwriting
26 May 2014
All major companies are thinking about what they can do with 'big data.' Swiss Re has been running data analytics projects with its life insurance clients for the last twelve years. Looking back on his journey, William Trump shares five lessons learned in dealing with data and building predictive models.
In 2002, Swiss Re built its first predictive underwriting model for life insurance. We applied statistical techniques to past data – in this case banking data and life insurance data – to find new health 'predictors.' These were then used to pre-select a group of customers to target an offer of life insurance with minimal underwriting.
Swiss Re started using predictive techniques with the simple conviction that we should be using data intelligently in the pursuit of an improved customer experience. Our starting place was data-rich organisations. A breakthrough was finding such organisations where underwriting data was also available. We were emulating what banks and some general insurers were doing with their data. These companies used what they already know about customers (eg data on past purchases) to derive a competitive advantage – for example by pre-approving customers for specific products, or by targeting certain messages to customers at different times.
Swiss Re has recently developed its 14th predictive underwriting model – most of which have been built in the past 5 years, and most of them for bancassurers. As with any pioneering work, we have learned many lessons along the way and our processes have changed accordingly. Many of our findings can only be shared with our clients; however, there are five broad guidelines we have developed in building our models.
1. Start with the end in mind
This may sound obvious, but there are bad examples within and outside the financial services of people 'doing big data' without a clear business goal in mind. We start by asking the question: what are you seeking to achieve?
There are many outcomes life insurers may be seeking to achieve with their data. These include: reducing the underwriting process for healthy customers; improving retentions; increasing responses to marketing activities; or differentiating on price between customer groups. Targeting clear goals is crucial in order to acquire the necessary data and run the appropriate analytics.
2. Be realistic about limitations
There have been some far-fetched promises about what big data can achieve; but the effectiveness of data is frequently constrained. Swiss Re has invariably found that the strongest health predictor models are those that are built from scratch on a bespoke basis. By definition this means that past data is needed of sufficient quantity and quality – and this is not always available.
Bancassurers have heavily invested in bringing their data to the point where it is an asset. As a consequence, they are now exceptionally well placed to take advantage of their data – and this is where Swiss Re has done most of its work to date. Typically, their wealth of past sales data – including underwriting decisions and claims – means that, when matched to banking data, unique insights can be learned by applying our statistical techniques to the anonymised data-file.
Strikingly, we have observed vast differences in the quantity and the quality of data held by our clients, even by banks.
3. Imperfect data is no excuse
Insufficient or poor quality data is frequently used as an excuse for not doing anything at all. This is a mistake - as Eric Siegler states in his book on predictive analytics: "a little bit of prediction goes a long way"1. Even those insurers with relatively weak data can – and should – be using it to make some improvement to their processes. We recommend exploring what can be done – rather than what cannot – with the data that is accessible.
Just five years ago, life insurers would ask whether they should be doing this type of analytics work, whereas now they are asking how they can do this. This is a good sign.
Life insurers frequently ask about different types of data-sources, and how they could be used to predict health, purchase or lapse. These are our learnings to date:
- Agent/broker data: Only suitable for predictive modelling if the data is held by all brokers in a consistent way – which is typically not the case. Even then, questions remain about the depth of the data, ie how much information is held on each customer.
- General insurance data: Suitable for predictive analytics of risk, purchase and lapse; but we find that this data is typically quite shallow, with only a few variables held consistently on each customer – maybe 15-20 in total. Thus, we need to be realistic about the likely strength of a predictive model built this way.
- Third party data sources: These can have very strong predictive power in markets where the data source has good breadth (covers most of the adult population) and depth (a number of data fields). The US and the UK have proved best for this kind of information.
- Loyalty card / supermarket data: This is frequently as strong – if not stronger – than banking data. The challenge is persuading these providers to extract/share their data.
4. Data cannot do everything
Many of us who work in predictive analytics are at risk of believing that, given the right data, we could make perfect predictions. Unfortunately, this will never be the case. For a start, models make mistakes (eg the customer with a low propensity to lapse decides to cancel their policy). In addition, we find that customers frequently do not behave the way we expect them to.
Life insurers must develop models suitable for their data. The response to a new online product cannot accurately be predicted by using past data for products sold face-to-face with agents.
Analytics can help improve the process but strong sales methods and messages remain important. Strong results can only be expected when good analytics come together with strong sales (or retention) processes.
The growing body of evidence emerging from the field of behavioural economics is a helpful reminder that we are not the fully-rational individuals we think we are –or claim to be. With this in mind, Swiss Re takes a 'test and learn' approach to all our data modelling. Only then will we determine the true drivers of customer behaviour.
5. Big data – lots of talk, little action
All life insurers are now converts to the idea of big data in principle. In practice, the life insurance industry as a whole has been slow in making use of the data they have access to.
Our work has proven that much can be achieved through matching life insurance data with other descriptive data (eg bank, general insurance data) in order to predict health, purchase or lapse. Some life insurers have led the way on this. We hope that many more will follow.
All the above are just some of our learnings to date and not any sort of final conclusions. Data modelling and predictive underwriting is a constant learning process and we are continually working with our clients to maximise its effectiveness.
If you would like to learn more about the life insurance modelling at Swiss Re, please contact: William Trump.
1. Siegel, E. (2013) Predictive Analytics: The Power to Predict Who Will Click, Buy, Lie, or Die. Hoboken, New Jersey: Wiley
Download full article (PDF 148 KB)