Despite being a classic data driven business, the insurance industry is often regarded as slow in comparison to Amazon and Netflix. But one thing is for sure: insurers are catching up with advanced analytics methods such as predictive modelling. Ms Mandy Luo of ReMark demonstrates the power – and the value – of prediction.
The value of data analytics has long been clear, but the effective and advanced use of data is fast becoming the differentiator of business competitiveness. The elixir of any contemporary business is owning the customer experience, and this requires understanding and anticipating customer behaviour.
Indeed, data science thought-leaders, such as Dr Eric Siegel, illustrate how Big Data predictions lie at the heart of nearly everything, giving us the “power to predict who will click, buy, lie, or die” (Siegel, 2016).
The insurance industry is often regarded as slow in comparison to Amazon and Netflix, despite being a classic data-driven business. But one thing is for sure: insurers are catching up with advanced analytics methods such as predictive modelling. More specifically, the industry is starting to recognise the need to focus on data-driven behavioural insights throughout the customer lifetime journey.
Auto insurance leads the way in the use of telematics and data analytics, but life and health insurers are also beginning to explore how predictive modelling can be used in marketing, pricing and risk selection (Dion, 2011).
But while innovation has transformed certain sectors in particular markets, the industry as a whole still has some way to go.
Yet, go that way it must. Predictive analytic capability will be the driver of success – as the following case studies demonstrate. In short, providing deep behavioural insights at every stage of the customer journey will become imperative.
The power of prediction
Predictive purchase plus Persistency model – Growth market
A Colombian insurer offers Personal Accident cover to its policyholders and partnering bank customers through a telemarketing call centre. Despite strong response performance guided by a predictive response model, post-sales portfolio analysis reveals challenging 12-month lapse rates. While lapse rates do tail off substantially as expected after the first policy year, there is a clear need for a strategy that addresses both customer acquisition and persistency.
Three models were tested – Logistic, Stepwise Logistic, Decision Tree – using data from two sources:
- ~ 60,000 portfolio data with over one year actuarial exposure and “lapsed or not” flag.
- sales data with depersonalised customer attributes: demographics, affluence, credit and loan information.
These two sources are combined to obtain comprehensive customer attributes linked to response and persistency. The target variable is the persistency flag, defined as “Yes” if the policy is still in force after 12 months or “No” if otherwise.
Predictive model scores persistency at anonymous customer level, which, together with the model’s response score, creates a “dual” scoring for each customer. For a combined strategy, the dual scoring divides the customers into four segments, each with tailored strategies to improve the KPI.
This targeted matrix implementation has increased retention by 15-30%, without compromising total GWP revenue.
Predictive underwriting and Claims model – North America
A US life insurer offers 10, 20, 30-year Term and Whole Life cover, underwriting over 500,000 lives in the past 10 years. Around 3,500 death claim-records offer valuable claims experience to review the UW engine, address anti-selection and enhance the purchase experiences of “healthy” customers.
After filtering out irrelevant, repetitive or under-populated data fields from the 500,000 depersonalised records with a death claims flag (Yes/No), 20 data fields remain, including: demographic, geographic, product and distribution information. There are also third-party data such as prescription drugs and motor vehicle records.
The modelling target is mortality occurrence, defined as “1” if policyholders died during the observation period (2006-2015) or “0”. Due to the expected high correlation between age/gender and mortality rate, two modelling approaches are tested with Logistic Regression.
Method 1 produced four age-banded sub-models and Method 2 – with normalisation attempts to remove age and gender bias – proves effective to examine attributes outside age/gender, and is more stable contributed by a much larger total sample.
Scoring, segmentation and profiling help to demonstrate that the top decile has a mortality rate more than twice the average, while the lowest decile has a mortality rate less than half the average. The model differentiates the riskiest 10% – those over five times more likely to claim than the healthiest 10%. (See Chart 1)
This modelling could enhance the underwriting process, suggesting more rigorous underwriting scrutiny on high mortality segments, while providing increased cover and cross-sell with less painful underwriting for “healthier” customers.
Predictive recovery model – France
While long-term sick leave in France is covered by the state insurance programme, a quick recovery benefits everyone. A French rehabilitation service company provides recovery treatment through its network of medical professionals. While the company’s recovery rate of 70% is well above the European benchmark of below 50%, it is keen to better predict patients’ recovery and enable a targeted, personalised coaching approach to further enhance recovery rates.
Approximately 5,000 patient records with depersonalised data fields are reviewed. As “Recovered or Not” may be related to treatment duration, only records falling into a defined coaching period of 12 months are considered for modelling.
Logistic Regression identifies the strongest predictors as salary information, provincial region and sickness causes. The relevant coefficient estimate is used to score the next patient and predict their chance of recovery during the defined 12-month treatment period.
Ranked scoring shows that when long-term sick employees are referred to the rehab company, some of them have 87-98% chance of recovery within a year, and others less than 40%. The model is validated by scoring the next patients, in which over 75% of success cases were captured by the top four deciles.
The model not only identifies where preventative action is most urgently required, but offers valuable insight to optimise resource allocation, pricing strategies and other business inputs.
Beyond the surface
The value of experience cannot be overestimated in the application and interpretation of models. Textbook knowledge of statistics is not enough to build an effective predictive model. Without constant examination of the rationale behind model application to ensure that the insights extracted are relevant, models are no more than observations of the past and hypotheses for the future.
Solely relying on data and models is futile. Experience tells us that behaviours change and they change fast. People are often irrational when it comes to gauging risks (Jia, 2013) and regular validation of models is necessary to address such uncertainties.
Finally, a modeller needs to be on the lookout for relevant external data to append that would enhance the quality of the model. Credit data, for example, has been widely recognised as a powerful predictor of propensity to buy, lapse or claim. What’s more, wearables and personal quantification engender visions of future hybrid propositions between life and health.
The bottom line is that the importance of predictive modelling in commercial insurance has shifted. Once an innovative “secret weapon”, it is now considered to be mere table stakes. (Batty, Kroll & Stehno, 2009). A
Ms Mandy Luo is ReMark’s Chief of Actuary and Head of Data Analytics.