Insurance, innovation and behaviorial evolution

Version française ici


A conjunction of elements led me to reflect lately on the link between insurance, innovation and behavioral evolution! It is an opportunity to give you my point of view on the issue.

Insurance is often considered reactive. Indeed, it reacts to events called claims to compensate for the residual loss.

Campaigns, sometimes famous (I think about road safety), are intended to transform behavior upstream to reduce downstream losses. For the insurer, it is an initial investment that finds its ROI in reducing the indemnification. We are here in a model in which the insurer is a preventer and no longer just a compensator.

However, a lot of debate has taken place in recent years to give the insurer a more proactive societal role upstream to reduce risks and thus change behavior. The embedded telematics (with Pay-How-You-Drive offers for example) has often been seen as an incentive for good practice: if you drive better, I make you pay less. Ditto for connected bracelets of health: if you play sports, you will have a reduction.

However, we note that these approaches do not always work or that they are not effective. However, do not we say “chase the natural will come back at a gallop”? The challenge is therefore twofold:

  • Find new ways to influence behavior.
  • Adjust offers to better stick to individual behaviors rather than standardizing responses.

Nudges at the service of behavioral change

Context

It will probably not have escaped you that the last Nobel Prize in economics (2017 ) was awarded to research on the so-called behavioral economy. To learn more, read here . This is also called benevolent manipulation.

insurance innovation and behavioral evolution
Cover page of the book Guide to Behavioral Economics, free download here . Thanks to Yuri for giving me the paper book.

I will not attempt here to detail fully the possibilities offered for insurance, because it deserves to think seriously about it. Anyway, I invite you to read the very good article from one of my competitors on the issue. Here is the example cited:

For some American motor insurances, the declaration of the mileage achieved in the year (n) conditions the premium for the following year (n + 1). The average rate of under declaration is estimated at 15% compared to the Km actually made. A Nudge has been tested to try to improve the statements. Instead of signing the declaration on honor at the end of the document, the researchers put it on the heading of the document, before the declaration. This simple change results in a 10% increase in the number of Km declared, a gain for the insurer of $ 48 per insured. The simple reorganization of the structure of the document, the architecture of choice, makes it possible to modify the behavior of the insured and thus to increase the income.

Some examples of nudges for insurance

Here are some examples of nudges that are or could be relevant to insurance depending on the objectives:

  • Reducing risks: The incentive to drive better.
    • The solution implemented by Liberty Mutual and relayed here consists of a mixture of gamification and information to drivers. Thanks to embedded telematics, the solution allows everyone to consult their driving information and to consider how to drive better and therefore how to reduce their premium, the discount rate being updated live every day for 3 months before being fixed.
    • By extending the exercise outside the pure scope of intervention of the insurer, we can note that tracing, on the road, shorter white stripes gives the impression of going faster and encourages you to slow down without even realizing it.
  • The incentive to subscribe:
    • A study relayed in the book above (London Economics and YouGov in 2013 for the FCA) showed the impact of the presentation of the insurance offer in parallel with another product. This is the case for example when buying the insurance when buying a mobile or cancellation insurance.

 assurance innovation and behavioral evolution

Beyond the questionable ethical aspect, this shows the influence of the distribution strategy on the choices of insured.

  • The choice of priorities: Last-mile problems (read for it this article and that one)
    • Segmenting the populations can be done according to a probability of underwriting. By identifying 3 segments (low, medium and high effort) using artificial intelligence, we can then determine who has the strongest probability of traveling the last mile to the subscription. These populations, for which the effort is small, will be treated differently and maybe receive, when the time comes, a little nudge! This can be a text message, a small message on the screen or a welcome email when giving the little extra that will be enough to convince them.

I will come back to these questions later with other detailed use cases to imagine more precisely how to use these methods.

Adjust offers

Most of the current offers are very standardized. They take into account a need considered identical for all insureds. At the time of individualization, why could not we consider adapting the coverage more finely according to the real risks of the insured. I can already hear the reactions: “This is the essence of our job”, “We already do it”, the contracts offer options and choices of levels to ensure according to his needs. “Very good, but in truth, can one think otherwise?

The Wilov speech is very interesting on this point. Most auto insured drivers drive less than 50 days a year. Yet they are covered full time. By inventing the “Pay-When-You-Drive”, their starting premise is to charge drivers only when they drive! The price is not necessarily much lower, however, the feeling of the insured is much better (provided that the user experience is the appointment of course!).

The approach of Inspeer goes in a similar direction. Starting from the premise that changing the behavior is complicated, Emmanuelle Mury and her team are working on the notion of affinity groups that go around the use. Clearly, they identify similar behaviors, and create the corresponding supply. I will come back to this in a dedicated article!

In short, the notion of supply is still too much seen today as unique, identical for all. Indeed, IT systems did not allow to easily and quickly deploy variants and customization, or the illusion of personalization was at the marketing level. It’s time to move to a higher level!

 

And you what do you think? When do we start to revamp your offers?

Lexis Nexis – Best practices for predictive modelling

best practices for predictive modelling

Version française ici


Lexis Nexis published this summer a white paper on the best practices for predictive modelling , or more precisely on the steps to follow to implement this type of solutions for small commercial.

According to them, 4 steps are necessary for a product creation of this type:

  • Ideation
  • Design and development
  • Implementation
  • Monitoring

So far nothing transcendent, isn’t it? Let’s check it out!

Ideation

Successful ideation assumes that two conditions are met: a strong sponsorship and a cross-functional team.

The responsibilities of this team are as follows:

  • Identify and validate the business problems to solve
  • Generate ideas on how to solve these issues with predictive models
  • Select the best ideas
  • Highlight the benefits of predictive models
  • Calculate implementation costs
  • Determine the ROI and justify the use of predictive models in relation to another solution
  • Establish acceptance of the topic among the teams.

Design and development

The report is focused on contracts / products for small businesses. The suggestion is then to go through an “insurance score” to analyze and estimate the risk and to price it, according to a probability of losses.

3 steps are needed:

best practices for predictive modelling

  • Data mining : nature of data, sources, refresh frequency, etc. For example, it is possible to use historical loss experience data ( Note: obviously …! ), but also credit data, or public data about the company. In a more detailed way, the geographical location is relevant. ( Note: at this point, note that we do not use anything complex! )
  • Model creation and validation : This is to determine, on the basis of a set of data, patterns or correlations that recur. We are here in deductive mode, we start from data to deduce a model. The challenge is to identify which data plays a role in achieving the desired goal . Then, it is possible to test the identifier models on data and production processes, to ensure that, when capturing the data, it is possible to categorize a new customer using the defined models.
    • Here is the kind of report that can be generated to define a number of groups to automate the subscription with 3 possible actions: acceptance (right), visa application (center) or automatic refusal (for worst groups).

    best practices for predictive modelling

  • Regulatory review : This aspect specific to the American market (but finally quite close to the regulatory aspects valid everywhere), suggests to compare the required data with the specificities of each state, and to apply, where applicable, restrictions.

implementation

The implementation is based on a few key steps

good practice predictive models

Monitoring

Finally, monitoring the relevance of the model assumes, on the one hand, to track the use that is made of this product, but also to measure its effectiveness.

best practices for predictive modelling

Monitoring

Finally, monitoring the relevance of the model assumes, on the one hand, to track the use that is made of this product, but also to measure its effectiveness.

best practices for predictive modelling

Tracking : Scores must be tracked when applied and when modified before application. In these latter cases, it is important to understand why and possibly modify the model iteratively to improve it.

Efficiency : The most important thing is to make sure in the long run that the model is good for achieving the business objectives that were originally defined. If this is not the case, it must either recalibrate (keep the mechanisms, but readjust the valuations), or rebuild it!

Small bonus on best practices for predictive modelling

Moreover, on this subject and always by Lexis Nexis, I invite you to consult this video, which includes some of the fundamentals:

best practices for predictive modelling

VA²CS – Detection and prediction of falls

Detection and prediction of falls

Creative Specific Software has developed an innovative solution for the detection and prediction of falls for the elderly. This solution is called VA²CS.

Halfway between connected object and artificial intelligence, it allows to offer a real service of prevention of a significant risk! <! – more ->

The need

For the record, there are in France 1250 falls per day, for 30 deaths. The need to detect the fall is therefore not new, and several solutions already exist. However, most existing solutions rely on bracelets, pendants or watches, and this, without major innovation for 25 years. The latest solutions are based on ground-based sensors and teleassistance boxes ( see the Harmonie Mutuelle solution – Orange ), but the cost remains generally important.

There are several major limitations to these solutions:

  • The first reaction of a falling person is not to press a button, whatever it is, but to try to get up
  • Once standing, the 2nd reaction is not to try to call for help because they do not want to disturb. The person is not going to be taken under examination immediately, with all the consequences on health aggravations in particular.
  • In the event of a fall with loss of consciousness, not all solutions can automatically send help
  • Most calls via teleassistance boxes are calls of convenience (against loneliness, need to talk to someone, need for psychological assistance, etc.). The average duration of these calls is almost 3 minutes.

Moreover, in the event of a trigger or an appeal, the tele-assistants have the legal obligation today to warn the helpers (under risk of not helping a person in danger). When the family is not available, this generates firefighters’ intervention costs (180 €). When we add the potential damage to the home, when it is not always relevant, because there is not always a fall, the consequences are heavy.

The solution

Fall Detection

With this in mind, and after working with Dr. Jean-Marie Vetel, one of the designers of the french GIR grid , Ramzi Larbi and his team have developed a new solution.

This involves equipping the living area with sensors (cameras), and then, thanks to the analysis of images captured in real time to detect falls.

When it is detected, a photo is sent automatically to the remote assistance center. It takes on average about ten seconds to give a certain answer to the action to take:

  • the fall is real and you have to send help,
  • either there is no fall, it is a false alarm, and the alert can be closed.

The desire is to obtain a solution at a reasonable price, the offer is based on conventional cameras and a connected box that contains intelligence and algorithms.

VA²CS already has 1500 EHPAD rooms in France, within the largest networks (Korian / Orpéa), as well as 600 private individuals. A partnership also exists abroad with Tunstall, the world leader in teleassistance for the elderly. The algorithms are patented in France and North America.

Prediction of falling

In order to complete its offer, the company has worked in partnership with Orpea to identify key factors in prediction of a fall. Then, she developed the corresponding algorithms. It is now able to predict a fall with a precision of about 60%!

Implemented in testing at Orpea, this solution has reduced falls by 50%!

Extensions

The solution was then supplemented more conventionally by intrusion detection coverage or facial recognition of visits to notify malicious visitors to relatives.

My opinion on detection and prediction of falls

This solution was not developed at first with a desire to meet the needs of insurers. However, it is totally in line with the current concerns of the sector.

The potential consequences of this type of solution for public health and for insurers are in my opinion significant.

Indeed, being able to predict a fall makes it possible to send notifications to qualified personnel or a relative to intervene. This is obviously valid in specialized institutions (EHPAD) to support staff reduced at night, but use cases can go further.

Thank you Ramzi Larbi for this exchange and good luck!

Note: I do not have a commercial relationship with the company providing this service.