How would you explain the fairness of using a particular rating factor? You can’t rely on just statistics anymore. While the super-complaint tied fairness to the overall premium, now long running challenges to certain rating factors are re-emerging. Underwriters need to prepare for this.
Fairness and pricing has become a much more debated subject in insurance markets. In the UK, the super-complaint into household insurance pricing pushed the issue firmly onto the regulator’s radar. Yet remember this was largely about the discounts and loadings applied to ‘a premium’, in relation to the likelihood of a policyholder remaining loyal to a particular insurer.
Insurers should expect the debate to now move on to the constituent rating factors that together create that premium. You can see the beginning of this in the growing debate about whether it is fair to use credit scores in the rating of a household or motor policy. Questions can expected about location data soon too.
Some in the market are picking up on this and weighing up their options. One or two insurers are starting to reject credit scores, using this as a marketing tactic to differentiate themselves as an insurer to be trusted.
Selling Something the Buyer can Understand
What the emerging debate on credit scores points to is a wider interest not just in whether a premium is fair, but in whether a rating factor is fair. It is a debate that is moving alongside an already established debate on the understandability of policy wordings. Insurers working on the understandability of policy wordings should consider broadening this to include the understandability of the rating of the associated premiums. After all, why make a policy wording understandable, if the consumer then has no way of knowing if that wording is being rated fairly.
Why go to all this effort, some of you may ask. Well, insurers working on the understandability of policy wordings are doing so as part of a wider strategy to build more trustworthy relations with customers. They’ve dropped the ‘just trust us’ approach and turned instead to ‘this is why you can trust us’. The former can work but not for any great length of time and now in progressively more limited circumstances. The latter is more sustainable and more affirmative.
Another reason is that a wider debate about the ethics of using certain types of data for certain purposes now often turns to insurance as a classic example of the problems that can emerge. In other words, in the wide ranging reading I do on data and ethics, insurance just keeps on turning up as the case study par excellence.
Framing an Explanation
This is therefore building expectations on underwriters to explain their rating. As I said earlier, ‘just trust us’ no longer works. ‘ Show us your workings’ is becoming more the norm now (remember those Colorado laws I looked at here).
So how can underwriters move from explaining their use of credit scores, to having a framework for explaining the much wider range of data being used in rating? This is what I’ll turn to now, drawing on some interesting research published last year in the United States.
A lot of this article is orientated around a paper published in January 2021 by Barbara Kiviat, a professor at Stanford University. She’s an economic sociologist who studies how moral beliefs and other cultural understandings shape markets. Her paper is “Which Data Fairly Differentiate? American Views on the Use of Personal Data in Two Market Settings” (you can read it here).
Much of the paper is based around the findings of a large survey she conducted into consumer views of sixteen rating factors and their use in motor insurance and consumer lending. Some of the rating factors were common to both markets, while others related to just the one market.
And she focussed the responses being sought by confirming to all participants that the use of each rating factor was confirmed by statistical analysis and that each type of rating factor did predict insurance claims or loan nonrepayment. Any questions relating to consent were scoped out in the same way.
All respondents were asked to rate the fairness of using each type of rating factor, from very fair to very unfair. For some rating factors, respondents were then asked to explain their answer.
The Importance of Context
What Prof. Kiviat found was that Americans perceive some types of data as fair to use and other types as overwhelmingly unfair. So while there is a consensus that data can be used in market contexts to differentiate, this did not mean that all data can be used in that way. Context mattered, enormously.
For example, participants found that motor insurers using data about speeding tickets was fair, while lenders using it was unfair. Pretty much the reverse was the case with income. So people weren’t thinking about the fairness of firms using personal data in a generalised way. Instead, people judged that use to be fair in some contexts, and unfair in other contexts.
The survey found that responses fell into three broad clusters. Some data was permissible and other data was proscribed. So for example, respondents felt it was very fair for insurers to use accident history, but very unfair for them to use shopping history.
In between permissible and proscribed data was an ‘unsettled’ middle, sometimes with strong but conflicting views. Around one in three people felt that it was very fair or somewhat fair for a lender to use data about the number of past addresses, but at the same time, one in three felt this was very unfair or somewhat unfair. For motor insurers, one in five felt that data about someone’s driving was very fair or somewhat fair to use, but at the same time, one in five felt such data was very unfair or somewhat unfair.
Only Some Behaviours are Fair
Now some of you might think that the behavioural nature of some of this data would matter. Not necessarily, Prof. Kiviat found. Some behavioural data fell very firmly into the proscribed cluster – for example, web browsing history, social media use and retail purchases. Other behavioural data fell firmly into the permissible cluster – for example, speeding tickets and timely bill payment.
Prof. Kiviat found some interesting variations in responses relating to telematics data. Half the respondents thought that it was somewhat or very fair for a motor insurer to use data about how a person drives (whether they slam on the brakes, turn sharply, etc.). However only about a third of respondents thought that it was somewhat or very fair for a car insurer to use data about where or when a person drives. Americans think it is fair for insurers to hold them accountable for some, but not other, behaviours.
More than Statistics is Needed
So what does this tell us? Other research by Prof. Kiviat has found that moral claims based on statistical relationships were fragile. To be seen as legitimate, they need to be backed up by more intuitive explanations of why two things are related. What this tells us then is that insurers cannot justify the fairness of using such and such a rating factor by claims of mathematical relatedness. They must rely much more on ‘logical relatedness’. In other words, the connections people draw through reasoning about how the world works.
When Prof. Kiviat explored the qualitative feedback from participants in her survey, she found that a rating factor was judged as unfair where the data it drew up conflated situations and behaviours. So an insurer might see the number of past addresses as a red flag for evictions or troubled finances. Respondents saw this as too narrow an interpretation, citing a host of family and career reasons for moving. They saw it as unfair to interpret some types of behavioural data without knowing additional context.
Another example might be driving at night. Was the person returning from an evening of partying, or returning from a late shift at work. The fact that they are driving at night is not enough to make a financial judgement about their character or behaviour.
Water as a Motor Rating Factor
I’m minded by all this of my favourite motor rating factor, that I found in use by a US insurer a few years ago. It is whether you drink bottled water or tap water. The fact that you might decide (or change) your drinking habits for environmental, political, economic or social reasons is not judged relevant by this insurer. What matters to them is that there is a mathematical relatedness between premiums / claims, and the source from which you get your drinking water.
Now, some of you will be thinking something along the lines of “well, this is business. What happens with premiums and claims matters”. And on one level, you’d be right. However, another way of looking at this is from the perspective of the customer, who is the person giving you their business, trusting you to deal with them and their business in ways that are fair. They think it is very unfair to use shopping data in the rating of a motor premium. An insurer needs to be sensitive to both aspects: premiums / claims to be sustainable as a business; customers to have a business in the first place.
Expect More Transparency
There is a lot of fascinating stuff in Prof. Kiviat’s paper, but I want to now consider how insurers might use what her research has found. That it’s worth making use of her findings is, I believe, pretty clearly evidenced by events of the last few years, at least here in the UK insurance market.
There is a clear trend towards greater transparency around how premiums are determined. The super-complaint and Parliamentary concerns are behind this. The regulator is having to respond, not least to fulfil the promises it made to the Treasury Committee a couple of years ago. Insurers need to prepare for the ramifications this greater transparency will have.
I believe every insurer should be able to justify the fairness of each of their rating factors. To begin with, this should be done internally, with the expectation of having to eventually submit those justifications, in some form or another, to an external audience.
Then it should be integrated into relations with business partners, such as data brokers and software houses. They are after all a big source of new sources of data relating to underwriting and pricing.
Listening to Customers
These explanations should bring together both a quantitative justification for that rating factor’s use, and a qualitative justification. The quantitative justification needs to be layered, showing how the statistical outputs vary at different levels of context. Those different levels of context need to be framed around outputs from the qualitative work done to understand how the rating factor sits in the real world inhabited by the product’s target customers.
That qualitative work needs to be drawn from real target customers. Opinions from within the insurer would not be good enough. Nor would the opinions of data brokers and software houses be good enough either. The person within your firm assigned to act as the ‘consumer’s voice’ is a starting point, but only for initial scoping. Your need to listen to your actual customers.
A good starting point for such an exercise would be the use of credit scores, for three reasons. Firstly, there is lots of qualitative and quantitative research out there to inform your considerations. Secondly, it’s a rating factor that can be expected to need explaining sooner rather than later. And thirdly, that’s because there are consumer groups out there now advocating against its use in insurance.
The Wider Landscape
As I've said before, the super-complaint was just the end of the beginning of the issue of 'fairness and pricing'. The focus is now moving onto the fairness of individual rating factors. It's part of course of the wider ethical landscape relating to data and analytics. That's because rating factors need data. Understanding that wider landscape should form part of every insurer's digital strategy.