There’s a lot of talk at the moment about insurance undergoing a transformation. Some point to big data, and others to analytics and algorithms. While both of those play their part, they are not themselves the transformation. That is happening within insurance itself and goes by the name of personalisation.
A lot of people in the market see personalisation as the future of insurance. They talk about it bringing insurers and consumers closer together, in a digital, mutually beneficial relationship. Yet it might not be so simple. There are signs that personalisation could in fact delivery something quite different and far less beneficial. Could it take insurance into a ‘digital winter’, in which the market struggles to perform its expected function and consumers feel like they’re increasingly being left to fend for themselves?
So what is personalisation exactly? It involves basing the premium, cover and claim around the individual, rather than around a pool of similar people. This means pricing a policy according to the particular individual being insured. It means providing cover according to the particular risk being presented. And it means varying that price and cover according to the data streaming on a real time basis into insurers’ systems.
That personal angle has a nice feel to it. You wouldn’t be paying for the costs of claims made by that accident prone neighbour. Your premium would be for your own risk, untrammelled by the misdeeds of others. What’s more, you won’t have policies listing lots of complicated ‘perils’. Yet there is more to these benefits than you think.
Historically of course, insurance has always recognised the individual, but seen them through the lens of risk pooling. Now that lens is progressively disappearing. All that big data, all those algorithms are allowing insurers to see each policyholder through their own personal lens.
The reason why I don’t think personalisation like this is the future of insurance is because it comes with three inherent flaws, which I’ll outline here.
Identity and Reality
Personalisation relies on knowing lots of things about each of us, in order to individualise our premiums and claims. That means lots and lots of data, most of what will be unstructured: in other words, data that is not organised in a pre-defined manner. As a result, it will contain irregularities and ambiguities, which will need to be ironed out in order to meet data quality standards. Yet the techniques for ironing out unstructured data can themselves introduce their own irregularities, such as gender bias (more here).
Once in a more structured state, the data needs to be tagged, a process pretty similar to sticking labels onto it. This then allows data analytics to sweep through the dataset, looking for patterns that align with the firm’s pricing model. Personalisation can result in simple portfolios having over 1000 data categories and over 100,000 micro-pricing segments. We’re talking about quite some granularity here.
Each of these stages (unstructured data, tagging, segmentation) introduces interpretation and translation risks. Yet these are largely managed by powerful tools at the heart of big data analytics, such as the correlation. Now, a correlation expresses the possible strength of a relationship between two data objects. It does not express the actual relationship between them: that is what causality does. So what we have is a trend (personalisation) that orientates insurance decisions around the individual, but which also relies on analytical techniques that are unable to accurately represent the individual. This means that while personalisation is all about the person, the extent to which personalisation produces an accurate representation of a person is always open to question. It introduces a tension therefore between the identity the insurer has constructed for you, and the reality of you as a person.
One could well ask of course how different this is to the underwriter's historical use of pooled profiles or proxy measures. They seem to land the underwriter with much the same problem. Yet proxy errors emerge from a pool of risk, while personalisation errors emerge from individual risks. Their impact is therefore much greater.
It's also important to remember that all data exists in a context. Consider whether you drink bottled or tap water. For some US insurers, this is a pricing factor for motor insurance. For policyholders, this piece of data sits in a personal context that is social, environmental, economic or political.
And your water preferences will change according to a decision in one or more of those contexts. It will not be influenced by anything to do with your motor insurance, and in turn will not influence how you drive. Yet the motor insurer notices that change in water preference and changes your premium as a result. It has created its own context for your decision that is likely to be unconnected with the reality of how you go about day to day life.
Added together, what emerges is a record used by the underwriter to price your policy that bears questionable resemblance to who you are in reality. It is an identity held together by statistical methodology. A lot of statistical methodology is very neat and precise, built upon mathematical techniques widely recognised for their objectivity. Yet they cannot, and never will, represent anything other than a partial you. Personalisation will never actually achieve its goal.
Let’s remember the fundamental reason why people buy insurance. It is to spread the cost of loss over time, by replacing a loss that is uncertain by time and extent, with a payment that is regular and predictable.
Now consider where personalisation is taking insurance. It is framed around a huge expansion in pricing segmentation, and a contraction of the risk pool to one. If that stream of data powering the real time underwriting of your policy signals a change in any of those micro pricing segments, then this will be reflected in your premium in some way. It has to, for that is what pricing on a personalised basis means.
As a result, consumers find their insurance changing, from being priced on a relatively stable basis, to being priced on an ever changing basis. For many people, a lot of the time, those changes will be negligible. However for many people, some of the time, life does have its ups and downs, and those occasions will produce more substantive pricing movements. So personalisation is going to herald a new insurance market, in which a regular and predictable price will become a thing of the past, replaced with episodic swings between negligable and substantial price movements.
And what’s more, consumers won’t be able to understand what has triggered a change in their personalised premium, because there will have been so many possible data points that could have been responsible. Insurers may give policyholders a nice online dashboard, but if there are a 1,000 underwriting factors at play, will policyholders really be given risk reduction insights, or just pricing signals? It may feel like transparency, but is it really?
And should you claim of course, your personalised premium will respond accordingly. And this will especially be the case for people who only claim very occasionally, who might have enjoyed low premiums for some time and now face hugely increased premiums. What this adds up to is a public who just don’t understand how their policy is priced, what they can do about it and how it might look at next renewal. Insurance will feel inherently unstable, at which point, some may ask: is this what they buy insurance for?
Fairness and Equality
Let’s move on and remember that personal insurance in the UK is a market, full of firms competing for a profitable share of a big market. Yet while this has often (but not always) delivered widespread access to insurance, personalisation is starting to change that. A reliable source is of the opinion that stratification is already emerging in some insurance markets, caused by insurers chasing the more obvious (and more immediate) sources of profit.
Now, some of you may think that this is just life: we can’t always have what we want, and besides, if we send the right pricing signals, then the market will move to meet them. There’s some credence to this, but not enough. Firstly, insurers can ignore or misinterpret pricing signals, depending on how their strategies have shaped their corporate mindset. And in switching almost exclusively to digital pricing signals, insurers have orientated their pricing ear around one particular signal used by many, but not all.
For some years now, there have been political concerns over financial inclusion. Those concerns now recognise that while particular groups of consumers regularly experience exclusion by financial markets, many more people will at certain points in their lives experience some degree of financial vulnerability as well. The recent ‘Insuring Womens’ Futures’ report provides a clear example of how women experience this.
Personalisation will sense those periods of vulnerability, yet largely interpret them through the lens of price and profit. More of us than has commonly been thought will experience products becoming more expensive or less complete. Social sorting, as it is called, will become more widespread.
And personalisation will result in the data rich being priced with more attention and enthusiasm than the data poor. And adverse selection is going to cause the data poor to more often be underwritten as a high risk, on the basis that known risks should not cover the uncertain costs of ‘less known’ risks.
Situations like these give rise to debates about fairness, but they’re often discordant. Insurers emphasise the fairness of merit, while others emphasises the fairness of access and need. Fairness is, as the regulator has admitted, hugely important and hugely complex. That’s why it’s disappointing that those involved seem to have difficulty finding common ground upon which to build.
While fairness is certainly raising questions about the alignment of the insurance market with the needs of the public, debate is likely to be triggered more immediately by another factor: the propensity of the algorithms behind personalisation to generate discriminatory outcomes (more here). It is of course not at all the intention of insurers to deliver quotations that seem to be discriminatory, but that is irrelevant. If the discriminatory outcomes are generated, then the issue of bias within underwriting systems will have to be addressed. It will be a difficult debate.
When describing what all this will look like when added up, I often use the analogy of a cake. If your underwriting knife, fuelled by personalisation, keeps on cutting that ‘insurance cake’ into ever more slices, then the result will no longer be a cake, but a pile of crumbs. The essential nature of what you started out with becomes lost.
The ambition may have started out as the dawn of a new insurance, yet the outcome is beginning to look like something other than insurance as we've known it for so long. The focus of personalisation has evolved as well, away from a risk/profit balance, towards a profit only basis. Prices are being based more on what people are prepared to pay; claims are starting to be settled on what claimants are prepared to accept. Yet if you no longer price on risk, if you stop basing claims settlements on insured losses, then it is no longer insurance, but a financial instrument at best akin to a savings plan, at worse a complex and barely trusted irrelevance.
Regulators are noticing that insurance is starting to produce, in their words, an 'uneven distribution of benefits'. When the FCA’s Director of Competition was asked last week about this, she talked about the regulator ‘thinking a lot about its implications for market confidence’.
Compare this with the fate of the UK’s short term credit market. Success in that market had become linked with the use of algorithmic pricing, based around huge numbers of data categories. However, public disquiet about pricing models and outcomes led to that market being labelled as dysfunctional and subject to pricing and product regulations that have made it a mere shadow of its former self.
Is personalisation pushing insurance towards a similar fate? There are signs of this. As one artificial intelligence expert said recently: “If you replace a lot of people with one algorithm, it becomes a single point of failure.”
Consider the example of flood insurance in the UK. Insurers built huge databases that gave them per-property granularity. Alongside this were positioned models of flood patterns. As Statements of Principle between insurers and the UK Government still allowed risk related premiums to be charged, the market moved prices and excesses steadily upwards for properties deemed to be in areas of heightened flood risk. A study for the Scottish Government found that of the 290,000 households in those flood risk areas, 41,000 would be unable to afford insurance. Yet as price increases moved from steadily upwards to exponentially upwards, even that estimate looked too small.
In the end, a solution was found. Policyholders deemed to be in areas of high flood risk were offered insurance at normal rates and excesses, with household policyholders across the country paying a small extra premium each year towards a reinsurance pool. So as personalisation took hold of the flood insurance market, the business model to emerge as the solution for the Government, the market and the public, was one big insurance pool. The irony was not lost on some.
Last year, I was part of a panel of speakers at a leading EU data conference that addressed the question as to whether big data was steering insurance towards a cliff or a superhighway. It looks like on this occasion, advances in personalised flood pricing drove the household market close to the cliff, before it was negotiated away from the drop just in time.
The Shape of Things to Come?
I believe the future shape of insurance will not be formed around personalisation because it is a solution that ultimately serves the market far more than it serves the consumer. It involves too much push, and not enough pull. It’s built upon inherent partiality, and will progressively feel exclusionary, rather than complete and inclusionary. Therein lies its fatal flow.
At the moment though, insurers feel emboldened by how much they can find out about their customers. For the first time since door to door agents, insurance feels closer to their customer. Yet proximity is not the same as intimacy. Is a personalised approach to pricing causing customers to want to get closer to their insurer? Some may have felt this initially, but it doesn’t seem to be gaining wider traction.
There are signs that granular underwriting can create a lot of customer friction. People either don’t notice personalisation, or they experience it as unfair in some way. And they perceive its downside as out of proportion to its upside. Not a great recipe for success.
What will success look like for the insurance market of the future? That I will cover in a forthcoming post that looks at the structure of insurance innovation and the signs of, and conditions for, its future direction.
In the meantime, what about the present? Let’s roll the personalisation ball forward a few years. More data and more sophisticated analytics will yield more granularity of underwriting. Yet this will be at much expense and yield diminishing returns. Competition will be secured through microscopic advantages from customer segments defined ever more tightly. The difference between success and failure will feel much narrower, more fraught and less certain. Will investors start to look quite differently at the sector?
Looking on will be a regulator increasingly concerned about market confidence in the sector, and a public increasingly perplexed, occasionally angry, at some of the unexplained outcomes the sector is generating. Add in Parliamentarians unable to reconcile the personalisation message from the sector with the stratification being experienced by some constituents, and you have a cauldron of disruption, fuelled not by innovation but mistrust. I think the future for insurance is bright, but it is not personalisation.