Skip to main content

How to make mortality improvement assumptions a model of consistency

A new framework for future mortality – part 6 of a 7-part series

Insurance Consulting and Technology
Insurer Solutions

By Richard Marshall | November 19, 2019

Part 6 in our series discusses how the proposed new framework aligns better with internal models of longevity trend risk, or with the longevity components of economic capital models.

Insurers typically predicate their calibration of a longevity trend risk model upon how their assumption-setting process would respond to new information. However, there is often a significant gap between the practices described in their model documentation and the reality at the point of setting a new best-estimate assumption.

So, how can we make sure they are more consistent with one another?

New information risk

Our recent Risk Calibration Survey shows that for most UK internal model companies, ‘new information risk’ – essentially changes to assumptions driven by news, not by data – contributes significantly to the overall longevity trend risk stress. The same is likely true across European Economic Authority (EEA) member states, as well as for longevity trend risk contributions to economic capital outside of the Solvency II regime (wherever these are based on a one-year value-at-risk metric).

Most approaches to modelling new information risk consider a range of scenarios in which one or more drivers of mortality are ‘stressed’ to reflect the impact that an item of new information could have on expected future improvements. Examples are:

  • Faster than expected decreases in the prevalence of smoking
  • Unexpected increases in healthcare, social care and public health spending, or
  • The announcement of novel pharmaceuticals or surgical advances

Scenario- and cause-specific mortality variations are often captured in cause-of-death models; the link between each driver and cause-specific mortality will be parameterised to capture effects such as lags between a change in the driver and the corresponding change in mortality, and the relative risks of mortality in the presence or absence of a driver (e.g. obesity).

Unlock More

About the series

In this series of articles, we discuss the impediments to understanding mortality improvements and the dangers of herding and group-think in this process. In response to these issues, we introduce a new framework for the development of mortality improvement assumptions.

Best-estimate improvements in the UK, however, typically take the form of one of the CMI projection models, with only infrequent variations from the core set of parameters (some insurers vary the long-term rate by age in a non-core manner). Typically, companies may use information garnered over the year to help decide what their long-term rate or period smoothing parameter should be, but they do not make out-of-model adjustments to reflect the impact of information received. Ultimately, the shape of the improvement surface is determined by the parameters in the model. Annual revisions rarely directly reflect the expected impact of any new information received.

Our alternative approach

By contrast, in our proposed framework, suppose that some extreme new information emerged which would change the insurer’s views of medium-term improvements. This information would be directly incorporated into the annual assumption-setting exercise, resulting in a change in the medium-term view. The change in the best-estimate assumptions one year on, due to the new information received, exactly mirrors the effect that receiving that information is assumed to have within the new information risk model.

Not only is the information incorporated within the new assumptions, but the change in the best-estimate assumptions (in the absence of any other information) reflects the shape of the modelled impact of that information, rather than it simply being incorporated into the short-term assumptions (period smoothing parameter) or long-term improvements.

The key difference is that the assumed responses (in the risk model) to the receipt of new information are consistent with the actual changes which will take place when setting the best-estimate assumptions.

Model risk

Extending these ideas to modelling, insurers typically allow for ‘model risk’ by comparing the central projections (using best-estimate parameters) from a range of models which represent alternative best-estimates of future improvements. Reasonableness of the outputs (given current information) is often largely ignored; the focus is on exceeding some regulatory hurdles for the magnitude of the stress.

Within our framework, we only reflect and communicate our views using the final (e.g. CMI) model, not set them. This means that model risk applies not at the level of this final model, but rather to the models used when developing the short-, medium- and long-term views.

Our framework emphasises the views of the insurer, not the output of a single model. This means a much greater focus on the effects of new data and information when calibrating a trend risk stress and a much smaller contribution from pure model risk.

Moreover, model risk exists insofar as the insurer’s views would be influenced by their modelling (e.g. choice of stochastic model to inform an expert judgement as to their short-term expectations, or alternative driver-based models in the medium- and long-term). However, where this generates implausible model results (on the basis of available information) these alternatives would not be adopted at the point of recalibration. Such model outputs should not contribute to the allowance for model risk.

Our approach tackles a common fallacy: the assertion that an insurer will change the structure of their best-estimate improvements model without receiving new information which strongly indicates the need for (or even mandates) this change. Insurers would test the plausibility of any new model that implied higher improvements against the information available to them. Assuming that they have allowed for pre-existing information within the current best-estimate assumptions, only new information should drive the adoption of such a model (as a response to a new information risk stress).

Use in ORSA scenarios

The framework also supports better ORSA (Own Risk and Solvency Assessment) scenario modelling.

Suppose a board risk committee wanted to see the effect of a social, political or technological stress, such as an unexpected complete ban on smoking/vaping, or the development of a treatment for Alzheimer’s disease. Under a traditional modelling approach, an analyst might determine the impact in a cause-of-death model and apply this as an out-of-model adjustment to the best-estimate. However, were the event to occur, at recalibration only a change to one of the best-estimate model parameters might actually be made. There is a disconnect between the theoretical scenario impacts and the impacts on the best-estimate assumptions if that scenario were realised.

Using our framework, the modelled impact of an ORSA scenario of this type would be consistent with the resulting evolution of the best-estimate assumptions were that scenario to materialise, giving senior managers more confidence that any planned management actions to allow the firm to respond to such scenarios are based on relevant information.

This can only improve an insurer’s understanding of risk derived from its capital model and support better decision-making by senior management.

Next time

In the final instalment of this series, we will consider how adopting our framework might affect financial information disclosed to the markets both in the short term and over the longer term. We will discuss the implications for results reported under:

  • Solvency II
  • IFRS 17

Related Capabilities

Contact Us