In the first of a series of six blogs, Gordon Woo, catastrophist at RMS, Will Davies, chief actuary at Insurem, Matt Attard, senior analyst at Marsh analytic solutions and Sophie Weisenberger, senior consultant at RPC Tyche, consider outcome bias and its impact on actuarial analysis in the context of COVID-19. They are part of G19, one of the workstreams of the IFoA’s Covid-19 Action Taskforce
The focus of the G19 workstream is on insurer and consumer behaviour and unintended bias. The key discipline underpinning the study of human behaviour is social psychology. Daniel Kahneman is the psychologist whose exposure of cognitive biases has radically changed the thinking of risk stakeholders. This blog series looks at six biases highlighted by Kahneman with the first blog focussed on outcome bias.
Suppose an unusual risk event occurs. Future perception of this kind of event will depend on what the outcome turns out to be. If the outcome is benign, then the perception will tend to be fairly relaxed in respect of risk preparedness. However, if the outcome is a notable loss, the perception will change to concern, with a need for more urgent risk preparedness.
In 2009, an influenza pandemic emerged from Mexico. The pandemic virus was highly contagious, and spread around the world quite rapidly. However, the case fatality rate was low - somewhat less than for regular seasonal flu. Public perception of this pandemic was driven by the comparatively benign outcome. The UK government was heavily criticized for buying at considerable expense a large quantity of an influenza antiviral drug, Tamiflu, which ultimately was not really needed.
COVID-19 is the worst pandemic to have struck for a hundred years. The underlying virus SARS-CoV-2 is highly transmissible, and the case fatality rate is somewhat greater than 1%, which is ten times higher than seasonal flu.
Preparedness for this pandemic catastrophe was limited, despite the fact that, in the first two decades of the 21st century, there were five near-miss pandemics each of which had a case fatality rate of 10% or more. These were SARS in 2003; H5N1 influenza in 2005; MERS in 2012; H7N9 influenza in 2013; and Ebola in 2014. In each case, a different throw of the pandemic dice might have resulted in a slightly more transmissible virus, which might have generated a global lockdown economic disaster.
Countries such as Vietnam, Taiwan, Singapore and South Korea which had a bad outcome from SARS and MERS were relatively well prepared for COVID-19, especially in respect of diagnostic testing capacity. With a cumulative population about three times that of UK, the cumulative death toll from COVID-19 has been less than five hundred. A British aid nurse was seriously ill with Ebola, but otherwise, UK was not as badly impacted as other countries by emerging infectious diseases of the past two decades. This contributed to a sense of market complacency for some insurance coverages, which is an outcome bias.
Application to Actuarial Work
Considered in the broadest sense, actuarial estimates are highly subject to outcome bias as they are typically based on analysis of historical data- that is, the outcomes of historical events. In most analyses, little effort is given to imagining what alternative outcomes didn’t occur and so are not included in the data being used to set the current estimate. That this limitation exists is often acknowledged in actuarial processes, for example with the application of ‘contingency loads’; in casualty pricing or management margins in reserving. Stress and scenario testing addresses this issue, but it may not cover near-misses adequately.
An occasion when this analysis of near-misses could be carried out is dealing with ‘Events Not in Data’; which we are required to assess when setting Technical Provisions under Solvency II. This seems to present a good opportunity to engage in deep-thinking about the plausible-but-not-yet-realised alternative outcomes of insurance events. It is difficult to separate this analysis from the capital-setting process but that is a discussion active elsewhere and so we will not delve further into that. An area of insurance modelling that seems to deal well with outcome bias is property catastrophe modelling. For a given exposure set, in addition to applying known historical events, a series of plausible-but-not-yet-unrealised events is also modelled.
The operation of outcome bias also affects the probability weighting applied to events considered and thus their respective likelihood of occurrence, relative to other outcomes. In reserving, outcome bias is particularly prevalent when deriving Best Estimates. Best Estimate reserves attempt to capture an estimate of the possible cash flows weighted by their respective probabilities of occurrence, using all available information at the valuation date. This estimation basis is supposed to be free from conservatism and other biases. Nevertheless the operation of outcome bias could be reasonably evidenced, if the judgement and consideration of information driving the inclusion of contingent cash flows and the estimation of their respective probabilities are influenced by the ex-post outcome of the contingency, which may be rather benign.
The following example can perhaps shed some light on the matter: Prior to COVID-19 trends for loss development factors in estimating ultimate claims development may have had limited consideration of possible settlement or development delays, say due to reasons such as operational disruptions or perhaps delays due to systems migration etc. However post-COVID, claims development decelerated, due to lock-downs, inaccessibility of loss adjusters to the loss site, and mobilising of remote-work facilities following office closures.
Based on this outcome, development factors may be biased towards a slower immediate development due to the weighting of possible disruption due to the prevalence of the pandemic at the time of reserving. The impact on loss development would have been particularly material on liability classes as court proceedings would have halted or been materially hindered. However, the operation of outcome bias here lies in the possibility that that another bout of infections may not necessarily lead to the same reactions by public health authorities and insurers. Accentuating the probability of a slower loss development of losses may be a manifestation of outcome bias.
We may see outcome bias emerge in the future as a result of COVID-19. Those insurers that have experienced smaller losses than expected might give less attention to pandemic risk in future. It could be that this is somewhat justified if they truly have little exposure to pandemic risk, or it could be that the nature of their business mix means that, while they have little exposure to COVID-19, they could still experience large losses from a pandemic that manifests differently. For example, while COVID-19 has affected the health of young people to a lesser extent than the elderly, other pandemics in the past have had the opposite effect. This just reinforces that analysis of near misses is important to understanding the risk profile.