The statement that human influence has warmed the atmosphere, ocean and land, has been a “high certainty” statement for some time now, at least since 2005.
Climate models have replicated trends in historical changes in global temperatures very well with changes in greenhouse gases and aerosols included, and these same models show little change in temperatures without them1.
However, this high certainty in global warming science should not be confused with the inherent uncertainty in modelling the impacts of climate change on business operations, particularly losses associated with extreme weather.
Uncertainty is a real and necessary part of modelling climate risks. In financial markets uncertainty brings volatility, which is very important in hedging, so why is a measure of uncertainty so hard to find when using climate risk metrics? Uncertainty shows us how good (or not) a single model is, or how many convergent or divergent views of risk there are between models.
Presently, there is an insufficient representation of uncertainty in commercially available climate risk metrics. In fact, the lack of disclosure on uncertainty and methodologies is, in itself, becoming a point to regulate.
For example, the Climate Financial Risk Forum (CFRF), co-chaired by the Prudential Regulation Authority (PRU) in the UK, recently released new guidelines on tackling financial-related climate risks2. The CFRF guide recommends firms clearly explain the limitations of the data they use in their disclosures and describe data gaps in the metrics employed.
Risk comes in three dimensions: hazard, exposure, and vulnerability. Risk modellers tend to spend more time focusing on the hazard (climate) component, but we know from historical loss information that the cost of extreme weather events is as much about how and where we choose to live and do business, as it is about underlying changes in the hazard.
There are five stages at which uncertainty can creep into modelling physical climate risks:
- The physical modelling of atmospheric variables, and the interactions between these variables across time and space by climate models
- The decisions on which future emission scenario to use
- The modelling of natural hazard footprints and assignment of frequency levels to these footprints, by hazard models
- The modelling of exposure (where the assets are, their value to the company, and their construction)
- The modelling of vulnerabilities (the physical damage, or loss of revenue, expected from the intersection of hazard events with a portfolio of assets)
In stage one (climate modelling), uncertainties relate to inter-model disagreement (model spread – i.e., the divergence of opinions between climate models), and the downscaling and bias-correction of climate simulations to make them useable for local applications. It is important to appreciate that for some variables these uncertainties impact the magnitude of the change (e.g., temperature) while for other variables, it is the direction of the change that is uncertain (e.g., rainfall). These uncertainties are amplified as more extreme events are considered.
In stage two (climate scenarios), uncertainties relate to the future trajectory of economic activity, the scale and timing of mitigation efforts (e.g., net zero targets) and the pace of technological development. Two widely used sets of climate scenarios are the IPCC’s Shared Socioeconomic Pathways (SSPs) and the Network for Greening the Financial System (NGFS) scenarios. In each case, the probability of each scenario is unknown.
In stage three (hazard modelling), uncertainties relate to the veracity of environmental datasets such as digital elevation models and the extent to which hazard model algorithms can replicate environmental processes. There are also uncertainties related to the statistical fit of modelled events to extreme value distributions, and the stochastic nature of climate variability, in assigning probabilities to hazard events.
In stage four (exposure modelling), uncertainties relate to the values and characteristics attached to in-situ assets, and the extent to which supply chains can be modelled (because physical climate risk can be inherited from third parties, as well as experienced directly).
Finally, in stage five (vulnerability modelling), uncertainties relate to the number of observations available to construct damage-loss curves specific to the hazard and asset type.
What to do?
We need models that can transparently express this cascading uncertainty in financial terms. Aggregation is also important for bottom-up approaches as it acts to smooth some of the assumptions made in the asset-level modelling, especially for geographically well-distributed portfolios. Reporting physical climate risk at a company level for risk scoping may be fit-for-purpose, assuming there are no systematic biases in the assessment, but basing financial decisions on address-level climate data is almost always misguided.
The stochastic loss modelling framework has a lot to offer in this space. Stochastic modelling techniques have been used widely within (re)insurance for the past 30 years, developed with the aim of putting a dollar-loss value on extreme weather risk. They are highly complex, probabilistic models that actively use (sample) uncertainties that are generated along the way. They are a mature, widely used framework that can contextualise a climate change signal within an envelope of uncertainty.
Risk is, at the end of the day, subjective and everyone (and every-model) will assess risk differently. Risk managers would do well to take an equally broad view of climate risk and embrace uncertainty in the process.
Aon wishes to thank Prof Andy Pitman, Director of the ARC Centre of Excellence for Climate Extremes (CLEX) at UNSW for his peer-review of this paper.
 IPCC (2021a). Climate Change 2021: The Physical Science Basis, Summary for Policymakers. Intergovernmental Panel on Climate Change, pp. 42.
 The Climate Financial Risk Forum October 2021 www.fca.org.uk%2Fpublication%2Fcorporate%2Fclimate-financial-risk-forum-guide-2021-data-metrics.pdf