Joint-outcome prediction markets for climate risks

Experiments (CMIP) run by the World Climate Research Programme is a framework to com- pare and understand differences between CGCMs. The most recent CMIP6 exercise involved 100 models from 50 different modelling centers [2]. Interpreting what the differences in model predictions from different CGCMs might be telling us about uncertainty in climate forecasts is complicated. Ensembles of simulations from different models cannot be interpreted as proba- bilistic forecasts, not least because the models are not independent as they share common “ancestors”, or they have been influenced by the design of other models [3]. While much of what is understood about the climate system is codified in CGCMs, climate scientists also pos- sess more tacit expertise about the strengths and weaknesses of different models and what aspects of their predictions are likely to be most reliable. Synthesizing the output of different models and combining it with the knowledge of experts in a structured and quantitative way is one of the challenges confronting the production of consensus forecasts for policymakers and planners. One of the most important determinants of climate on a timescale of decades is the radia- tive forcing that results from GHGs in the atmosphere. As part of the CMIP experiments, models are run with specified concentration pathways of GHGs ranging from ones where the concentration of carbon dioxide (CO2) in 2100 is similar to its present value (400 ppm) to sce- narios where the concentration of CO2 has increased to more than 1200 ppm. Concentrations

of other GHGs (e.g., methane and nitrous oxides) are also specified in these scenarios. The amount of warming over the next century, according to CGCMs, is highly sensitive to GHG concentrations, ranging from 1.5˚C in the more benign scenarios to more than 4˚C should CO2 exceed 1000 ppm. Climate forecasts are thus conditional on the GHG concentration sce- nario used by the models. Producing an unconditional forecast of future climate requires a view on the probability of different GHG concentration pathways. The likelihood of different atmospheric concentrations of GHGs will be strongly affected by policies to reduce GHG emissions, and these policies will, at least partly, be influenced by predictions of future climate. In the context of inflation forecasting and interest rate setting this kind of interdependency has been called the “circularity problem” [4–6]. Circularity cre- ates ambiguity in the interpretation of unconditional forecasts: if only modest warming is pre- dicted over the next thirty years, this could be because forecasters do not believe the climate is sensitive to GHG concentrations, or it could be because they believe there will be dramatic reductions in GHG emissions. The former interpretation means little action is needed from policy makers whereas the latter interpretation implies a robust policy response is expected. Policy makers need to be able to differentiate these situations if the forecast is to be useful for informing policy. This issue is typically addressed by conditioning physical climate forecasts

on scenarios representing different emissions pathways. The challenge of predicting the likeli- hoods of the different emissions scenarios is treated as a separate problem that, arguably, has not received the same attention as physical climate forecasting. For example, the likelihood of the RCP8.5 scenario from the Intergovernmental Panel on Climate Change (IPCC) is a matter of ongoing dispute among climate scientists [7–9]. As mentioned above, the resolution of CGCMs is typically around 60 km, but many deci- sions concerning adaptation and resilience to climate change require information on smaller scales. The process of producing higher resolution predictions from CGCM simulations is called downscaling. Several different approaches are used for downscaling. Regional coupled models (RCMs) simulate the dynamics and thermodynamics of the atmosphere at a higher res- olution than CGCMs, but over a limited area [10]. They can be “nested” within a CGCM from which they get their boundary conditions. In contrast, statistical approaches use simulated quantities from a CGCM as predictor variables to predict quantities, such as temperature and rainfall, at higher spatial resolutions. Statistical approaches can include anything from PLOS ONE Joint-outcome prediction markets for climate risks PLOS ONE | https://doi.org/10.1371/journal.pone.0309164 August 30, 2024 2 / 14 participating in the markets. The manuscript was prepared after MR had left Winton and Winton did not play a role in the decision to publish, or in the preparation the manuscript.

Competing interests: The authors have declared that no competing interests exist.