Joint-outcome prediction markets for climate risks

traditional linear modelling up to more sophisticated machine learning methods, such as neu- ral networks [11]. What they have in common is that the models must be fitted or trained using CGCM simulations of past climates and observations at the required granularity. They typically make the assumption that the relationship between the larger scale climatic variables, directly modelled by the CGCMs, and more local climate, will remain the same, even in a changing climate. In addition to public-sector CGCM modelling centers, there is now an increasing number of private sector companies—climate service providers (CSPs)—that produce high resolution climate forecasts of physical climate risks. Private sector companies have contributed to inno- vation in short-range weather forecasting for many years but there is an important difference between short-range forecasting and climate prediction: providers of short-range weather forecasts can establish track records of skill in a reasonable amount of time on which they can be judged. Users of long-range climate forecasts, however, cannot use historic track records to select a source of forward-looking climate information. The long “discovery time” associated with climate forecasts creates an information asymmetry between the user and the provider, who is likely to have a better understanding of their quality. This is known as the “lemon prob- lem” in economics, a reference to the market for used cars [12]. When users cannot ascertain the quality of forecasts it is hard for them to place a value on the forecasts. In response, provid-

ers might be reticent to invest in improving forecast accuracy. Markets suffering from infor- mation asymmetry can thus experience a self-reinforcing decline in the average quality of products, culminating in market failure. The information asymmetry can be mitigated by sell- ers being held accountable for quality through voluntary warranties or statutory protections (colloquially known as “lemon laws” in the market for used cars). Agreeing to be paid based on accuracy allows providers to signal confidence in the quality of their forecasts. Paying fore- casters according to a proper scoring rule can incentivize them to only make a prediction if they believe they have skill, even for a single forecast [13]. This trinity of problems affecting climate forecasts—distributed knowledge, circularity of predictions, and information asymmetry—could all be addressed using carefully designed “prediction markets” [14]. Prediction markets are designed to elicit and aggregate information, rather than transfer the ownership of assets or risk. They have been advocated as a way of aggregating information to produce collective forecasts, particularly when information and expertise is dispersed [15]. These markets are typically built around contingent contracts, known as “Arrow-Debreu” securities, that pay out 1.00 unit if a specified event occurs. These contracts are traded and, assuming the prevailing price (between 0.00 and 1.00) reflects the expected payout, it can be interpreted as a “market-based” probability of the specified event occurring. Prediction mar-

kets can continually assimilate new information as it becomes available to participants. Also, if the units represent real money, there is an incentive for participants to provide accurate and timely information, meaning that prediction markets can be an “incentive compatible” method for eliciting participants’ true beliefs. Prediction markets are neither a replacement, nor an alternative, to predictive models and expertise; they are a mechanism for synthesizing the outputs of different types of models and other forms of knowledge. Mechanisms to synthesize disparate approaches to forecasting are required more than ever due to the proliferation of new methods, such as machine learning, for climate prediction, as well as continued improvements to more traditional models. Predic- tion markets with expert participants have outperformed survey methods for predicting the reproducibility of scientific research [16] and they have been suggested as an appropriate tool to aggregate the diverse expertise required to make climate forecasts and to give those forecasts credibility [17–20]. Markets for year-ahead global temperature have been run on commercial PLOS ONE Joint-outcome prediction markets for climate risks PLOS ONE | https://doi.org/10.1371/journal.pone.0309164 August 30, 2024 3 / 14