Joint-outcome prediction markets for climate risks

To test the viability of the concept, a series of markets with prediction horizons of months, rather than years, but with two-dimensional outcome spaces and similarly high levels of granu- larity were constructed. Instead of targeting GHG concentrations and global temperature these markets predicted monthly temperature and rainfall for the United Kingdom. These var- iables were chosen to provide an interesting prediction problem on seasonal time scales, and one that would attract participants with comparable backgrounds and expertise to those that would be desirable in longer-range climate markets. Another attractive feature of seasonal forecasting for the purposes of this experiment is that there are several approaches to producing forecasts on horizons of a few months ahead. Ocean temperatures can be used as statistical predictors of seasonal temperature and rainfall while fully dynamical CGCMs can also be used [25]. Seasonal forecasting therefore provides an example where different types of information—different types of models and expertise about their relative merits—should be synthesized to produce a prediction. Market design Automated market maker To test the viability of a granular joint-outcome market, six joint markets to predict the monthly averaged daily high temperature and total monthly rainfall for the United Kingdom, for the months of April to September 2018, were constructed. The UK Met Office publishes these monthly summary statistics. Temperature was partitioned into intervals of 0.2˚C, rang-

ing from 0˚C to 25˚C, with open intervals covering temperatures below 0˚C and above 25˚C. Rainfall was partitioned into intervals of 5mm from zero to 200mm, with an additional interval covering totals above 200mm. This partitioning gave 127×41 = 5207 distinct outcomes, which were mutually exclusive and comprehensively exhaustive (see Fig 1). To overcome potential liquidity issues with so many outcomes, an LMSR automated market maker was used [22]. This algorithm will always quote a price at which it will buy or sell indi- vidual outcomes or combinations thereof. The LMSR market making algorithm uses the cost function C q ð Þ ¼ b log X m i¼1 eqi =b ! where m is the number of outcomes, qi is the market maker’s exposure to outcome i, and b is a liquidity parameter that determines how much prices change in response to a given change in the maker maker’s exposure. The parameter b also determines the maximum net loss that the market maker can suffer. For the temperature-rainfall markets values of m = 5207 and b = 500 were used. Each participating team was endowed with 1000 credits which were transferable between the six markets. This transferability allowed participants to deploy their credits at the prediction horizons where they believed they had the greatest informational advantage. The prevailing (marginal) prices the market maker assigns to each outcome, pi, are given by the derivative of the cost function, that is pj ¼ @C @qj ¼ eqj=b Pm i¼1 eqi =b Note that these marginal prices are normalized so that Pm j¼1 pj ¼ 1, which is why the out-

comes are defined to be mutually exclusive and comprehensively exhaustive. It is also why each pj can be formally interpreted as the probability of outcome j. PLOS ONE Joint-outcome prediction markets for climate risks PLOS ONE | https://doi.org/10.1371/journal.pone.0309164 August 30, 2024 5 / 14