Incentivizing Data Sharing for Energy Forecasting: Analytics Markets with Correlated Data

Reliably forecasting uncertain power production is beneficial for the social welfare of electricity markets by reducing the need for balancing resources. Describing such forecasting as an analytics task, the current literature proposes analytics markets as an incentive for data sharing to improve accuracy, for instance by leveraging spatio-temporal correlations. The challenge is that, when used as input features for forecasting, correlated data complicates the market design with respect to the revenue allocation, as the value of overlapping information is inherently combinatorial. We develop a correlation-aware analytics market for a wind power forecasting application. To allocate revenue, we adopt a Shapley value-based attribution policy, framing the features of agents as players and their interactions as a characteristic function game. We illustrate that there are multiple options to describe such a game, each having causal nuances that influence market behavior when features are correlated. We argue that no option is correct in a general sense, but that the decision hinges on whether the market should address correlations from a data-centric or model-centric perspective, a choice that can yield counter-intuitive allocations if not considered carefully by the market designer.