Modern economic theory is largely founded on game theory. When explaining or predicting social outcomes, economists rely heavily on equilibrium notions, especially Nash equilibrium, i.e. a situation in which no party has any incentive to unilaterally change their strategy.
To understand when equilibrium assumptions are justified, we need methods of predicting the outcome of learning processes in games. Progress has been hampered by lack of data, and estimation methods that suffer from bias and low statistical power. Additionally, existing models make predictions about learning conditional on knowledge about initial play.
We will provide a new data set, new models of learning, and new methods of estimating and evaluating the models. The data set includes thousands of participants online, playing symmetric two‑player three-strategy games selected be representative of the entire space of possible best-response structure. We supplement this with a selection of other widely studied games and confirm results in laboratory experiments.
On the modelling side, the project develops fully portable predictive learning models that take only the game form as input, integrating a parametric model of initial play with standard learning dynamics, adding behavioural inertia, and incorporating transfer across game. Methodologically, we introduce new estimation and evaluation tools based on cross‑validation and end‑state prediction, supported by extensive model/parameter recovery exercises.
