An active line of algorithmic research over the past decade has developed techniques for analyzing systems of self-interested agents. A crucial challenge in analyzing such systems is to predict aggregate properties at large scales; this requires drawing conclusions about global phenomena in systems whose behavior is currently only well-understood at the level of individual agents or pairs of agents. Deriving conclusions about macroscopic properties of systems described at a microscopic level is important in both computing and the social sciences. Market prices, for instance, arise from the microscopic interactions of individual traders. Understanding both the normal functioning of markets and their failure requires methods that can bridge the gap between these different scales of resolution.
This project uses ideas about networks and learning from algorithmic game theory to bridge the micro-macro gap. The research on networks considers theories of bargaining and trade in which participants are constrained by a network structure. This includes models for the distribution of power among agents in a network, as well as models in which prices in a market arise strategically through the interaction of market-making intermediaries in a network. The research develops models of market failures, particularly the kinds of cascading breakdowns of trust that played a crucial role in the global financial crisis in 2008. The project employs learning models to capture how perceived counterparty risk -- the ability of one's trading partner to complete a transaction --- spreads through a market.
The research on trust in financial markets can potentially contribute to broader policy debates about methods for restoring trust in markets. Currently there is a lack of analytical techniques that can tractably manipulate non-trivial learning dynamics to uncover the resulting network-level consequences, such as cascades. The research will provide tools for analyzing the determinants and evolution of trust in financial markets.
The project will also inform the development of introductory courses that cut across many disciplines, providing undergraduates from a wide range of backgrounds with a computationally grounded perspective for reasoning about the behavior and consequences of networks of interacting agents.
