Tariff threats: how social media shapes expectations?

A team of researchers has developed a multi-agent model to understand how tariff threats influence households' economic expectations. The model uses 300 households from the Michigan Surveys of Consumers, transformed into persistent large-language-model agents exposed to social media information over several simulated months. Calibrated to reproduce demographic and distributional patterns, the agents show how factors like immediacy, rate salience, message complexity, and sender identity shape inflation and unemployment expectations. A second experiment reveals that central bank explanations can coordinate beliefs, though their effect on average expectations depends on message content. This approach allows for structured exploration of policy communication strategies, subject to human validation.

The developed model employs a multi-agent system (MAS) to simulate complex dynamics in economic expectation formation. Each agent represents a real household, modeled as a persistent large-language-model (LLM) agent. These agents are exposed to streams of information on social media platforms during simulated periods, allowing observation of how heterogeneous content influences collective expectations. Initial calibration is based on demographic data and real responses collected from the Michigan Surveys of Consumers, including specific distributional patterns. Experiments show that message structure — such as the immediacy of communicating a tariff threat, the numerical salience of rates, semantic progression, and narrative complexity — significantly modulates the variance and mean of inflation and unemployment expectations. Additionally, sender identity (e.g., a politician versus a central institution) interacts with message content to determine communication effectiveness. A key feature is that open-ended responses from agents allow tracing of underlying psychological mechanisms, such as perceived attention, ambiguity, and causal narrative construction. A second experiment demonstrates that explanations provided by central banks can serve as belief-coordination mechanisms, especially when messages are clear and coherent. However, the impact on average expectations remains sensitive to the specific content of the message, suggesting that not all institutional communications are equally effective. The proposed framework does not replace human validation but offers a controlled environment to test hypotheses about policies and communication, reducing reliance on costly repeated surveys.