ForecastAgentSearch: Multi-Expert AI Agents Predict Geopolitical Events

Geopolitical forecasting is notoriously difficult, requiring nuanced understanding of local contexts, dynamic signals, and uncertain futures. ForecastAgentSearch, introduced at SIGIR 2026, tackles this by framing it as a multi-expert agent search problem. The system first analyzes the forecasting query to understand task context, then searches a library of specialized LLM agents. Each agent is profiled for regional knowledge, domain expertise (e.g., economics, conflict), reliability, and complementarity—allowing the system to select a diverse, high-quality team.

Once the best agents are retrieved and ranked, they provide independent analyses. A coordination module aggregates these perspectives, producing a final forecast with explicit reasoning and uncertainty estimates. This contrasts with single-model forecasting by leveraging specialization and offering transparency. The paper outlines key design challenges: how to build robust agent profiles, design effective retrieval and ranking algorithms, and ensure coordination without bias. The authors also propose evaluation protocols for future work.

While still preliminary, ForecastAgentSearch points toward a future where AI forecasting systems are not just black boxes but searchable, modular collections of expert agents. This could make geopolitical predictions more reliable and interpretable for analysts in government, finance, and intelligence—reducing reliance on ad-hoc human judgment or monolithic models. The approach is presented as a foundation for more accountable agent-based forecasting.