Agentic Time Machine as an Infrastructure for Future-Event Forecasting

Researchers introduce Agentic Time Machine (TM), an infrastructure that reconstructs past web states to enable efficient evaluation of AI agents on event forecasting tasks. A multi-agent framework using this system achieves top performance on FutureX benchmarks and Polymarket predictions, demonstrating that offline evaluation correlates strongly with live forecasting results.

The development of Agentic Time Machine addresses a fundamental challenge in AI research: evaluating forecasting agents efficiently while maintaining environmental realism. Traditional live benchmarks suffer from slow feedback loops spanning weeks or months, while static database replays lack the dynamic nature of real-world information environments. TM bridges this gap by reconstructing historical web states, enabling researchers to rapidly iterate and test forecasting approaches at scale.

This work emerges from growing interest in AI agents capable of reasoning about future events, a capability increasingly demanded in financial markets and policy analysis. The FutureX benchmark has become a focal point for this research, attracting teams to develop sophisticated approaches. The proposed planner-solver-aggregator framework represents a meaningful architectural innovation—decomposing forecasting into parallel analytical angles before aggregating diverse evidence sources mirrors how professional forecasters and analysts operate.

The technical contribution has practical implications for cryptocurrency and prediction market participants. Accurate event forecasting directly influences trading decisions, market pricing, and risk management across DeFi protocols and platforms like Polymarket. The framework's demonstrated ability to achieve top rankings on both offline evaluation and live leaderboards validates a replicable methodology that could enhance forecasting accuracy across these domains.

The infrastructure itself may prove more valuable than any single forecasting model. By enabling rapid experimentation with reproducible results, TM could accelerate development of better forecasting agents. This standardized evaluation environment reduces barriers for researchers and practitioners seeking to improve prediction capabilities, potentially driving broader improvements in how AI systems handle temporal reasoning and evidence synthesis across financial and geopolitical domains.

• →Agentic Time Machine enables efficient evaluation of forecasting agents by reconstructing historical web states, eliminating the slow feedback loop of live benchmarks.
• →A multi-agent planner-solver-aggregator framework achieves top performance on FutureX and Polymarket by decomposing forecasts into parallel analytical perspectives.
• →Offline evaluation scores using TM correlate strongly with live forecasting results, validating it as a reliable sandbox for agent development.
• →The infrastructure approach may prove more impactful than individual models by enabling rapid iteration and standardized evaluation for forecasting research.
• →Results suggest LLM agents can achieve competitive prediction accuracy on financial and geopolitical events when properly architected and evaluated.