Market alpha and structural risk

Research plan · Work in progress

This proposed study asks whether language models and research agents can connect scattered operational, regulatory and supply-chain signals into useful probability scenarios. Summarizing public information is a different task from identifying what would change an expectation.

The question

Can a model explain how several individually ordinary facts could combine into a material change, consider competing explanations, and assign probabilities that hold up on unseen cases? More agents or more retrieved documents would not, by themselves, answer that question.

Planned evaluation

• Freeze each case at a stated information cutoff. Exclude later documents and outcomes from model inputs.
• Compare a single model, a multi-agent research setup, and an explicit scenario-tree approach using the same evidence and declared budgets.
• Include cases with useful signals, misleading signals and no defensible conclusion. Require source links and allow abstention.
• Score probability calibration and source accuracy separately from writing quality. Compare against simple baselines, not just other elaborate agent setups.

Pending work

Assemble the case families, specify the scoring rules, and run held-out evaluations. No market benchmark or investment-return result is reported here. The worker-protection red-team cases motivate the broader multi-hop reasoning question; they do not establish financial forecasting performance.