When the market closes red and you remember you have a “set and forget” strategy.
§ Research · methodology · macro brief
The framework is grounded in published multi-agent LLM research. Predictions are verified against future price bars, the learning loop adjusts weights from closed-trade verdicts, and a free macro brief lets observers read what the macro digest reads.
§ 01 · Foundation paper
Multi-agent LLM frameworks demonstrate consistent outperformance over single-agent baselines and buy-and-hold strategies across diverse market conditions in academic backtests.
The paper shows that distributing analysis, debate, and decision across specialised LLM agents yields more robust trading decisions than a single monolithic model — particularly under regime shifts where bull-vs-bear review exposes weak theses.
View on OpenReview* Academic research results. Past performance does not guarantee future returns.
§ 02 · Methodology
A condensed map of how every cycle ends in a verifiable prediction. The full audit lives at docs/learning-audit.md.
Every candidate is scored on signal convergence, fundamental quality, catalyst proximity, technical setup, and macro alignment. Weights are learnable from closed-trade verdicts.
Bull and Bear researchers each build their strongest case and stress-test the other's three best points. The Consensus Judge decides on a split (e.g. 60/40) and emits a conviction rating.
Every prediction carries a horizon (1d/5d/30d) and is verified against price bars at-or-after horizon, with a 7-day tolerance to cover holidays/weekends. Verified rows feed agent_accuracy_rolling.
When a position closes, the three-round Strategist · Critic · Judge debate runs. The Judge assigns WIN/LOSS verdicts and proposes concrete parameter shifts (sector confidence, scanner weights, conviction gate).
§ 03 · Free macro brief
A short, structured note on geopolitics, central banks, sector rotation, and supply chains. The same digest the macro intelligence service generates for the trading agents.