Does the agent actually beat the market?
Out-of-sample weekly rebalance over a recent window. We publish the raw decisions so you can audit every call.
Evaluation protocol · Concordal-20×78 (target)
Designed target: 20 tickers (10 US + 6 A-share + 4 crypto) × 78 weeks of out-of-sample weekly backtesting — each ticker-week yields one full decision (LLM trace, 5 analyst rationales, bull/bear debate, risk-panel vote, manager rationale, realised 5-day forward return). The full 78-week run is being executed in batches — the live activity above and the curves below show the real published sample so far, not the complete dataset.
Loading…