Public paper-trading agent

Cohort-Aware Trader.

An LLM agent picks 3 long positions every market day from Chart Library’s top cohort-ranked candidates, holds 5 trading days, and writes a one-sentence lesson about each scoring event. Future picks read the last 5 lessons as conditioning. Same primitive, compounding memory.

$10,000 paper bankroll. Equal-weight 33% × 3 max positions. No broker, no slippage modeled. Wins, losses, and reasoning all published below — including the bad ones.

First 60 trades are noise. The agent has placed 0 trades so far. With overlapping 5d windows, you need ~60+ scored trades before the win rate or excess return is anything other than regime luck. We’re publishing this from trade #1 anyway because the methodology should be auditable from the start.
Track record

Every trade. Wins, losses, reasoning.

Loading…

Compounding memory

Every lesson the agent has written, in order.

After each scoring event, the agent reflects on its picks vs SPY and writes a one-sentence lesson. Future picks read the last 5 as conditioning. The list below is the full memory.

How this works

Decision

Each market day at 22:30 UTC the agent pulls the top 10 cohort-ranked candidates. It enriches each with cohort_analyze (n=300 historical analogs, full forward distribution, conformal-calibrated bands), reads the last 5 lessons it wrote, and Claude Haiku picks 3 with conviction labels and 1-sentence reasoning.

Hold

Equal-weight 33% per position, 3 max. 5 trading days. No stops, no targets — the cohort signal either plays out or it doesn't. We're testing the signal, not the risk-management layer.

Score + reflect

At the 5-day mark each trade closes at the daily-bar close. Once all trades from a decision date are closed, the agent writes one short reflection — pattern noticed, bias flagged, or signal that worked. That lesson conditions future picks. The memory compounds.

Educational use; not financial advice. Paper trading without a broker; results assume you could buy and sell at the daily-bar close. Real execution would have slippage. Past pattern performance does not guarantee future results.