Milestone: my open-source trading platform fa-quant got its first strategy through all 8 statistical checks and into paper trading.
20 years of ETF data. Out-of-sample Sharpe 0.70, max drawdown 13.7%.
pip install fa-quant
rl-pricing v2, day one:
- 5 modern RL agents: BC, IQL, CQL, Decision Transformer, CVaR-PPO
- point-in-time SEC filings and news, lookahead-checked in CI
- R5 pre-registered: re-testing FinRL-DeepSeek
Can they beat equal weight? CI is scoring them now.
github.com/BichengWang/RL…
New in ai-decision-system: jev-decision-eval now calibrates model confidence. Fit isotonic or Platt on one labeled set, gate on held-out cases, and keep both runs as records anyone can verify.
Write-up:
I built an open benchmark for AI coding-agent teams. First run, 48 GitHub issues filed after the model's cutoff:
coder+reviewer: 29/48 (60%), $0.10/issue
3 coders+judge: 28/48
1 agent: 27/48
Every patch, log and grade is public.
medium.com/p/181d76eee1c0
I just launched RL-pricing, an open lab that re-tests trading-AI claims.
Live today:
- rl-pricing 1.0 on PyPI
- a 4-task benchmark, 58 entries
- forward scoring on new data only
- a live LLM ledger, committed before each US open
medium.com/p/81b449ec45a5
Our deep RL trading paper reported 55.5% a year. I re-ran it 75 times with next-open fills, fees and 5 seeds per agent. None reached it.
Best re-run: 46% a year. Holding the same 20 stocks in equal weight: 29.7%.