WHAT NEW IN X10.31 SAT

X 上的 AI 合集档案 · 22 个合集 · 61,986 条

jevVIDEO

12 个 Claude Code agent 把 64 篇量化论文做成可交互实验室

用 12 个 Claude Code agent 复现 64 篇量化论文:Haiku 读 PDF、Opus 写策略代码、Jev 只做裁判给每个策略输出存活/死亡/信息不足的概率,并对比论文内、扣成本后、发表后三个数字。

xinces
@xinc3s14 天前收录原推发布于 2026.10.15
xinces @xinc3s原推发布于 2026.10.15
在 X 上打开原推

THIS IS GETTING F*CKING DANGEROUS: 12 CLAUDE CODE AGENTS TURNED 64 QUANT PAPERS INTO A LIVE LAB THAT DOES THE ONE THING THE PAPERS NEVER DID every strategy from every paper, rebuilt in your browser. pick one, change the costs, the dates, the universe, and watch the edge move in real time each seat got one job: → HAIKU 5.5 reads all 64 pdfs: signal, universe, window, rebalance, costs. a page number on every line, gaps marked "unknown" → OPUS 5.5 turns each paper into code. one isolated agent per strategy, all on the same price data → JEV is the judge. it can't write a single word. it returns one probability per strategy: survives, dies, or not enough info → auditors rerun each paper's own main table. if the numbers don't match, it stays red → ui agents build one panel per paper: the claim, the rules, the deviation, what's still unchecked here's why it's dangerous. not for you. for the papers the biggest replication study tested 447 published anomalies. 64% weren't even significant once microcaps were cut and stocks weighted by size and the ones that work decay. across 97 predictors, returns were ~26% lower out of sample and ~58% lower after the paper came out so the lab shows 3 numbers for every strategy: in the paper. after costs. after the publication date the third one never makes it into the abstract every paper sells you the edge. the date on the cover tells you when it started dying paste this into claude code with your pdfs ↓ "Turn these papers into an interactive strategy lab. 1. Haiku reads each PDF: signal, universe, dates, rebalance, costs. Cite page numbers. Anything missing is 'unknown'. 2. One isolated agent per strategy builds it in code. All share one price dataset and one backtest engine. 3. Auditors rerun each paper's main table. A mismatch beyond tolerance is a red flag, never hidden. 4. Jev scores each strategy: survives / dies / not enough info. Thresholds live in code. 5. For each strategy show: the paper's result, after realistic costs, and after the publication date. 6. Read-only. No brokerage, no orders. Start with one paper and show me the plan before anything runs."

这条是引用着另一条写的:x.com/xinc3s/status/2108204467774509059

x.com/xinc3s/status/2110535607877308528采集于 2026.10.16 10:06