dsh-ai-hedge-fund — an AI hedge fund of investor legends, on DeepSeek Harness
Buffett, Munger, Lynch and Wood value your ticker independently — the PM aggregates, names who moved the verdict, and records dissent. Every claim carries an evidence tag. 让巴菲特、芒格、林奇、木头姐各自独立估值你的标的——PM 聚合裁决、点名动因、 异议在案。每个主张带证据标签,输出可审计的 IC-MEMO。
Inspired by virattt/ai-hedge-fund — restructured as a native DeepSeek Harness experience. Not affiliated; independent community project. Sister project of dsh-TradingAgents (the bull/bear debate committee).
Why this exists
| ai-hedge-fund | dsh-ai-hedge-fund | |
|---|---|---|
| Runs where | Python lab | Your DSH terminal — the model key you already have |
| Calls/session | 10+ | 7 (≈40s) |
| Personas see each other | sequential agents | independent valuations, blind to each other — no herd drift |
| Output | trade tables | IC-MEMO — auditable open format |
| Dissent | — | recorded, never deleted — PM must name unrebutted objections |
| Abstention | — | PASS is first-class — “证据不足,拒绝表态” |
Quick start (zero extra keys)
git clone https://github.com/vicfei/dsh-ai-hedge-fund && cd dsh-ai-hedge-fund
python3 scripts/hedge_fund.py --ticker DEMO # offline smoke, ~40s
python3 scripts/fetch_card.py --ticker NVDA # real US data card
python3 scripts/hedge_fund.py --ticker NVDA --data cards/NVDA.json
First smoke verdict (2026-11-16, temperature=0) — the value/growth split on synthetic data:
| 人格 | 立场 | 信念 |
|---|---|---|
| 沃伦·巴菲特 | PASS | 70 |
| 查理·芒格 | PASS | 72 |
| 彼得·林奇 | HOLD | 55 |
| 凯瑟琳·伍德 | HOLD | 58 |
PM: PASS @ 68 — "巴菲特与芒格关于转换成本与定价权的论点压制了木头姐的 '降价即赛道被验证';两位 HOLD 均为等待验证而非看多……" 而林奇与木头姐的 异议被完整记录在案("可能错过 GARP 机会"/"低估飞轮期权价值")。
The fund
analyst ──► 4 personas value INDEPENDENTLY (blind, method-faithful)
│
risk guardrails (constraints only)
│
portfolio manager ──► IC-MEMO
(names its influences; dissent recorded)
Persona packs are config: value wing (Buffett/Munger) vs growth wing (Lynch/Wood)
by design — real disagreement is the product. --personas selects a subset.
Status & roadmap
v0 engine smoke-tested (see memos/). Real-data adapters (US & A-share, zero keys,
CN-cloud friendly ladder) inherited from the sister project. See ROADMAP.
Credits & disclaimer
Inspired by virattt/ai-hedge-fund — independent native re-implementation, MIT. Not investment advice. Demo data is synthetic; real-data modes cite sources.
No comments yet. Be the first to write one.