Daisy Financial Research — Autonomous Stock / Company Research Skill

中文 | GitHub | Releases

What it does

A multi-platform agent skill for finance research. Given a stock/company/sector topic, it plans the research, pulls structured data from Tushare, searches the web via Brave / Bailian MCP, runs Python for math and valuation, and produces a sourced, reproducible Markdown + HTML (+ optional PDF) report.

Design borrows from virattt/dexter — iterative agent loop (plan → gather → validate → answer) — but packaged as a cross-platform skill, no separate CLI.

Key capabilities:

• Agent-native CLI (v2.1.0+): every script auto-emits a stable {ok, data, meta} JSON envelope when stdout isn't a TTY, supports --schema introspection and --dry-run, with documented exit codes 0–5.
• Cross-session decision memory (v2.2.0+): append-only Markdown log with pending → resolved lifecycle, atomic rewrites, win-rate / mean-alpha stats. Format wire-compatible with TradingAgents' memory.py.
• AKShare HK fallback (v2.3.0+): closes the documented pro.hk_daily_basic Tushare gap with PE/PB/PS snapshots and ROE/EPS/BPS time series — no Tushare token, lazy-imported.
• Borrowed prompt library (v2.4.0+): five reference docs adapted from TradingAgents — Bull / Bear / Synthesis debate, Aggressive / Conservative / Neutral risk debate, reflection prompt, decision schema, China-market analyst framing, technical-indicator cheatsheet. See references/.
• Auto-resolve workflow (v2.5.0+): dexter_memory_log.py auto-resolve fetches close[decision_date] and close[as_of_date] for the ticker plus the right benchmark (CSI 300 for *.SH/SZ/BJ, HSI for *.HK with AKShare Sina fallback, SPY for US via yfinance), computes raw + alpha + holding days, then resolves the pending memory-log entry in one call.
• Technical indicators + decision-level backtest + tighter agent loop (v2.6.0+):
  • scripts/technical_indicators.py — point-in-time SMA / EMA / MACD / RSI / Bollinger / ATR / VWMA via stockstats, with a strict look-ahead-bias guard (rows with Date > --as-of are dropped before computation). Auto-routes by ts_code suffix to pro.daily / pro.hk_daily / yfinance.download.
  • references/data-source-routing.md — canonical (market × data type) routing table consolidating A-share / HK / US primary calls and documented fallback chains. references/hk-ticker-name.json is a curated 5-digit → Chinese-name dict for ~30 HK majors; akshare_hk_valuation.py name --ts-code <code> is a zero-API local lookup.
  • dexter_memory_log.py backtest — risk-adjusted decision-level metrics over a --from / --to window: per-rating mean alpha, hit rate, alpha_t_stat, annualized_alpha_pct, Sortino-flavored ratio, plus the cumulative-alpha curve and its max drawdown. Deliberately not called Sharpe — the log records discrete decisions, not a continuous NAV.
  • dexter_memory_log.py record --rating now tolerates LLM-formatted input — accepts a canonical word, a markdown bold tag (**Rating**: Buy), or a full Portfolio Manager synthesis paragraph; rejects input with no 5-tier word loudly instead of silently defaulting to Hold.
  • dexter_scratchpad.py can-call <pad> <tool> <query> — soft loop-limit + similarity guard ported from virattt/dexter. Always allows (returns allowed: true) but emits a warning when the same tool has already been called >= --max-calls times in the pad, or when the new query is textually similar to a prior call (stdlib difflib, no embedding deps).
• Plan-first workflow with JSONL scratchpad recording every tool call, params, result, assumption.
• DCF valuation with sensitivity matrix and sanity checks.
• Bank / financial-sector valuation override (RoTE / CET1 / NIM / P/B / payout) instead of forcing DCF on the wrong frame.
• A-share + Hong Kong Stock Connect screening presets (dividend-quality, value, momentum, etc.).
• Three-layer report output (md → html → optional pdf), CSS already handles CN/EN font fallback.
• Brave MCP + Bailian WebSearch MCP for web context.

Multi-Platform Support

Prerequisites

# Python 3.9+
pip install tushare pandas requests
# Optional: AKShare HK valuation/fundamentals fallback (no Tushare token needed)
pip install akshare
# Optional: PDF output
brew install pandoc
brew install --cask mactex      # or basictex for a smaller install

Environment:

export TUSHARE_TOKEN=xxxxxxxx   # required for any Tushare call

Installation

Quick Start

# A-share dividend-quality watchlist + Markdown report draft
python <skill-dir>/scripts/screen_a_share.py --preset a_dividend_quality --top 50 --report

# Render the Markdown draft into the three-layer report
python <skill-dir>/scripts/financial_report.py ./financial-research/reports/<TIMESTAMP>_a-share-a_dividend_quality-screen.md \
    --title "A-share dividend watchlist" --slug a-div-quality --pdf

All output lands under ./financial-research/{reports,watchlists,scratchpad,universes,memory}/ in your cwd by default.

Agent-native CLI

Every script under scripts/ follows a uniform contract designed for both humans at a terminal and agents calling via subprocess. Same shape, eight scripts:

# Discover the script's parameter and output schema (preferred over --help for agents)
python <skill-dir>/scripts/screen_a_share.py --schema

# Preview the request shape — no Tushare call, no file written
python <skill-dir>/scripts/screen_a_share.py --preset a_value --dry-run

# Force JSON regardless of TTY state
DAISY_FORCE_JSON=1 python <skill-dir>/scripts/screen_a_share.py --preset a_value

Output auto-detection: when stdout is not a TTY, scripts emit a single JSON envelope. When stdout is a TTY, scripts emit the legacy human table. Override with --format json|table.

Success envelope:

{
  "ok": true,
  "data": { "trade_date": "20260430", "candidates": 50, "csv": "...", "preview": [...] },
  "meta": { "schema_version": "1.0.0", "request_id": "req_abc123", "latency_ms": 412 }
}

Error envelope:

{
  "ok": false,
  "error": { "code": "no_data", "message": "...", "retryable": true, "context": {...} },
  "meta": { ... }
}

Exit codes: 0 ok · 1 runtime · 2 auth · 3 validation · 4 no_data · 5 dependency.

Long-running operations (screen_hk_connect.py --with-momentum, financial_report.py) emit NDJSON progress events on stderr, one JSON line per phase, so an agent can detect liveness without blocking on stdout.

Output paths

Every script accepts --out-dir <root> to override the root; the subdir is appended automatically.

Hermes users: to keep the legacy ~/.hermes/reports/financial-research/<subdir>/ layout, pass --out-dir ~/.hermes/reports/financial-research to every script.

Development with uv

pyproject.toml lists the runtime dependencies. Reproduce the env locally:

uv sync                  # core: tushare / pandas / numpy / requests
uv sync --extra akshare  # also: akshare (HK valuation + HSI benchmark fallback)
uv sync --extra us       # also: yfinance (US tickers in auto-resolve / technical_indicators)
uv sync --extra ta       # also: stockstats (technical_indicators.py)
uv sync --all-extras     # everything

Auto-update

The skill checks <skill-dir>/.last_update on first use per conversation. If older than 24 hours, it silently runs git pull --ff-only. Failures (offline, conflict, not a git checkout) are ignored without interrupting the workflow.

Manual update:

cd <skill-dir> && git pull

vs no skill

Disclaimer

This skill produces data analysis and research records, not investment advice. All conclusions require independent judgement against the latest public information.

Support

If this skill helps you, consider supporting the author:

WeChat Pay

Alipay

Buy Me a Coffee

Give a Reward

Author

• Bilibili: https://space.bilibili.com/1107534197
• GitHub: https://github.com/Agents365-ai