stockproof · 有据

An LLM reads stock news, earnings reports and company announcements for you, with every point linked to its source. A research toolkit then checks each signal against years of prices, and the app tells you what that check found. US stocks, China A-shares and Hong Kong. English and 中文.

· 中文说明

Try it in 30 seconds. The demo runs offline on bundled real data and needs no API key:

pip install "git+https://github.com/taoqizhang55-code/stockproof"
stockproof demo

What it does

Daily stock watching

Screen and scorecards

Research and the quant lab

Your portfolio, and AI agents

Everything runs in one local web page (stockproof dashboard). The Home tab shows what you can do, with clickable examples; the Lab tab backtests ideas from a form, no JSON needed; the Portfolio tab checks up your holdings and remembers them for next time. The first time you use stockproof in DeepSeek Harness, the agent lists everything it can do and suggests trying the dashboard; after that, ask "what can you do?", type /stockproof for ready-made requests, or say "open the dashboard". Say "open the dashboard's lab tab" to jump to a page. Running stockproof with no arguments prints the same grouped menu, and in DeepSeek Harness typing / lists ready-made prompts (check up a portfolio, backtest an idea, why did it move…).

What an LLM can't do alone — and can with stockproof

Ask DeepSeek Harness "A 股业绩大超预期的公司，财报公布后第二天买入持有 20 天，扣费后能跑赢吗？". The model turns the question into a spec, the lab runs it on ten years of point-in-time A-share data, and a report opens in your browser. The answer was no: −3.6% a year (t = −2.0), positive in 3 of 4 years before 2020 and negative every year from 2020 to 2024. Ask for a variant and the agent reports that the significance bar went up with the number of tries.

In DeepSeek Harness, type /stockproof for ten ready-made requests:

Related MCP server: Stock SDK MCP Server

What the research found

Run on the author's data (2016–2026). Each result is a click away in the Research tab, with t-statistics and caveats.

The pattern holds across all of these: public information is priced fast. So stockproof helps you read and check evidence faster. It does not hand you a trading signal, and the app says this wherever it shows a score.

Principles

• Cited, not invented. The LLM may only use the numbered items it is given. Uncited points are dropped, and "unexplained" is an allowed answer.

• Scores carry their track record. ✓ useful in tests · tested, not significant · ? not tested yet.

• Point-in-time. EPS as first reported, real announcement dates, entry after the news. No peeking ahead.

• Any LLM. DeepSeek (default), OpenAI, Qwen, a local Ollama model, or any OpenAI-compatible API. Each item is read once and cached in SQLite, so daily use costs cents.

• No paid data. SEC EDGAR, Yahoo/Google News RSS, Nasdaq, and Eastmoney / Sina / CCTV / Shenwan via AKShare.

• Local and private. The dashboard listens on 127.0.0.1 only and refuses requests from other websites. API keys are read from the environment and never written to disk.

Install

Python 3.10+ (3.11+ for China markets).

pip install "stockproof[cn,mcp] @ git+https://github.com/taoqizhang55-code/stockproof"
export DEEPSEEK_API_KEY="sk-..."                              # or another provider: stockproof providers
export STOCKPROOF_SEC_CONTACT="Your Name you@example.com"     # US earnings: SEC asks for a contact
stockproof dashboard

Add stocks in the page (AAPL, 600519, 300750, 00700 …) and click Refresh news / Refresh earnings.

Commands

stockproof dashboard                         # everything in one local page
stockproof scan AAPL 600519 00700 --lang zh  # news mood report
stockproof earnings AAPL MSFT --open         # latest releases read by the LLM, with SUE
stockproof screen --top 30                   # whole-market A-share screen
stockproof screen --backtest --skip-screen   # test the screening rule on history
stockproof scorecard                         # evidence cards for the screen's top 10
stockproof scorecard 600519 NVDA             # …or for any stocks
stockproof research --market cn --open       # earnings event study + backtest (us / cn)
stockproof policy                            # 新闻联播 policy study (first run: a few hours of downloads)
stockproof daily                             # one refresh of the watchlist + daily digest
stockproof schedule --time 21:00             # run `daily` every evening (macOS)
stockproof calendar                          # who reports in the next two weeks
stockproof mcp --install-dsh                 # add to DeepSeek Harness on this computer (desktop app too)
stockproof lab caps                          # quant lab: what can be tested here
stockproof lab run '{"type":"earnings","filters":[{"field":"sue","op":">=","value":2}],"hold_days":20}' --open
stockproof lab confirm <id>                  # one shot on the locked holdout
stockproof portfolio "600519 100, 00700 300, AAPL 50"   # portfolio check-up
stockproof demo                              # offline demo, no key

Chinese codes can be typed in any common form: 600519, 600519.SS, 300750, 920001, 700, 0700.HK.

How the pieces work

A feed for MRK also returns stories about the German Merck KGaA, a feed for GE returns GE HealthCare news, and NVDA appears in every "stock market today" article (in one scan only 2 of 14 headlines were really about Nvidia). The model gives each headline a relevance score (0–1) and a sentiment score (−1 to +1), with a reason. Off-topic stories stay visible but are not counted. When several headlines cover the same event, the model marks them as duplicates, so the event counts once.

SUE = (this quarter's EPS − the same quarter last year) / standard deviation of the previous 8 such changes, using EPS as first reported. US quarters come from SEC XBRL, and Q4 = full year − Q1–Q3. A-share reports are year-to-date totals, so single quarters are derived by subtraction. Announcement dates come from the 8-K Item 2.02 filing (after 16:00 New York time counts as the next day) or from Eastmoney's 预约披露 actual disclosure date.

• Screen: fundamentals for all ~5,200 A-shares come from Eastmoney's 业绩报表 tables (one download per report period).

  • Filters: no ST stocks, market value ≥ 50亿, at least 2 years listed.

  • Score: the average percentile of five factors.

• Scorecard: six dimensions plus industry policy, with one LLM call per stock per day.

  • Earnings quality asks whether growth comes from the core business and is backed by cash, or from one-offs.

  • Announcements count each kind of event once, since one buyback can produce ten filings.

• Entry and holding: entry at the close after the reaction (earnings) or at the next open (policy), with capital split into slots and empty slots holding the market.

• A-share costs and trading rules:

  • Commission 0.025% per side, slippage 0.1% per side.

  • Stamp duty on sales: 0.1%, or 0.05% from 2023-08-28.

  • Price limits are respected, at 10% / 20% / 30% depending on the board.

• Benchmark: the equal-weighted average of the stocks studied, because today's CSI 300 members beat the index by 13% a year (survivorship).

Data sources and limits

• Sources: SEC EDGAR, Nasdaq, Yahoo/Google News RSS, and Eastmoney, Sina, CCTV and Shenwan via AKShare.

  • These are free, public, unofficial endpoints. They can change or rate-limit you.

  • stockproof slows down, retries, and stops with a clear message when a source refuses.

  • tools/*_probe.py checks which sources work from your network.

• LLM errors: models can misread. Every point links to the original, so check it.

• Not investment advice. The research above is part of the reason this tool doesn't give buy or sell calls.

Development

pip install -e ".[dev,mcp]"
pytest              # ~170 offline tests: fake feeds, fake LLM, fake AKShare; no key or network needed
python tools/make_demo.py   # rebuild the demo bundle from your own stockproof_data

MIT License.