Taurient Lite

English · 中文

An automated pre-market news brief for US equities, macro and AI/tech. Every trading day at 6:00 AM PT it scans global financial and tech news, merges duplicate stories, ranks them into three tiers, explains why each one matters, and publishes a web dashboard plus a Markdown archive.

It's inspired by Taurient, the in-house tool at the Australian hedge fund Minotaur Capital that uses several LLMs to scan about 35,000 news articles a week for investment ideas. Taurient Lite covers only the information-compression part of that. It doesn't pick stocks or give advice. Its job is to tell you what you need to know before the open in a few minutes of reading.

→ Live brief: taurient-lite-backend.onrender.com (The brief itself is written in Chinese. Free-tier hosting, so the first load may take a few seconds.)

What you get every morning

• Three-tier news: must read, worth knowing and background noise. Each item shows its publish date and a freshness badge.
• Deep metadata on top stories: an asset-impact matrix, a cross-source comparison and a timeline of how the story developed.
• Sector heatmap: one treemap per GICS sector. Tile size is market cap, tile color is the day's move, and sectors are ordered strongest to weakest. It shows where money is flowing today. A top-gainers list would mostly show micro-caps and single-stock news, while a treemap gives large companies the space they carry in the market.
• Momentum scanner: about 2,500 liquid US stocks are scanned each morning and put into four stages by how close each one is to where its move started (see below).
• Watchlist panel: tickers that appear in today's news are highlighted and link to their stories. Signed-in users can keep their own watchlist and filter the brief down to it.
• Stock pages: zoomable candlestick charts with draggable trendlines, company profile, key stats, analyst ratings, earnings and a raw headline feed.
• Calendar tab: upcoming events on a rolling weekly grid. Events without a confirmed date go in a separate "TBD" list instead of getting a made-up date.
• Indices and rates snapshot: S&P 500, Nasdaq, Dow, 10Y/30Y yields, WTI crude.

How it works

GitHub Actions ──► market data (Yahoo, FINRA, full-market momentum scan) ──► data/*.json
                                                                                  │
Claude Code routine ──► scan news, dedupe, tier, write "why it matters" ──► briefs/YYYY-MM-DD.json   ← single source of truth
                                                                                  │
taurient_lite/ (pure-stdlib renderer) ──► validate JSON ──► HTML dashboard + Markdown archive
                                                                                  │
backend/ (FastAPI on Render) ──► serves the latest brief with live quotes, stock pages, accounts

• Data and presentation are kept apart. The daily JSON is the only source of truth, and the web page and the Markdown archive are both generated from it. Restyling touches only theme.py. Changing news coverage touches only config.json.
• The rendering core has zero third-party dependencies. Dependencies flow strictly downward: schema → components → renderers → pipeline, so each layer can be tested on its own.
• Model judgment runs once a day. Numbers are always fresh. News triage needs an LLM, so it's a daily batch job. Prices don't, so the backend fetches live quotes and heatmap data (cached for 5 minutes). If a fetch fails, it falls back to the archived numbers and never returns a 502.
• Hard numbers come from primary sources: Fed dates from federalreserve.gov, earnings dates and filings from SEC EDGAR, short interest from FINRA.

More detail is in docs/architecture.md and docs/frontend-architecture.md (in Chinese).

Momentum scanner, and an honest backtest

News is a lagging signal. By the time a catalyst makes the headlines, much of the move has usually happened already. The scanner reads only price and volume (OHLCV) and asks how far into its move a stock is:

Stage	Meaning
Early	Broke a 50-day high within the last 3 days, on volume ≥ 2.5× the 20-day median
Continuation	Within two weeks of the breakout. The trend holds, but the best entry has passed
Base	Not broken out yet, range tightening. A watchlist candidate
Extended	> 35% above the 20-day MA, or > 10 days since the breakout, or doubled off its base

This is not a trading signal. The full backtest covered 2,515 liquid US stocks over two years and 5,090 early-stage signals:

• The win rate is barely better than a coin flip (about 51.5%). Expected value is +1.79% over 30 days, but the median is only +0.49%. The returns come almost entirely from a few big winners in the tail.
• The idea that the earliest signals are the best ones was disproved by the data: the "earliest" quartile averaged −0.19% over 20 days, and the "most extended" quartile averaged +1.92%.
• Take-profit rules make the win rate look better and the returns worse. A +5% take-profit raises the win rate to 73.6% but cuts expected value from +1.79% to +0.41%.
• Caveats: the test period (2024–2026) was mostly a bull market, the sample has survivorship bias, and the numbers are before transaction costs.

The stages are there to state facts about where a stock is in its move, not to predict returns. What the scanner gives you is time. It flagged BE as early stage at $23.92 on 2025-06-30, months before it became a headline. The stock is now around $258. The scanner didn't predict that. It saw the move in price and volume before it reached the news. The full tables are in README.zh-CN.md.

Quick start

git clone https://github.com/400-tang/taurient-lite.git
cd taurient-lite

# Re-render an existing brief (no dependencies needed)
python3 render.py 2026-09-18          # → site/index.html + briefs/2026-09-18.md

# Run the test suite (standard library only)
python3 -m unittest discover -s tests -t .

# Run the web backend locally
pip install -r backend/requirements.txt
uvicorn backend.server:app --reload   # → http://127.0.0.1:8000

Other entry points:

python3 scan_momentum.py                     # daily momentum scan over data/universe.txt
python3 scan_momentum.py --refresh-universe  # rebuild the liquid-stock universe (weekly)
python3 apply_momentum.py <date>             # merge scan results into a day's brief
python3 fetch_sectors.py                     # sector map + market caps for the heatmap
python3 fetch_short_interest.py              # official FINRA short interest
python3 import_watchlist.py                  # import a TradingView watchlist export

To generate a full brief with Claude Code: claude "Generate today's brief following RUNBOOK.md".

Project layout

taurient-lite/
├── RUNBOOK.md             Daily procedure the scheduled agent follows
├── config.json            Coverage, item targets, freshness limits, watchlist
├── taurient_lite/         Rendering pipeline (zero third-party deps)
│   ├── schema.py          Data model and validation
│   ├── components/        UI components
│   ├── theme.py           Design tokens and stylesheet (light and dark)
│   ├── momentum.py        Momentum stage logic (pure functions)
│   ├── treemap.py         Treemap layout (pure geometry)
│   └── ...
├── backend/               FastAPI service on Render (live quotes, stock pages, Supabase auth)
├── tests/                 381 tests, standard library only
├── briefs/                Daily JSON (source of truth) and Markdown archive
├── data/                  Market snapshots produced by GitHub Actions
└── .github/workflows/     Scheduled market-snapshot and momentum-scan jobs

Automation

• Claude Code cloud routine: runs on weekdays. The cron is 0 13,14 * * 1-5 (UTC), and a time gate in the prompt makes only the run after 6:00 AM Pacific do any work. That handles daylight-saving time without editing the cron twice a year.
• GitHub Actions: fetches market snapshots and runs the momentum scan before the routine starts. The routine's sandbox can't reach Yahoo but can reach GitHub, so the Actions runner fetches the data and commits it to the repo, and the routine picks it up when it clones.
• Render: deploys backend/ from render.yaml (New → Blueprint in the Render dashboard).

Design

The layout is modeled on a morning newsletter rather than a trading terminal: a single column about 65 characters wide, Newsreader for headlines, IBM Plex Sans for body text and IBM Plex Mono for every number. Cobalt blue is used for structure and interaction only. Red and green are reserved for market direction. The up/down colors were checked for color-blind readability. In dark mode, arrows, signed values and intensity bars mark direction as well as color.

Disclaimer

This is a news-summary tool, not investment advice. It makes no buy or sell recommendations and no price predictions. Every number in a brief has to trace back to a cited source, and numbers that can't be verified are left out.