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.

This is the beginning of the README. Read the rest at the original. Read the full README ↗