Hedge-Fund-Grade Stock Analysis In Plain English

One researcher open sourced the kind of chart-reading AI that Wall Street's systematic funds pay fortunes for. Wrapped in a Claude skill, you can run hedge-fund-grade technical analysis with one prompt in plain English, then turn the prediction into a newspaper-style tearsheet. Three repos, one stack. Research workflow only, not financial advice.

What You'll Learn

• What Wall Street actually pays for: the open source AI trained on 12 billion candlesticks from 45 exchanges that reads charts the way ChatGPT reads English
• Why Claude and ChatGPT cannot do this on their own: they were trained on text, not market data, so anything technical they give you is guessed
• The Claude skill that wraps Kronos: install it once and run /kronos AAPL 1d from any Claude conversation in plain English
• The Kronos Ledger: a one-page newspaper-style tearsheet with forecast, technical indicators, SEC EDGAR fundamentals, and a top-line house call
• How to read the output: the confidence interval is the signal nobody explains, plus the routine that keeps the stack from turning into a gambling habit

What This Guide Is (And Isn't)

Wall Street wins because they have AI that actually reads price action. Until recently, that AI was either proprietary, gated behind eight-figure data licenses, or both. One researcher quietly open sourced a model that does the same job. This lesson walks you through what it is, how to install the Claude skill that wraps it, how to layer in The Kronos Ledger to turn the prediction into a publishable report, and how to pair the whole thing with Claude on the fundamentals side. This is a research workflow, not a trading system, and not financial advice. Stocks carry real risk. All three repos are MIT or Apache licensed and free.

The operator frame: the point of the stack is to make you slower and more grounded, not faster and more impulsive. If you find yourself using it to chase moves, you are using the wrong tool. Reset.

What Wall Street Actually Uses (And Why Claude Cannot Do This)

Renaissance Technologies, Two Sigma, Citadel, and the rest of the systematic side of Wall Street do not pick stocks by reading articles. They run ML models over OHLCV data. Their edge is that they have models trained to read price action directly. Until recently, you could not get one of these models without an eight-figure budget and a team of quants.

Why ChatGPT and Claude cannot fill the gap: large language models learn from text. They never saw raw OHLCV data, candlestick patterns, or order flow. When you ask them to evaluate a chart, they pattern match against text descriptions of charts they read in articles. They are not reading the chart. They sound confident because they are great at structure. They are guessing because the chart side was never actually evaluated.

The tell: every viral "analyze any stock in 5 minutes with Claude" video lands on the same three-prompt structure. The output reads rigorous because Claude is great at structure. It reads shallow because the technical layer was never actually evaluated.

Meet Kronos

Kronos is an open source foundation model from researcher Shiyu Chen. It treats K-line data (the open, high, low, close, and volume of every candlestick) as its native language. Instead of running text over price action, it tokenizes the price action itself and runs a transformer over the structure. It reads charts the way ChatGPT reads English.

The numbers that matter: trained on more than 12 billion candlestick records from 45 exchanges, covering equities, indices, and crypto. Over 27,000 GitHub stars. MIT licensed, free for commercial use. There is a live Bitcoin demo, a fine-tuning pipeline, and a local webui you can run on your laptop. On the published benchmark, it outperformed every prior-generation prediction model across multiple horizons.

Kronos repo: https://github.com/shiyu-coder/Kronos (you do not interact with this directly; the Claude skill in the next section pulls it down for you).

What it is not: Kronos is not a trading bot, not a guaranteed signal, and not a license to skip risk management. It is a model that outputs a price-move prediction with a confidence interval. Treat the output as one input into a research process, the same way a quant team would.

The Claude Skill: One Prompt, Plain English

Running Kronos in raw Python is fine, but it is faster and more honest to wrap it in a Claude skill so you can ask in normal language and get the technical read back in plain English. The skill pulls the OHLCV for any ticker, runs Kronos against it, and returns the prediction plus the confidence interval as text Claude can reason over.

Claude skill repo: https://github.com/coopersimson96/kronos-claude-skill

How to install:

# Clone the skill into your Claude skills directory (one time)
git clone https://github.com/coopersimson96/kronos-claude-skill ~/.claude/skills/kronos

# Confirm Claude sees it
claude

# In any Claude conversation, try:
# run kronos on AAPL
# /kronos NVDA 1y 1d 30
# /kronos BTC-USD 3mo 4h 48

The first run auto-installs Kronos plus a venv, torch, transformers, and yfinance. It takes 3 to 7 minutes. Every subsequent call is 5 to 15 seconds.

How to read the output: a narrow confidence interval means the model is relatively certain across its forward distribution. A wide interval means the model is hedging. When the interval is wide, the prediction is noise. Ignore it. Treat single forward predictions as low confidence and look for the same direction across multiple timeframes before you give the signal any weight.

The interval rule: narrow band across at least two timeframes = look. Wide band = ignore. A single prediction on a single timeframe is not a signal, it is a starting point for further research.

The Kronos Ledger: Turn The Prediction Into A Real Report

The Claude skill gives you a forecast and a confidence band as plain text. The Kronos Ledger turns that raw output into a one-page, newspaper-style tearsheet you can scan in 60 seconds. Forecast chart up top, technicals dashboard below (Bollinger Bands, SMA 50, SMA 200, RSI, MACD, ATR, volume), fundamentals pulled live from SEC EDGAR (latest 10-K and 10-Q), a news cycle block with sourced bullets, and a top-line BUY / SELL / HOLD house call with the reasoning underneath. Everything lives in a single self-contained index.html file, so you can fork it, point it at any ticker, and re-deploy in minutes.

• Live demo (NVDA): https://kronos-nvda.vercel.app (open it on your phone for the reel-mode layout)
• Ledger repo: https://github.com/coopersimson96/kronos-nvda-tearsheet

How to install and run on your ticker:

# Clone the Ledger
git clone https://github.com/coopersimson96/kronos-nvda-tearsheet ~/Desktop/kronos-ledger
cd ~/Desktop/kronos-ledger

# Open the README for the data refresh commands
open README.md

# Re-generate the Kronos forecast that drives the chart
python3 ~/.claude/skills/kronos/scripts/run_kronos.py <TICKER> 1y 1d 30

# Then re-skin index.html with the new ticker (find/replace NVDA + paste new data)
# Or deploy as-is with the Vercel CLI (use the production flag from the Vercel docs)
vercel

The Ledger frame: the Ledger is the output layer. It does not replace your decision-making, it makes the existing research legible. One page, one ticker, one house call you can disagree with. Print it before earnings if you want to be honest with yourself about why you owned it.

Layer 2: Claude For The Fundamental Side

Kronos handles the chart. Claude handles everything else. The key is to lock Claude down so it only operates on the company's own documents and stops trying to use its general training to speculate.

The Project setup: in Claude, create a new Project for each ticker. In the Project instructions, give Claude a single grounding rule: only use the documents in the knowledge base, do not use outside knowledge, do not speculate, and if information is missing say so explicitly.

Then go to the company's investor relations page. Download the last 5 years of annual reports, the most recent quarterly earnings, every earnings call transcript, and any current investor slide decks. Drop all of it into the Project. You now have an AI analyst who knows exactly the company you are evaluating and nothing more.

The two memos that earn the work: ask Claude to write two short memos. One bull case for owning this stock, one bear case for losing money on it. Save both. When the next quarter drops, you re-read them. Did the company deliver what the bull case assumed, did anything in the bear case get worse, what would invalidate either side.

The grounding instruction: without the "only use provided documents, do not speculate" rule, Claude will pattern match against its general training and the memos will quietly hallucinate. Lock this in the Project instructions, not the chat.

The Weekly Routine

The stack only matters if you actually run it on a schedule. The routine that turns the tools into a real research system:

1. Pick a watchlist of 5 to 10 tickers. Any more and you cannot keep up with the quarterly refresh. Start narrow.
2. Run the Kronos skill on each ticker weekly. Ask Claude in plain English. Note the prediction direction and the confidence interval. Flag the ones with narrow bands across multiple timeframes.
3. Generate a Ledger for any flagged ticker. Re-skin index.html with the new ticker symbol and run the data refresh commands. You now have a one-page tearsheet for that name.
4. Open the Claude Project for the same ticker. Re-read the bull case and bear case memos. Ask Claude what would invalidate the current bear case. That answer is your exit thesis.
5. Quarterly earnings refresh. When the new 10-Q drops, paste it into the Project. Ask Claude: did management deliver, did the numbers improve, did anything in the bear case get worse. Update the memos and re-publish the Ledger.

The filter vs amplifier test: if the stack is filtering you out of tickers that did not deserve attention, it is working. If you find yourself overriding the stack because a stock is running, the stack is no longer working. Reset.

Go Deeper

• Printable PDF version of this guide: https://drive.google.com/file/d/1BXp68CDNfNkTN0aWhz-2SSknZt7f3Ng5/view?usp=drivesdk

Want the whole system installed live, with Cooper on the call? Build Your Own AI, the two week sprint runs the first two weeks of August.

Or get a new working tool every Monday inside The AI Operator community on Whop.