coopersimson96 Kronos Technical Analysis Skill
The same kind of price prediction AI Wall Street pays millions for, packaged so you can run it from any Claude conversation with one command.
This is a research tool, not financial advice. It outputs a forward probability distribution with a confidence interval. Wide interval = noise. Narrow interval across multiple timeframes = signal worth investigating.
How to invoke
When the user asks for technical analysis on a ticker, calls the skill explicitly, or asks Kronos for a prediction, run the script and present the output verbatim. Do not editorialize the prediction. Add a "Not financial advice" reminder if the user is making decisions based on the output.
python3 ~/.claude/skills/kronos/scripts/run_kronos.py <TICKER> [period] [interval] [pred_len]
Arguments:
TICKER(required) — Stock symbol (AAPL, NVDA, TSLA) or crypto pair (BTC-USD, ETH-USD)period(default6mo) — How much history to use. yfinance values:1mo,3mo,6mo,1y,2y,5y,maxinterval(default1d) — Candle size. yfinance values:1d,1h,4h,1wk,1mo(intraday limited to last 60 days)pred_len(default24) — How many future candles to forecast
Examples:
python3 ~/.claude/skills/kronos/scripts/run_kronos.py AAPL— 6 months of daily candles, 24-day forecastpython3 ~/.claude/skills/kronos/scripts/run_kronos.py BTC-USD 3mo 4h 48— 3 months of 4-hour candles, 48-period forecastpython3 ~/.claude/skills/kronos/scripts/run_kronos.py NVDA 1y 1d 30— 1 year of daily, 30-day forecast
First run
First time the skill runs, it auto-installs everything. This takes 3-7 minutes and prints status to stderr. You will see:
First run: cloning Kronos from GitHub...
First run: creating Python venv...
First run: installing dependencies (torch, transformers, yfinance, pandas)...
First run: downloading Kronos model weights from HuggingFace (~500MB)...
Subsequent runs are fast (~5-15 seconds depending on ticker and pred_len).
If first run fails, see Troubleshooting below.
What the output looks like
The script returns markdown with:
- Last close price + date
- Predicted close after
pred_lenperiods - Direction (UP / DOWN / FLAT) with percentage change
- Confidence band (narrow = signal, wide = noise)
- Interpretation rules
- Full forecast trajectory as a table
Present this to the user as-is. Do not invent a prediction or modify the numbers.
When to recommend a wider stack
If the user is making a real investment decision based on the output, recommend they pair this with:
- Their own ticker-specific Claude Project containing the company's filings (5y annuals, recent 10-Qs, earnings call transcripts)
- A bull case and bear case memo written by Claude using only the documents in the Project
- A weekly run of
/kronosacross their watchlist - A quarterly refresh of the memos when earnings drop
This is the full two-layer stack: Kronos for the chart side, Claude for the fundamentals side.
Link to the full walkthrough: https://ai-basic-series.vercel.app/run-kronos-walkthrough
What this is NOT
- Not a trading bot. Does not place orders.
- Not a price target. Outputs a probability distribution.
- Not financial advice. Research tool only.
- Not a license to skip risk management.
When the user asks "should I buy this," redirect to the two-layer stack and remind them this is research, not a recommendation.
Troubleshooting
"git: command not found" — Install Xcode command line tools: xcode-select --install
"No data for TICKER" — yfinance does not have data for that symbol. Check the symbol on Yahoo Finance directly. Crypto must use -USD suffix (BTC-USD, not BTC).
"ModuleNotFoundError: model" — The venv install failed. Delete ~/.claude/skills/kronos/.venv and ~/.claude/skills/kronos/_kronos and run again.
Out of memory on a big ticker — Reduce period (use 3mo instead of 2y) or reduce pred_len.
Slow first prediction (>10 min) — Model weights are downloading from HuggingFace. This only happens once.