Kronos Technical Analysis Skill
Run hedge-fund-grade technical analysis on any stock or crypto ticker from any Claude conversation — one command returns a price prediction with a confidence interval.
Data flow is largely transparent: public OHLCV via yfinance, first run auto-clones GitHub and downloads ~500MB weights from HuggingFace. Positives: no malicious instructions, no credential collection, explicit 'not financial advice'. Deducted for: automatic git clone + pip install + model download on first run with no user confirmation step; no hash verification of third-party model weights; unpinned dependencies.
Docs are self-consistent; parameters, examples and troubleshooting (missing git, no data, failed venv, OOM) are covered, with failure paths explained. Deducted for: the actual script is not present in evidence, so auto-install logic and error feedback cannot be verified; no tests; failure feedback relies on prose only.
Triggers, arguments, examples and 'What this is NOT' boundaries are fairly clear, including the crypto -USD suffix detail. Deducted for: complete dependence on yfinance (Yahoo) and HuggingFace downloads, which are poorly reachable from mainland China networks; no Chinese-language support stated; triggers are English-only.
Docs are reasonably layered (usage/output/troubleshooting in SKILL.md) and README credits the upstream MIT-licensed model. Deducted for: no license metadata for the skill itself, no version/changelog, maintenance responsibility is a personal attribution only, and first-run auto-install carries hidden assumptions.
Claims markdown prediction + confidence interval output that can be presented directly; framing is sensible. Deducted for: static review cannot verify output quality or usability; sample output uses placeholders ($XXX.XX); marginal value over calling Kronos directly is mostly packaging convenience.
Auditable links to the upstream model and data source; sources credited in README. Deducted for: no tests, no CI evidence, no real execution records; marketing claims like 'hedge-fund-grade' are unverifiable; facts and inference are not separated.
- First run automatically clones a third-party repo, installs unpinned dependencies (including large torch) and downloads ~500MB of model weights; review and confirm before running, ideally in an isolated environment.
- Core function fully depends on yfinance and HuggingFace, which may be unreachable from mainland-China networks without workarounds.
- Model output is a probability distribution, not a price target; do not use it as investment advice; 'hedge-fund-grade' claims are marketing language.
- The repo provides no license for the skill itself; confirm terms before commercial use or redistribution.
What does this skill do, and when should you use it?
This skill packages the open source Kronos foundation model (trained on 12B candlesticks from 45 exchanges) as a Claude skill. Say something like "run kronos on AAPL" in a conversation and it pulls recent OHLCV data via yfinance, runs Kronos for a forward prediction, and returns plain-English markdown with the predicted close, direction, percentage change, and a confidence interval. It is explicitly a research tool, not financial advice. First run auto-installs dependencies and downloads roughly 500MB of model weights.
Fetches historical OHLCV data for a stock symbol or crypto pair (e.g. AAPL, BTC-USD) via yfinance; runs the Kronos foundation model inside a Python virtual environment to produce a forward forecast; and returns markdown containing the last close, predicted close after pred_len periods, direction with percentage change, a confidence band (narrow = signal, wide = noise), and the full forecast trajectory table. History period (1mo to max), candle interval (1d/1h/4h/1wk/1mo), and forecast length are all configurable.
- A retail investor who wants a quick technical-side prediction on a stock like NVDA inside Claude without setting up any ML tooling themselves
- A crypto trader forecasting short-horizon BTC-USD moves on 4-hour candles
- A researcher running /kronos weekly across a watchlist as one input into a broader research process
- An advanced user pairing Kronos chart predictions with a Claude Project containing company filings — the two-layer stack for chart-side plus fundamentals analysis
- A conversational user who states a forecast horizon in plain English (e.g. "next 30 days") and lets Claude infer the arguments
What are this skill's strengths and limitations?
- Wraps a large-scale trained open source foundation model into a one-phrase skill — no ML background needed
- Fully automated first-run setup; fast (~5-15 seconds) on subsequent runs
- Output includes confidence bands and interpretation rules to distinguish signal from noise
- Supports both stocks and crypto with flexible period, interval, and forecast-length parameters
- Honestly scoped as a research tool, with built-in "not financial advice" guardrails and risk-management reminders
- First run downloads ~500MB of weights and heavy deps like torch, taking 3-7 minutes
- Data depends on yfinance (Yahoo Finance); some symbols lack data and intraday candles are limited to the last 60 days
- Repository license is unknown, leaving commercial-use risk unclear
- No test suite; issues like OOM or failed venv installs require manual troubleshooting
- Marketing framing like "Wall Street's billion dollar AI" is unverified — prediction accuracy has no independent benchmark in the source
How do you install this skill?
Place the skill folder at ~/.claude/skills/kronos/ (containing SKILL.md and scripts/run_kronos.py). No manual installation is required: on first invocation the script clones the Kronos repo, creates a Python venv, installs torch/transformers/yfinance/pandas, and downloads ~500MB of model weights from HuggingFace (3-7 minutes). The repository license is unknown — verify before adopting.
How do you use this skill?
In any Claude conversation: "run kronos on AAPL" (defaults to 6 months of daily candles, 24-period forecast), "kronos BTC-USD 3mo 4h 48", or "predict NVDA with kronos for the next 30 days" (Claude infers args). You can also call the script directly: python3 ~/.claude/skills/kronos/scripts/run_kronos.py <TICKER> [period] [interval] [pred_len]. Subsequent runs take ~5-15 seconds.
How does this skill compare with similar options?
Built on the same author-ecosystem Kronos open source model (shiyu-coder/Kronos, MIT licensed); this skill is essentially a conversational Claude wrapper around that model. The author also recommends combining it with a manually curated Claude Project of company filings as a two-layer stack.