Kronos Foundation Model and Day Trader Skill Pack for Claude Code Guide
A guide to pairing Kronos, an open-source K-line foundation model, with Day Trader's 11 Claude skills for chart scanning, 10,000-session simulation, and paper-t
Kronos is an open-source foundation model for financial candlesticks that was accepted at AAAI 2026. A related workflow pairs it with the Day Trader pack of 11 Claude skills for chart scanning, trade planning, simulation, and paper trading. In an X post, @carm1nee attributed Kronos to Tsinghua University researchers and Day Trader to Jake Schincariol.
Image source: https://github.com/shiyu-coder/Kronos
Both projects use the MIT License. Day Trader uses plain Python and free public data from Yahoo Finance charts, Nasdaq Trader symbol lists, SEC EDGAR, and the House Clerk. Its paper-trading workflow requires user approval before an order.
Kronos: Decoder-Only Foundation Model for Candlestick Sequences
Kronos is an open-source decoder-only foundation model pre-trained on candlestick (OHLCV) sequences across more than 45 global financial exchanges. Unlike general-purpose time-series forecasting models, Kronos is engineered specifically to manage the high-noise characteristics inherent in financial markets using a specialized two-stage architecture.
First, a dedicated tokenizer quantizes continuous, multi-dimensional K-line data into hierarchical discrete tokens. An autoregressive Transformer then models these token sequences to generate probabilistic forward paths. Pre-trained weights available on Hugging Face cater to varying resource requirements:
- Kronos-mini: 4.1M parameters, 2048 token context window (uses Kronos-Tokenizer-2k)
- Kronos-small: 24.7M parameters, 512 token context window (uses Kronos-Tokenizer-base)
- Kronos-base: 102.3M parameters, 512 token context window (uses Kronos-Tokenizer-base)
Through Python and the KronosPredictor class, users can pass historical price data frames and generate multi-path future projections by adjusting sampling temperature and top-p nucleus parameters.
Day Trader: 11 Claude Skills Powered by Standard-Library Python
The Day Trader repository provides 11 modular Claude skills. Its tools require Python 3.8 or newer, use standard-library Python without packages to install, and read public data from Yahoo Finance charts, Nasdaq Trader symbol lists, SEC EDGAR, and the House Clerk.
After installation, users invoke the tools as Claude Code slash commands. /trader-scanner scans for opening-range breakouts, flags, and oversold bounces; /trader-news tracks headlines and filings; /trader-entry watches a setup bar by bar.
Exit points and risk parameters are established before entry. /trader-exit sets the stop and target, /trader-budget sizes the position to the budget and maximum loss per trade, and /trader-risk checks the order against configured limits.
Time Machine Simulation and Paper Trading Safeguards
Time Machine (/trader-timemachine) simulates setups before capital is risked. It draws 5-minute candles from the ticker's recent bars, replays 10,000 simulated sessions, and also checks past sessions in which the same setup fired.
When evaluating these tools, users must account for clear operational boundaries:
- Research-code scope: The Kronos repository provides forecasting examples and scripts for fine-tuning on user datasets.
- Simulation Rather Than Prediction: Time Machine uses the stock's recent 5-minute bars. Earnings reports, halts, or headlines can produce conditions absent from that history.
- Paper Trading by Default: Day Trader defaults to paper trading. Only
/trader-brokercan lead to an order, after/trader-riskpasses it and the user explicitly approves it. - Execution vs Price Guarantees: Stop-loss bracket orders represent conditional execution instructions and do not guarantee filled prices in volatile or gapping markets.
Sources
- GitHub Repository: shiyu-coder/Kronos
- GitHub Repository: Jakeschincariol/day-trader-skill
- Carmine on X: Carmine (@carm1nee) on X