Vibe-Trading AI trading desk illustration

Links & Resources

A research lab at the University of Hong Kong (HKUDS) open-sourced a full quant trading desk that runs from your terminal. Type what you want in plain English - a team of AI agents researches the market, drafts the strategy, runs the backtest, and reviews the risk.

It's called Vibe-Trading, has 18K+ GitHub stars, ships with 456 pre-built quant factors and 18 live data sources. MIT licensed. Free.

The Specs

  • Org: HKUDS (HKU Data Intelligence Lab)
  • Stars: 18K+, 2.9K forks
  • License: MIT
  • Install: pip install vibe-trading-ai
  • LLMs: 13+ providers (OpenAI, Claude, Gemini, DeepSeek, Qwen, Ollama...)
  • Stack: Python 3.11+, React 19, FastAPI, MCP

Multi-Agent Trading Teams

Instead of one model doing everything, Vibe-Trading runs 29 swarm presets - coordinated teams of specialized agents. A few:

  • investment_committee - Bull/bear debate → risk review → PM final call
  • global_equities_desk - A-share + HK/US + crypto researcher → global strategist
  • crypto_trading_desk - Funding/basis + liquidation + flow → risk manager
  • earnings_research_desk - Fundamental + revision + options → earnings strategist

Every worker is grounded with fetched market data, not model priors. That's the difference between this and asking a chatbot for stock tips.

Under the Hood

Every research run flows through a five-layer pipeline:

Plan → Ground → Execute → Validate → Deliver

The Ground step fetches point-in-time data before reasoning. The Validate step runs Monte Carlo permutation tests, Bootstrap confidence intervals, and Walk-Forward validation - so you see whether an edge is real or a fluke before risking a dollar.

456 Pre-Built Alphas

Vibe-Trading ships an "alpha zoo" drawn from published research - Microsoft Qlib 158, Kakushadze's Alpha101, Guotai Junan 191, and the academic FF5/Carhart family. One line to bench them all on your universe:

vibe-trading alpha bench --zoo gtja191 --universe csi300 --period 2018-2025 --top 20

It categorizes every factor as alive, reversed, or dead on your data.

The Sleeper Feature: Shadow Account

Upload your real broker export and Vibe-Trading diagnoses your actual trading behavior. It parses exports from Tonghuashun, Eastmoney, Futu, and generic CSV, then profiles you against known leaks:

  • Disposition effect - selling winners too early, holding losers too long
  • Overtrading - churning that eats you on fees and slippage
  • Momentum chasing - buying tops
  • Anchoring - fixating on entry price

It extracts your implicit rules, backtests them as a "shadow" strategy, and shows where you broke them.

vibe-trading --upload trades_export.csv
vibe-trading run -p "Analyze my trading behavior and compare it with my actual trades"

Quick Start

pip install vibe-trading-ai
vibe-trading run -p "Backtest a BTC-USDT 20/50 moving-average strategy for 2024, summarize return and drawdown, export the report"

Bring your own LLM key - or run Ollama locally and pay only for compute. Exports to TradingView Pine v6, TDX, MetaTrader 5, and vnpy. Live trading connectors default to paper mode with a kill switch and audit ledger.

Honest Caveats

MIT-licensed code, but you still supply LLM keys and data-provider keys. Monte Carlo and walk-forward reduce overfitting risk - they don't eliminate it. Backtested performance is not live performance. Nothing here is financial advice.

Also: there's a phishing Discord impersonating the project. Only use the official HKUDS link above.